André B.

Dorsman
Kazim Baris Atici
Aydin Ulucan
Mehmet Baha Karan Editors

Applied Operations
Research and
Financial Modelling
in Energy
Practical Applications and Implications
Applied Operations Research and Financial
Modelling in Energy
André B. Dorsman · Kazim Baris Atici ·
Aydin Ulucan · Mehmet Baha Karan
Editors

Applied Operations Research

and Financial Modelling
in Energy
Practical Applications and Implications
Editors
André B. Dorsman Kazim Baris Atici
VU University Amsterdam Hacettepe University
Amsterdam, The Netherlands Ankara, Turkey

Aydin Ulucan Mehmet Baha Karan

Hacettepe University Hacettepe University
Ankara, Turkey Ankara, Turkey

ISBN 978-3-030-84980-1 ISBN 978-3-030-84981-8 (eBook)

https://doi.org/10.1007/978-3-030-84981-8

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Switzerland AG 2021
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Foreword by Otto Jager

For Everyone on This Planet

President John F. Kennedy in 1961 proclaimed that the United States would place
a man on the moon by the end of the decade. In 1969, Neil Armstrong took his
‘one small step’ on July 20th—the Americans made it with five-and-a-half months
to spare. After a devastating flood in 1953 that took 1836 lives, the Dutch decided
to mostly close the North Sea delta and build a storm surge barrier. Twelve building
projects and 43 years later, the coastline is reduced from 700 km to only 80 km.
The energy transition is often compared to the man on the moon-mission or to
the Delta works—mostly to emphasize the vastness of the challenge we are now
working on. Exceptional and impressive as those achievements forever will be, the
energy transition surpasses them by far. The change to a renewable energy system will
have serious impact on every economical sector and on the life of every individual
in our society. Armstrong’s ‘small step for a man’ needs an upgrade: ‘The energy
transition will be an upswing for everyone on this planet’.
It is exactly that aspect that makes this transition such an immense challenge:
it depends on the cooperation of all people. More than ever before, decisions on
remodelling the energy system will be affected by social, technological and economic
trends. The world of energy will be more dynamic, changes will be more drastic and
people will be more involved in renewable solutions. Extra challenging, but also an
open door for new chances—illustrated by the success and potential of green bonds
in financial markets.
Of course, the energy sector is used to looking forward for at least a decade and
anticipate on substantial changes. In this process, we need all the information and
analyses available to make the right decisions. These studies on Applied Operations
Research and Financial Modelling in Energy contribute to a better understanding of
policy implications of the proposed or applied methodologies. They also show the
value of using the right models and methods for decision making. I’m convinced that

v
vi Foreword by Otto Jager

these perspectives from Operations Research and Finance will have a positive effect
on the quality of our decisions—strategic, tactical, and operational.

Otto Jager
Chief Financial Officer
TenneT TSO B.V.
Arnhem, The Netherlands
Contents

Introduction: Applied Operations Research and Financial

Modeling in Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
André B. Dorsman, Kazim Baris Atici, Aydin Ulucan,
and Mehmet Baha Karan
Optimization Methods on Electricity Generation and Transmission
Expansion Planning Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Mahdi Noorizadegan and Alireza Shokri
Demand-Driven Electricity Supply Options of Electric Vehicles:
Modelling, Simulation, and Management Strategy of Public
Charging Stations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Elvin Coban and Gokturk Poyrazoglu
A Review on Smart Energy Management Systems in Microgrids
Based on Power Generating and Environmental Costs . . . . . . . . . . . . . . . . 51
Özgür İcan and Taha Buğra Çelik
Measuring Efficiency and Productivity Change in the Turkish
Electricity Distribution Sector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Yetkin Cinar and Tekiner Kaya
Price and Volatility Forecasting in Electricity with Support Vector
Regression and Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Mahmut Kara, Kazim Baris Atici, and Aydin Ulucan
Forecasting the Hydro Inflow and Optimization of Virtual Power
Plant Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Sezer Çabuk, Özenç Murat Mert, A. Sevtap Selcuk-Kestel, and Erkan Kalaycı
Comparing the Renewable Energy Technologies via Forecasting
Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
Fazıl Gökgöz and Fahrettin Filiz
Business Cycles and Energy Real Options Valuation . . . . . . . . . . . . . . . . . . 173
Turalay Kenc and Mehmet Fatih Ekinci
vii
viii Contents

Understanding the Electricity Switching Behavior of Industrial

Consumers: An Empirical Study on an Emerging Market . . . . . . . . . . . . . 201
Murside Erdogan, Selin Metin Camgoz, Mehmet Baha Karan,
and M. Hakan Berument
Does the Market Value Clean Innovation? Evidence from US
Listed Firms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
Antoine Dechezleprêtre, Cal B. Muckley, and Parvati Neelakantan
The Power Grid: From a Technical to a Finance Issue. Who Bears
the Financial Risk? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
André B. Dorsman and Kees van Montfort
Introduction: Applied Operations
Research and Financial Modeling
in Energy

André B. Dorsman, Kazim Baris Atici, Aydin Ulucan,

and Mehmet Baha Karan

1 Introduction

Decisions in the energy sector are generally complex with multiple conflicting objec-
tives. Accumulating demand, increasing competition, rising awareness on environ-
mental issues, together with evolving rules and regulations are all binding the energy
sector with environmental, social, and financial pressure to keep the production and
distribution processes under control. The planning in the sector usually involves
many sources of uncertainty and risk, varying time frames, and a large number
of stakeholders with different views, which makes the application of Operations
Research (OR) methods particularly suitable. There exists a vast literature on energy
sector applications of OR methodologies. This is due to fact that optimization and
rational decision-making are vital to building up more sustainable energy manage-
ment systems in such a dynamic and competitive business environment. In this vein,
financial decisions are one of the main legs to be handled since energy investments
are usually capital intensive.
Financial Modelling (FM) is a scientific approach for decision making and its
methods are capable of supporting financial decision making at different levels of
various sectors as well as the energy sector. It is possible to identify a wide spectrum
of application areas in energy finance that may include but not limited to pricing

A. B. Dorsman (B)
VU University Amsterdam, Amsterdam, Netherlands
K. B. Atici · A. Ulucan · M. B. Karan
Department of Business Administration & Energy Markets Research and Application Center,
Hacettepe University, Ankara, Turkey
e-mail: [email protected]
A. Ulucan
e-mail: [email protected]
M. B. Karan
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1

A. B. Dorsman et al. (eds.), Applied Operations Research and Financial Modelling
in Energy, https://doi.org/10.1007/978-3-030-84981-8_1
2 A. B. Dorsman et al.

and hedging decisions, understanding the market dynamics and managing demand,
measuring the effectiveness of regulations, assessing the feasibility and efficiency of
investments, supervising cash flows, capital budgeting, allocating resources and eval-
uating risk. This is achieved by the employment of an extensive range of quantitative
tools such as deterministic and stochastic optimization, multiple criteria, multiple
objective and fuzzy decision-making methods, simulation, econometric modeling,
statistical inference, and contemporarily by application of learning algorithms.
Applied Operations Research and Financial Modelling in Energy (AORFME)
aims to contribute to the both academic and practitioner sides of the energy sector by
offering several modeling applications followed by policy implications to aid various
decisions faced in the sector. The chapters of the book mainly aim to focus on a
variety of energy decisions, to present a quantitative perspective on these decisions,
and to provide policy implications of the proposed or applied methodologies. For
this purpose, we bring a group of OR and Finance researchers together and present
a collection of chapters that contribute to the applications in the field of Energy.

2 Operations Research and Financial Modelling in Energy

The volume comprises twelve chapters on the application of Operations Research and
Financial Modelling in the energy sector grouped into three main parts. Within the
scope of the book, a vast array of problems devoted to energy markets is addressed
such as electricity generation & transmission planning, location and price deci-
sions, smart energy management systems, efficiency evaluation, price & volatility
forecasting, power plant pricing, the potential of renewable investments, valuation
issues, consumer decisions, and financial risk analysis. The book presents research
on both macro and micro levels as well as research focus on market dynamics,
renewable energy, pricing, and capacity decisions.
The above problem areas have been undertaken by the authors of the chapters
with a variety of methods. Methods implemented involve a range of techniques
of Operations Research and Financial Modelling from optimization to forecasting
and from conventional statistical methods to machine learning methods. Examples
include Mixed Integer Programming, Simulation-based Optimization, Data Envel-
opment Analysis, Time Series Forecasting, Malmquist Productivity Indices, Arith-
metic Brownian Motion, Principal Component Analysis, Logistic regression, Deep
Learning Methods, Support Vector Regression, and Random Forest.
Regarding the areas of application/industries, the volume presents research
devoted to different types of energy problems. The application areas also involve
variety with respect to means of production and networks. The chapters include
research on integrated natural gas & power networks, hydroelectricity power plants,
electricity distribution companies, electric vehicles, wind-hydro energy technologies,
and smart energy management systems in microgrids. There also exist chapters on
energy options, electricity markets, attitudes of industrial consumers, and valuation
of clean innovation.
Introduction: Applied Operations Research … 3

3 Issues Covered in This Book

The book is divided into three main parts listed below, each presenting a number
of chapters that focus on the abovementioned problem instances, methods, and
application areas of research:
Part A. Applied OR I: Optimization Approaches.
Part B. Applied OR II: Forecasting Approaches.
Part C. Financial Modelling: Impacts of Energy Policies and Developments in
Energy Markets.
Part A of the book consists of four chapters in which several optimization related
applications on energy are addressed: (i) electricity generation and transmission
expansion planning, (ii) demand-driven electricity supply options of electric vehicles,
(iii) smart energy management systems in microgrids, (iv) efficiency and productivity
change in the electricity distribution sectors. The methods vary from optimization to
simulation as well as their integrated use.
The book starts with the research of Mahdi Noorizadegan and Alireza Shokri
(Chap. “Optimization Methods on Electricity Generation and Transmission Expan-
sion Planning Problem”) on energy generation and transmission line expansion plan-
ning. After carefully reviewing the expansion problem domain in terms of problem
setting & modeling, types of uncertainty, and the solution methods, Noorizadegan and
Shokri propose a simulation-based optimization framework to handle the complex
problem of generation and transmission problems (GTEP) with key features and
methods inspired by their review. Their framework aims to capture the uncertainty
of both the electricity load and the power generation by renewable resources. The
framework suggests starting with an initial problem that leaves the complex compo-
nents out and simplifies the model. Then, simulation is suggested for improving the
constraints and the solution to this initial optimization problem. The authors point out
that the resulting framework is advantageous because instead of searching the entire
feasible region in a large-scale problem, it relies on decomposing it and introducing
the complexities of the problem to a simpler initial model step-by-step.
The next Chap. ”Demand-Driven Electricity Supply Options of Electric Vehi-
cles: Modelling,Simulation, and Management Strategy of Public Charging Stations”
presents research on a contemporary topic: Electric Vehicles (EVs) and their charging
stations (CSs). Elvin Coban and Gokturk Poyrazoglu discuss the challenges and
research opportunities in the demand-driven electricity supply options of electric
vehicles at public charging stations. This is a comprehensive research that covers
different aspects of the electric vehicle charging stations from location to pricing.
After reviewing the strategic, tactical, and operational level problems related to CSs,
a discussion on the existing modeling approach to locate CSs and their potential
extensions are presented. Following that, a simulation framework is described to
decide the number and type of chargers. Finally, the authors discuss different pricing
policies and potential future problems related to CSs. Last but not least, Coban and
Poyrazoglu’s research provides an extensive look at one of the important concerns
of the future: public charging networks.
4 A. B. Dorsman et al.

This chapter is followed by research on another contemporary matter: Smart

Energy Management Systems (SEMS). Ozgur Ican and Taha Bugra Celik, in their
Chap. “A Review on Smart Energy Management Systems in Microgrids Based On
Power Generating and Environmental Costs”, offer a review on renewable energy,
microgrids, and Smart Energy Management Systems. They rely on the optimiza-
tion methods used in improving SEMS and investigate the common grounds for
computational frameworks employed within these systems. They present a compre-
hensive discussion and table on the previous research on SEMS, their methods, and
a comparison of their results in terms of power generating and environmental costs.
The next chapter presents an application of efficiency measurement based on
linear programming. Yetkin Cinar and Tekiner Kaya’s Chap. “Measuring Efficiency
and Productivity Change in the Turkish Electricity Distribution Sector” looks at the
efficiency of the Turkish electricity distribution sector and its change over time using
well-known efficiency measurement methods of OR literature: Data Envelopment
Analysis and Malmquist Productivity Index. Relying on the large-scale privatization
experienced by the Turkish electricity sector starting from 2013, Cinar and Kaya
evaluate the efficiency in the post-privatization period. After a detailed review of DEA
applications in electricity distribution sectors, they measure the efficiency levels of
Turkish distribution companies, investigate the relationship of the efficiency levels
with several exogenous factors and assess the level of change over 5 years. The
chapter serves as a compact application of DEA and MPI supported by a review of
the related research which has been an interest since the 1990s.
Part B of the book is designed to present research on the forecasting methods
applied in the energy sector. Main research interests are (i) price and volatility fore-
casting in electricity markets, (ii) forecasting of the hydro inflow and optimization
of virtual power plant pricing, (iii) comparison of renewable energy technologies
via forecasting, (iv) valuation of energy real options with regime shifts. In this
part, the methods vary between the conventional econometric models to learning
methodologies.
Part B starts with a Chap. “Price and Volatility Forecasting in Electricity with
Support Vector Regression and Random Forest” by Mahmut Kara, Kazim Baris
Atici, and Aydin Ulucan. The chapter aims at contributing to the contemporary
research stream of machine learning applications to electricity markets for forecasting
prices and volatility. The chapter serves as a neat application of learning tools to the
electricity markets. The authors present two types of forecasting schemes (Price &
Volatility) using two types of modeling approaches (Support Vector Regression &
Random Forest) in Turkish day-ahead electricity markets. Within the scope of the
research, utilizing the hourly Euro prices January-2013 and September-2019 period,
a rolling data scheme is designed to produce hourly prices for 340 weeks considering
16 features. The volatility forecasting covers realized volatility values comprising
more than 2000 days and 10 features. The metrics of SVR and RF are compared
with each other in terms of each scheme as well as with the metrics of the naive
estimations.
Introduction: Applied Operations Research … 5

The next Chap. “Forecasting the Hydro Inflow and Optimization of Virtual Power
Plant Pricing” is on hydroelectricity and features two-part research focusing on hydro
inflow forecasting and virtual power plant pricing. The chapter combines forecasting
and optimization methodologies within a well-designed framework. The authors,
Sezer Cabuk, Ozenc Murat Mert, A. Sevtap Selcuk-Kestel, and Erkan Kalayci
propose a multiple-stage framework for hydroelectricity power plants that every
stage’s output is input to the next stage resulting in virtual power plant pricing. The
hydro inflow forecasting is accomplished by utilizing Seasonal ARIMA with eXoge-
nous variables (SARIMAX). The output of this stage, the forecasts, is used as the
input for the hydro optimization model to forecast the water capacity for the future.
On the other hand, they determine the price behavior using Monte Carlo simulations.
Once capacity and price have been modeled, the virtual hydropower plant values are
estimated.
The following Chap. “Comparing the Renewable Energy Technologies via Fore-
casting Approaches” by Fazil Gökgöz and Fahrettin Filiz focuses on evaluating
wind and hydro energy potentials through forecasting tools as conventional regres-
sion methods and deep learning methods as Long Short-Term Memory (LSTM) and
Gated Recurrent Unit (GRU). Utilizing Turkish wind and hydropower electricity
generation data, Gökgöz and Filiz discuss the strengths and weaknesses of several
forecasting methods in predicting electricity generation using renewable resources
with an emphasis that forecasting tools may serve as an effective tool for policymakers
in the sector.
Part B ends with the Chap. “Valuing Energy Real Options with Regime Shifts”
by Turalay Kenc and Mehmet Fatih Ekinci. The authors focus on the real options
approach to value energy projects since these investments possess a high level
of macroeconomic risk. After introducing, a basic real options valuation model
with regime shifts, they derive a framework based on the Arithmetic Brownian
motion (ABM) process with regime shifts for valuing the energy real options. The
proposed model is illustrated in numerical analysis with a detailed discussion of its
implications.
Finally, Part C is devoted to Financial Modelling. The part consists of three chapters
that financial modeling related applications on energy are presented: (i) analysis of
electricity switching behavior of industrial consumers, (ii) feasibility and potential of
renewable investments in Tanzania, (iii) valuation of clean innovation, (iv) the power
grid as a technical to a finance issue. The methods vary from Brownian Motion to
statistical tests to validate the various hypothesis.
Part C starts with the Chap. “Understanding the Electricity Switching Behavior
of Industrial Consumers: An Empirical Study on An Emerging Market” by Murside
Erdogan, Selin Metin Camgoz, Mehmet Baha Karan, and M. Hakan Berument.
The chapter is focusing on the supplier switching behavior of industrial electricity
consumers. The authors present empirical research based on survey data consisting
of items for risk of switching, cost of switching, the attractiveness of switching,
perceptions of the service quality, and market competition. The relation between
these items and the probability of switching from suppliers is established using a
6 A. B. Dorsman et al.

binary logistic regression model. The results of the research aim to shed light on the
decisions of electricity suppliers, regulatory agencies, and policymakers.
In Chap. “Does the Market Value Clean Innovation? Evidence from US Listed
Firms”, Antoine Dechezleprêtre, Cal B. Muckley, and Parvati Neelakantan aim at
bringing new insights to the corporate environmental-financial performance debates.
Utilizing the US patent data for the period 1995 to 2012, econometric modeling
is used to disaggregate the innovation measurements into clean, dirty, and other
components. The analysis reveals an important finding that environment-friendly
innovation pays off.
The final chapter of the book (Chap. “The Power Grid: From a Technical to a
Finance Issue. Who Bears the Financial Risk?”) is written by André B. Dorsman and
Kees van Montfort. In their chapter, the authors provide insight into financial relations
between various stakeholders of the Dutch electricity market. The authors provide
an understanding of the Dutch energy sector dynamics of clearing and the margin
requirements in financing. After establishing the key parties and system players, the
chapter discusses the quantification and bearing of the financial risks on the future
cash flows in the energy sector.

4 Concluding Remarks

To conclude, with a wide range of look into the energy sector decisions from
the perspective of OR and Finance perspectives, we hope that Applied Operations
Research and Financial Modelling in Energy (AORFME) contributes to the applied
research on energy-related issues and reaches its audience from the both academic
and practitioner sides of the energy sector.
This is the eighth volume in a series on energy organized by the Centre for Energy
and Value Issues (CEVI). In this volume, CEVI collaborates with Hacettepe Univer-
sity Energy Markets Research and Application Center. The previous volumes in
the series were: Financial Aspects in Energy (2011), Energy Economics and Finan-
cial Markets (2012), Perspectives on Energy Risk (2014), Energy Technology and
Valuation Issues (2015), Energy and Finance (2016), Energy Economy, Finance and
Geostrategy (2018) and Financial Implications of Regulations in the Energy Industry
(2020).
The editors would like to thank the authors for their valuable contributions and
the reviewers for their effort to improve the quality of this book project. We would
like to thank also the Springer staff for their continuous support.
Optimization Methods on Electricity
Generation and Transmission Expansion
Planning Problem

Mahdi Noorizadegan and Alireza Shokri

1 Introduction

Electricity is considered as the heart of modern economies and is predicted to have a

significant increase in its share in the global energy mix i.e., twice the rate of primary
energy demand (IEA, 2019). Solar and wind will have the highest growth rates among
other electricity resources from 2018 to 2040 (IEA, 2019). According the same report,
in a sustainable development scenario where electricity plays a larger role in energy
demand, renewable resources will account for two thirds of global electricity demand.
Therefore, given the energy mix transition towards electricity (particularly renew-
able sources), energy planning studies mainly focus on power Generation Expansion
Planning (GEP). In general, GEP seeks an optimal investment of power generation
units over a planning horizon to meet predicted/projected energy consumption (load)
subject to a variety of constraints and considerations. Moreover, transmission facil-
ities play important economic and technical roles in GEP as their installation cost
and technical constraints could have substantial impacts on generation expansion
decisions. Therefore, many studies combine these two problems and reviewed an
integrated generation and transmission expansion planning problem (GTEP). Whilst
there are different versions of GEP, the main decision variables include investment
schedule for generation units. While transmission expansion planning is included
in the problem, location of new power generation units and decisions for trans-
mitting power from generating locations to demand points/areas are also decided.

M. Noorizadegan (B) · A. Shokri

Newcastle Business School, Northumbria University, Newcastle Upon Tyne, UK
e-mail: [email protected]
A. Shokri
e-mail: [email protected]
M. Noorizadegan
Niroo Research Institute, Tehran, Iran

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 7

A. B. Dorsman et al. (eds.), Applied Operations Research and Financial Modelling
in Energy, https://doi.org/10.1007/978-3-030-84981-8_2
8 M. Noorizadegan and A. Shokri

However, decision variables are not limited to only these generation units and trans-
mission lines/corridors. Depending on the setting and assumptions, a problem may
include many other types of decision variables such as decommission decisions,
power generated by units, phase angles of voltages and currents, etc. For instance,
Micheliet al. (2020) consider decommissioning variables. Direct Current (DC) load
flow is an approximation of Alternating Current (AC) load flow and has been consid-
ered in some studies (Caunhye & Cardin, 2018) while it has been ignored in many
GTEP related studies. This involves the computation voltage angle which depends
on geographical properties and technical characteristics of transmission lines. Some
studies (Coester et al., 2018) incorporated less technical details and instead focused
more on economic analysis and environmental aspects of GTEP and GEP.
In recent years, technological and economic advancements in renewable sources
as well as environmental requirements have directed the focus of energy planning
problems towards GTEP with high renewable energy penetration. Despite their
advantages, renewable sources impose considerable complexities to power supply.
Although cost of renewable sources has substantially declined (e.g., 70% for solar
from 2010 to 2018), it seems that renewables still cannot effectively compete with
thermal technologies as their cost has also reduced (Feldman & Margolis, 2019;
Fu et al., 2018). Therefore, governments designed attractive incentive schemes to
encourage companies for investing in renewable sources. Levin et al. (2019) studied
incentives mechanisms under four categories: (1) investment support, (2) genera-
tion support, (3) quantity targets, and (4) carbon policies. There are various studies
for further investigation of incentive mechanisms (Alolo et al., 2020; Newbery,
2016). Because of their uncertain power generation, integrating renewable sources
into existing power systems which mainly consist of thermal units, is complex and
requires sophisticated planning and scheduling. Moreover, technical restrictions such
as rampage constraints of thermal units limit the utilisation of renewable sources.
For instance, Duck curve is a critical concept to address the impact of power gener-
ation by solar units on power systems (Denholm et al., 2015). Capacity factors of
renewable sources are another important component that can have a considerable
impact on economic and technical analysis in GTEP. Capacity factors are usually
estimated for an entire year. In fact, this type of complexity and limitation makes
renewable sources more expensive. The wide use of renewable sources imposes
another complication to GEP. The consumption of natural gas by gas-fired units is
significantly affected by the uncertainty of’ power generated by renewable sources.
In other words, the production of gas turbines needs to be adjusted with respect to
power generation changes of renewable sources to satisfy demand. As a result, the
uncertainty of renewables is transferred to the gas network. In order to maintain gas
pressure at a safe level in a gas network for other usage (e.g., residential and industrial
sectors), gas and electricity storage devices and gas compressors need to be installed.
Integrated gas and power networks have been studied in operational level (Fallahi &
Maghouli, 2020a). However, recently there has been an interest for this integration
in planning problem (Conejo et al., 2020). Such problems are in general non-convex
non-linear mixed integer problems (Esmaili et al., 2020). The source of non-linearity
is the gas flow equation known as Weymouth equation.
Optimization Methods on Electricity Generation and Transmission … 9

This book chapter provides a relatively comprehensive overview on GTEP and

suggests an optimisation modelling framework for GTEP that includes important
features. The rest of this chapter is organised as follow. In Sect. 2, we focus on the
mathematical modelling of a general GTEP where various types of objective func-
tions and constraints are discussed. In Sect. 3, we discuss two types of uncertain-
ties: demand and power generation of renewable resources, and equipment failure.
We suggest Interval Optimisation to deal with demand and generation of renew-
able resources and a cutting plane-based method for equipment failure. The both
approaches are conservative and consider the worst possible situations. In Sect. 4,
we briefly review the solution methods and suggest a simulation-based optimisation
framework for solving practical GTEP problems. Finally, we provide a summary of
this chapter in Sect. 5.
Notation—In this chapter, we use a simple notation
 D for
 simplicity. Bold face char-
acters and symbols indicate vector i.e., xtD := xikt ik
. Subscripts denote indices
while superscripts denote the type of variables (e.g., D stands for decommissioned
equipment). The main sets, indices, decision variables and parameters are defined as
per below. The rest of variables and parameters are defined where needed.
Sets, indices, superscripts
I : the set of locations (nodes),
t: index for time period (year),
Tt : a subset within time period t,
i, j: location indices,
h: index for hours of a day,
k: index for technology type (either power generation unit or transmission line),
l: index for fuel type,
G: used to denote the natural gas,
st: used to denote storage devices,
dis: used to denote discharging storage devices,
ch: used to denote charging storage devices,
D: used to denote decommissioned equipment,
N : used to denote new equipment.
Decision variables
 N
xtN := xikt ik
: binary variables representing whether at location i, a unit of
technology k at time period t is installed,
xtD := xikt D
ik
: binary variables representing whether a unit of technology k
located at location
 i at time period t is decommissioned,
yt := yi jkt
N N
: binary variables representing whether between locations i and
i jk
j, a transmission line of technology k at time period t is installed,
pht := [ pikht ]ik : continuous variables representing power generation at location
i with technology k at hour h in time period t,
10 M. Noorizadegan and A. Shokri
 
fht := f i jkht i jk : continuous variables representing power flow between locations
i and j using transmission line technology k at hour h in time period t,
Iht = [Iikht ]: continuous variables representing inventory of storage at location i
with technology k at hour h in time period t,
θi ht : continuous variables representing the phase angle at location i at hour h in
time period t,
disht := [disikht ]ik : continuous variables representing discharge of storage at
location i with technology k at hour h in time period t,
chht := [ch ikht ]: continuous variables representing discharge of storage at
location i with technology k at hour h in time period t,
πi ht : continuous variables representing gas pressure at location i at hour h in time
period t,  
ght := gi j ht i j : continuous variables representing natural flow between locations
i and j at hour h in time periodt,
M, MG : large enough numbers.
Parameters
Q := [Q k ]k : maximum  capacity of power generation unit of technologyk,
Qmin := Q min k k
: minimum level of power generation unit of technologyk,
F := [Fk ]k : maximum capacity of power flow for line of technologyk,
r := [rk ]k : increasing ramp rate for power generation unit of technologyk,
L lTt : maximum available fuel of type k in time period Tt
αh := [αkh
   ]k : capacity factor for hourly power generation unit of technology k,
x := x ik ik : indicator for existing unit at location i, a unit of technology k in the
beginning
   of the planning horizon,
y := y i jk i jk : indicator for existing transmission line between locations i and j
of technology k in the beginning of the planning horizon,
πimin , πimax : minimum and maximum permitted gas pressure at location i.
 

Note that xt = x + xtN and yt = y + ytN . We also eliminate the transpose sign in
the notation.

2 Mathematical Model

Key components of GEPR include but not limited to objective function, environ-
mental impacts of sources of power generation, reliability, resiliency, uncertainty,
operational restriction and consideration, and impact of power generated by gas
network. GTEP decides on facilities (generation units and transmission lines) with
effective lives of more than 30 years. Therefore, similar to long term planning prob-
lems, some technical details such as daily operational details are replaced with their
approximations or ignored. However, this could affect the electricity mix leading to
higher costs or in some severe cases infeasible situations in reality. On the other hand,
building and solving a GTEP with many details are formidable tasks. Therefore, a
Optimization Methods on Electricity Generation and Transmission … 11

reasonable trade-off between operational details and the problem complexity and
computational challenges is usually sought.

2.1 Objective Function

Whilst majority of studies consider cost-based models, some (Allahdadi Mehrabadi

et al., 2020; Lohmann & Rebennack, 2017) study a welfare or profit-based objective
functions. Electricity markets will have to be simulated to detect the electricity price
to model a maximisation of GTEP model. Simulating electricity markets with reason-
able details is a complex topic. However, some studies (Coester et al., 2018) applied
a rather simple methods such as Merit Order Curve. In such models, regulators also
play an important role in setting electricity markets. We refer to Cramton et al. (2013)
for further discussion on energy and capacity markets. The main components of cost
for GTEP include investment, decommission, operation, and fixed maintenance costs.
Investment cost—The investment for power system equipment is capital intensive
and usually involves long-term financial arrangements. Uncertainty of demand and
power generation by renewable sources complicates the risk assessment for investors.
Hence, some studies (de Oliveira et al., 2017; Simo et al., 2015) formulate GTEP as a
dynamic program. GTEP can be considered as a facility location problem with fixed
cost. This approach essentially requires a long-term planning horizon (to include the
full effective lifecycle of all equipment) in order to make a right balance between
operational and investment costs. However, due to complications such as disparity
of lifecycle of different equipment, it is not always possible. Instead of total fixed
investment costs, an equivalent annual cost for each equipment is computed. In this
situation, additional constraints are required to ensure of availability of a selected
facility for its entire lifetime. The investment cost of facility is computed towards
the end of planning horizon (Caunhye & Cardin, 2018; Ding et al., 2018).
Let Ftinv (xtN , ytN ) denote the investment cost function at time period t.
Decommission and upgrade cost—The main reasons to retire a unit are its high
maintenance and operational cost and its high rate of failure and unreliability. In
practice, even units may be used beyond their nominal effective lifetime when they are
properly maintained and looked after. Decommissioning some units such as nuclear
units incurs cost while decommissioning units such as gas turbines may lead to profit
as they have salvage values. In some cases, old units may not be decommissioned
and may be used to cover limited peak loads as an alternative for installing new units
for this purpose. The decision to keep old units or to decommission them should be
made through GTEP models. Moreover, units can be upgraded in an extra cost. In
some cases, upgrading old units can be more cost effective than investing in new
ones. For instance, gas turbines may be converted into combined cycle units or an
overhaul may significantly increase the efficiency of steam units. Upgrading a unit
usually involves considerable modelling complexities. Related constraints need to be
defined based on the type of upgrade. A simple solution is to define two new variables
12 M. Noorizadegan and A. Shokri

one for investment in the unit upgrade and one for decommissioning the old unit.
We use FtD,U (xt , yt ) to denote the cost function for decommissioning and upgrading
power generation units and transmission line at time period t. It is worth mentioning
that some units such as distributed generators have shorter effective lifecycle and
may be installed and also decommissioned within the planning horizon.
Operation cost—Alongside the investment cost, operation costs (i.e., variable cost)
comprise the main part of electricity cost. The operation cost includes fuel cost, water
supply, pollution and emission cost, start-up cost. All components of operation cost
may depend on the age of units. Thermal units can usually work with more than one
type of fuel, which increases the availability of units. However, the efficiency, and
emission produced by units depend on the type of fuel. For instance, due to restriction
of natural gas network in winter as the result of higher level of consumption, other
fuels such as Mazut are used in power plants. As such periods are usually short,
modelling other fuels can have a considerable impact on the electricity mix. The
reason is that when the natural gas limitation is enforced, a single-fuel GTEP model
would change the mix e.g., installing sufficient fuel-efficient units to ensure the feasi-
bility. Once fuel-efficient units are installed, the power generation plan and conse-
quently the operation cost would change. However, due to modelling and compu-
tational complexities, alternative fuels are generally ignored in GTEP. Temperature
and altitude also affect power generation by almost 10 percent (Sen et al., 2018).
Whilst it is not difficult to incorporate temperature and altitude in GTEP models,
their impacts are usually neglected. It is worth mentioning that the fuel consump-
tion function is not linear; but, a linear approximation is usually studied for more
simplicity. Start-up cost is often considered in GTEP models. Modelling start-up
requires constraints that link power generation in different hours. Such constraints
are complicating constraints and increase the computational complexity.
Although water supply is crucial for thermal units, it is not included in GTEP
studies as it is considered water is available everywhere. However, this is not a
correct assumption. Water supply at certain dry locations can be quite expensive
or in some areas impossible. We carried out a simple experiment to investigate the
impact of water supply cost and restrictions. We noticed that water supply in areas
with particular restrictions could play an important role in determining the electricity
mix.
Environmental consideration is among the most influential factors to shift elec-
tricity mix towards renewable sources. The emission cost is now a key part of opera-
tion costs, and is mainly considered for NOx, SO2 , CO, SPM, CO2 , CH4 and N2 O (Li
& Taeihagh, 2020). It is worth mentioning that the penalty for each one is different.
In addition to penalising, the production of some pollutants may be restricted. The
emission cost depends on several factors such as distance between power plants
and cities, population of cities, type of emission and pollution and type of fuel.
Moreover, some environmental consideration may forbid installing power plants in
special areas. This is important when the decision for transmission lines/corridors in
included.
opr
Let Ft (xt , yt ) denote the operation cost function at time period t.
Optimization Methods on Electricity Generation and Transmission … 13

Fixed and maintenance cost—There is usually a schedule for power plants and
unit maintenance, which depends on hours that each unit produces electricity in each
year. Some types of maintenance activities are short, but some are longer. The cost
for each type is therefore different. But it is common to consider a fixed cost for
annual maintenance for each unit depending on the technology of units. A fixed cost
is also considered for each unit, which does not depend on its performance. We use
f ix
Ft (xt , yt ) to denote the fixed and maintenance cost function at time period t.

2.2 Constraints and Technical Conditions

We classify the constraints and technical conditions of a GTEP model into three
groups: (1) investment related constraints, (2) capacity constraints, and (3) technical
constraints. These constraints are related to power generation units, transmission
lines, and gas network. A key factor in formulating a GTEP is time interval which is
usually hourly based intervals. But depending on the problem, 24 h in a day could
be split into 6 intervals. This will significantly reduce the number of variables and
constraints. In the following constraints, we consider hourly interval and present a
brief mathematical model for a general GTEP.
Investment related constraints—As mentioned in the objective function descrip-
tion, when the investment decisions are annually modelled, additional constraints are
required to ensure that once a unit is installed, it will be available for the rest of the
planning horizon. Analogously, we need to make sure once a unit is decommissioned,
it will be no longer available for production.

xtN ≤ xt+1
N
(1)

ytN ≤ yt+1
N
(2)

xtD ≥ xt+1
D
(3)

ytD ≥ yt+1
D
(4)

Capacity constraints—These constraints enforce capacity constraints for new and

existing power generation units and transmission lines.

phtl ≤ αh (Qxt + Q(1 − xtD ) (5)
l∈L

|fht | ≤ Fyt + F(1 − ytD ) (6)

14 M. Noorizadegan and A. Shokri


ηphtl ≤ L lTt (7)
l∈L

where η is the vector of fuel consumption rate for 1MWh corresponding to units
in vector phtl . When hydropower units are included in GTEP, additional constraints
for their power generation should be considered such as intakes and reservoir levels.
Since other entities (agricultural related organizations) are involved, hydropower
units may not be always available in particular during pick times.
Technical constraints—In an accurate model, all technical constraints in a Security
Unit Commitment (SUC) problem will have to be considered. However, as GTEP
is a long-term planning problem, only important conditions are studied. Given their
complexities, effective approximations for some constrains are developed and used
in GTEP. Here, we consider ramp rate constraints, start-up related constraints, and
DC power flow requirements. In order to formulate rampage constraints, additional
binary variables are needed for each unit. However, it may not be vital to include
such details for a GTEP. We suggest using the following ramp rate constraints:

pht − ph+1,t ≤ min{r, αh (Qxt + Q 1 − xtD − pht } (8)

The above inequality only enforces increasing ramp rate limits. A similar
inequality can be used for modelling decreasing ramp rate; but it is not crucial to add
the latter to the model. Equivalent ramp rate functions can be used to further simplify
the ramp rate constraints. Lohmann and Rebennack (2017) proposed an efficient way
of modelling unit start-up. They divided the power generation of each unit to two
L
parts: pht is a vector of variable for power generation of units up to their Qmin , and
U
pht is another vector of variables for power generation between Qmin and Q. Then,
L L
the difference between ph−1,t and pht approximates the start-up variable u ht . The
below inequalities compute the start-up variables

Qmin pht
L
+ Q − Qmin pUht = pht (9)

L
pht + ph−1,t
L
≤ uht (10)

pUht ≤ pht
L
(11)

The last inequality ensures that the load below minimum generation always
exceeds the load above minimum generation. It is trivial that without this inequality,
the start-up variable may be zero. AC load flow equations involve non-linear and
complex terms. Even in SUC problems which needs to be accurate, they are approx-
imated by DC load flow equations. In a long-term planning, a good approximation
may be sufficient. Therefore, we only enforce the key equations for load flow as
follows.
Optimization Methods on Electricity Generation and Transmission … 15


−M(1 − yht ) ≤ f ht − Sθ ht ≤ M(1 − yht ) (12)

where S is the matrix of reactance of lines and θ ht is the vector of difference of

phase angles at two ends of each line. We misuse the notation in the above inequality
for the notation brevity and simplicity. But it is worth mentioning that it should be
constructed so that we have f i j ht = si j (θi ht − θ j ht ) if a line is installed or exists.
We have observed that when the above inequality is removed, the solution of GTEP
significantly changes and may not be feasible for a real situation.
Storage Constraints—Storages substantially complicate the problem; because it
includes binding constraints that connect power generation of different hours (similar
to rampage constraints). Therefore, although they have become vital in power systems
with high renewable penetration, many studies still do not explicitly formulate them
(Chen et al., 2019). They are very important for technical purposes such as helping
to cover load rampage, stability of power network, and variation of renewables’
power generation. Storages are especially useful when the difference of the maximum
and minimum electricity loads is very high. In this case, there will be enough idle
units to charge storages during off-peak times to be used in pick times. Storages
conventionally are batteries and pumped-storage hydroelectricity. Recently, Power-
to-Gas (PtG) systems are used to produce gas during off-pick times, in order to be
used to generate power when required (Ban et al., 2017; Fallahi & Maghouli, 2020b).

Ih+1,t = Iht − γ dis disht + γ ch chht (13)

Iht ≤ Qst (Qxtst + Q(1 − xtD,st ) (14)

disht ≤ ιdis (Qxtst + Q(1 − xtD,st ) (15)

chht ≤ ιch (Qxtst + Q(1 − xtD,st ) (16)

where γ ch and γ dis are charging and discharging efficiency vectors, respectively.
Also, ιch and ιdis are charge and discharge rate vectors, respectively. Equations (13)
state the storage balance equation between two hours. Constraints (14–16) enforce
inventory, discharging and charging restrictions based on the existing, installed and
decommissioned capacity.
Gas Network—Natural gas is the main fuel used thermal units. The variation of
power generation by renewable sources changes the natural gas consumption of
thermal units and consequently the gas pressure in gas network. This could affect the
natural gas consumption of residential and industrial sectors. Therefore, it is impor-
tant to include the gas equation into GTEP to manage the impact of power system
with high renewable penetration on gas network. Below, we suggest a simplified
variation of gas network modelling. We use a non-vector notation for clarity.
16 M. Noorizadegan and A. Shokri
 
−MG 1 − y iGjt ≤ gi j ht gi j ht − φi j π 2j ht − πi2ht ≤ MG (1 − y iGjt ) (17)

πi ht ≤ π j ht ≤ πi ht (18)

πimin
ht y i jt ≤ πi ht ≤ πi ht y i jt
G max G
(19)

ht y i jt ≤ gi ht ≤ gi ht y i jt
G max G
gimin (20)

piG,min ≤ piGht ≤ piG,max (21)

where φi j is the parameter of natural gas pipeline,  is the compression ratio and
y iGjt = 1 if there exists a pipeline between node i and j, and otherwise, y iGjt = yiGjt (i.e.,
a decision variable). Constraints (17) state relation between gas flow and gas pressure
for new and existing pipelines. Constraints (18) model the impact of compressors on
the gas pressure. Constraints (19 and 20) respectively enforce the pressure and flow
restrictions on new and existing pipelines. Constraints (21) ensure gas production
restrictions on gas production nodes. Adding the above set of inequalities to GTEP
results in a non-linear program. There are various methods such as Newton method
and decomposition-based methods to deal with the nonlinear terms. The reader is
referred to Ding et al. (2018) and Fallahi and Maghouli (2020b) for further topics
on non-linear gas network related terms. Note that in these inequalities, we assume
that the gas flow direction is known in each pipeline. We also neglected modelling
line-pack and installing new compressors.
Balance Equations—Natural gas and power networks have to be separately balanced
at each node

pht + disht − chht + fht

in
− fhtout = dht (22)

G
pht + ght
in
− ght
out
= dht
G
+ ηphtl (23)

where l is the fuel index for natural gas, fhtout is a vector with element f iout
ht presented
asfht = f i ht i , f i ht =
out out out
f . Similarly, we havef in
= f i ht i , f i ht =
in in
  j∈I i j ht   ht
j∈I f ji ht , ght = gi ht i , gi ht = j∈I gi j ht ,ght = gi ht i , and gi ht =
out out out in in in
j∈I g ji ht .
The first balance equation ensures that electricity load at each node is satisfied. The
second balance equation connects the natural gas network to the power network.
The last term in this equation (ηphtl ) is the consumption of natural gas by power
generating units.
Optimization Methods on Electricity Generation and Transmission … 17

2.3 Final Deterministic Model

The summary of this section is a deterministic optimisation model as presented

below:
  f ix opr
min Ftinv xtN , ytN + Ft (xt , yt ) + FtD,U (xt , yt ) + Ft (xt , yt )
t∈T
s.t., (1 − 23)

The above model can be used as a base model for the next stage, which is to
consider uncertain parameters. The above problem has a diagonal structure based
on t. In other words, there is no constraint coupling variables for different t. The
objective function is also separatable based on t. As it will be explained in the solu-
tion method section, decomposition methods can be applied to such a structure. In
some studies (Moradi Sepahvand & Amraee, 2020), reserve and spinning reserve
are included in GTEP models. In security-constrained unit commitment problems,
reserve and spinning reserve are considered to respond unforeseen events such as
demand variations and equipment failure. A simple and practical way of computing
reserve and spinning reserve is to consider a certain fraction of load (Moradi Sepa-
hvand & Amraee, 2020). As reserve and spinning reserve are mainly operational
decisions, we do not independently address them in this model. In the next section,
we study uncertainty in GTEP which are due to two events: net load variation and
equipment.

3 Uncertainty

Electricity demand and power generation by renewable sources are two key sources
of uncertainty in GTEP. Power unit and transmission line failures are also uncertain
events in a power system. These two types of uncertainty are usually dealt with
differently. We briefly review main approaches for both types of uncertainties in this
section.

3.1 Uncertain Electricity Demand and Power Generation

by Renewable Sources

Stochastic programming (Ding et al., 2018) and robust optimisation (Jabr, 2013) are
the common approaches used to deal with power generation of renewable sources
and load uncertainties. The application of stochastic programming involves scenario
generation for possible electricity load and power generation by renewable sources
for the planning horizon. For a GTEP problem, the planning horizon is generally more
18 M. Noorizadegan and A. Shokri

than 15 years. Based on prediction/projection methods, a discrete set of possible real-

isations of each uncertain parameter is generated. Therefore, the number of scenarios
for hourly electricity load and power generation by renewable sources will be expo-
nential. As the first step to reduce the number of scenarios, only selective days are
considered for modelling (e.g., few days per month, or even per season for each
year). Another way of reducing number of scenarios is to merge 24 h of a day into
fewer time blocks. Then, scenario reduction approaches are applied to eliminate less
likely scenarios. However, solving a large multi-stage stochastic problem specially
for practical cases is still very challenging.
Alternatively, robust optimisation takes a less complex but more conservative
approach and plans for the worst cases. The worst cases can be formed prior to the
start of solution procedures. Multivariate statistical analysis based methods such as
“flying-brick” have been developed to deal with variable requirements of the look-
ahead generation capacity, ramping capability, and ramp duration for unit commit-
ment problems. For more details see Pourahmadi et al. (2020). We focus on Interval
Optimisation approach developed for unit commitment problems by Wu et al. (2012).
They used the concept of net load (NL) which is equal to total demand minus wind
generation output minus solar output generation. Net load is commonly used because
wind and solar generation, and demand have some similar characteristics such as
they are non-dispatchable, they depend on the weather condition, and they deviate
from forecasts (Makarov et al., 2010). Therefore, the electricity balance equation is
modified by the concept of net load. The key idea is to make sure that the installed
electricity mix is capable of responding to extreme situations which are illustrated
in Fig. 1. The worst situations are as follow: (1) power generation units including
thermal and hydropower units are able to increase their generation to satisfy the net
load from hour h where the net load is in its lowest level to hour h + 1 where the net
load is in its highest level. (2) power generation units can deal with duck curve from
mid-day towards night peak.
P E O
To this end, we need to define three power generation variable pht , pht ,pht , power
generation for pessimistic, expected and optimistic net loads. Then, the rampage

Fig. 1 Net load uncertainty intervals for a sample daily load

Optimization Methods on Electricity Generation and Transmission … 19

constraints need to be imposed for all combinations of these new variables (e.g.,
P
pht − ph−1,t
O
≤ min{r, αh (Qxt + Q 1 − xtD − pht O
}). When the ramp rate constraints
are revised as explained, it can be ensured that the duck curve is also addressed.

3.2 Uncertain Equipment Failure

Security and resiliency of a power system are usually defined by uncertain equipment
failures which can be due to technical failure, natural disasters, or sabotage. One
solution to deal with unforeseen equipment failures is to allocate a sufficient level
of spinning and non-spinning reserves, which is usually a topic for daily operation
(Morales et al., 2009). Another approach is to set the N-k security criterion. That
means if for any reasons, k equipment (mainly lines) fails at the same time, there
will be no power cut in the power system. This criterion can be imposed locally
with different values for k. As the number of equipment is high in a power system,
contingencies are limited to a pre-defined set of contingency scenarios (Qiming
Chen & McCalley, 2005). Then, binary variables or indicators and a set of related
constraints are used to model the N-k security criterion. This idea is applied within
bi-level programming, multi-stage robust optimisation and multi-stage stochastic
programming (Wu et al., 2016). There are also some probabilistic versions of N-k
security criterion (Sundar et al., 2018). But due to the complexity of probabilistic
constraints, this approach is not popular for GTEP.
As GTEP problems expand existing electricity networks, it may not be necessary
to define a binary variable for each line for the N-k security criterion. In other words,
it is very likely that there are already other routes to a demand bus if one line fails.
Studying the topology of the network could be very useful. Therefore, instead of
initially defining binary variables for each line, cutting plane methods can be used
to ensure the N-k security criterion with much less computational complexity. In a
cutting plane method, the N-k security criterion is first relaxed and the problem is
solved. Then, using a separation algorithm, it is checked to find a violation of the
N-k security criterion for each demand bus. If found, a cutting plane is constructed
to enforce the security criterion for that bus. In general, separation algorithms are
usually quite fast and it is relatively simple to identify violated constraints which
were relaxed (Nemhauser & Wolsey, 1988). For a GTEP, graph-based problems such
as maximum flow problem and shortest path problem could be used in designing
separation algorithms. Therefore, it is expected to achieve a better computational
efficiency in particular for real problems, as significantly less binary variables are
required in the model.
20 M. Noorizadegan and A. Shokri

4 Solution Method

There is a longitudinal study on GTEP in which the majority of them use Benders’
decomposition-based methods to solve their problems (such as Lohmann and Reben-
nack (2016) and Wu et al., (2016)). Therefore, this section reviews some princi-
ples of Benders’ decomposition and few important tips for implementing Benders’
decomposition particularly useful for solving practical problems.
Decision variables in a GTEP problem are naturally divided into strategic deci-
sions and operational decisions. This division paves the way for applying Benders’
decomposition where the investment and operational decisions are made in the master
problem and subproblems, respectively. In particular, Benders’ decomposition is
applied to two or multi-stage stochastic programming or robust optimisation varia-
tions of GTEP. Some studies (Lohmann & Rebennack, 2017) have further explored
the structure of their problem and proposed nested Benders decomposition reformu-
lations. As the operational problems are independent some time intervals, they can be
solved separately. Constraints such as available fuel and maximum amount of pollu-
tion produced by units are usually defined seasonally or annually. These constraints
link operational variables within a season or a year. In these cases, seasonal or annual
operational problems can be independently solved. Such further breakdowns can help
to reduce the computational efforts.
Connecting the master problem and the subproblems is done using optimality
and feasibility cuts. Optimality cuts approximate the impact of the master problem
decisions on the cost of the subproblem (Conejo et al., 2006). It is worth mentioning
that if there are binary or integer decision variables in the subproblem, standard
optimality which are developed for pure linear continuous subproblems cannot be
used. The reason is that optimality cuts are constructed based on the dual form of the
subproblem. The linear programming duality theorem does not hold for an integer
program (Nemhauser & Wolsey, 1988). This is a common mistake in studies about
GTEP problems. Further details can be found in studies about the concept of value
function (Guzelsoy & Ralphs, 2006; Trapp et al., 2013). Feasibility cuts are driven
when a solution of the master problem leads to an infeasible subproblem. When a
subproblem is infeasible for a master problem solution, feasibility cuts driven for that
solution only remove that solution from the search space. Nevertheless, it is possible
that next solutions of the master problem still lead to infeasible subproblems. Signif-
icant computational efforts thus would be spent on finding master problem solutions
with feasible subproblems. It will be computationally beneficial to avoid feasibility
cuts, if possible. Investment decisions e.g., power generation unit installation, are
made in the master problem. For a minimisation problem, in the first iteration, no
new investment is made to keep the master problem cost at its minimum level. This
will lead to some infeasible subproblem. To avoid dealing with feasibility cuts and
their drawbacks, valid inequalities that reflect the subproblem’s feasible region of
subproblems can be derived and added to the master problem prior to the solution
process. For power generation expansion decisions, a constraint can be formed to
Optimization Methods on Electricity Generation and Transmission … 21

enforce the sufficiency of the accumulative capacity to satisfy a selective peak load-
Similar valid inequalities can be formed for transmission lines. Moreover, storage
devices are complicating components of GTEP problems as they are only available
when they are charged. Because the investment cost of storage devices is usually
lower than other technologies, the solution of the master problem may include too
many storage devices at the first iteration. As a result, subproblems may be infeasible.
To overcome this issue, some valid inequalities approximating storage constraints
(charge and discharge processes) are useful in the master problem.
Given the significant number of variables and constraints, it may take several days
to solve a practical problem even with very powerful machines. The feasible region
can be reduced before the solution process by applying methods such as Merit Order
or even simple economic analysis to only include power generation technologies
which are likely to be a part of the optimal solution. In addition, decomposition
methods such as Dantzig Wolfe decomposition and column generation methods have
proved to be very efficient for optimisation problems with binary variables (Singh
et al., 2009). Nevertheless, they have not received much attention for reformulating
and solving GTEP problems (Flores-Quiroz et al., 2016).
A practical GTEP with fair number of details, which addresses concerns for inde-
pendent system operators may not be solvable with a reasonable time. Key aspects
of a power system such as frequency response control, power system inertia, various
types of losses (particularly important for distributed generators) are usually ignored.
As mentioned, gas network and storages are also mainly neglected in GTEP problems.
Simulation-based optimisation is a practical way of addressing all important compo-
nents, factors and constraints of GTEP and at the same time solving the resulting
problem. The application of simulation-based optimisation to GTEP is an indepen-
dent topic with many technical details. Here, we emphasise on its benefit for GTEP
and outline some general steps. For more detail, we refer the reader to Rodgers et al.
(2018).
As it is illustrated in Fig. 2, we can start with a GTEP optimisation model including
only key constraints. The GTEP optimisation model box may contain several sub-
boxes in relation to modelling and solution methods (e.g., cutting planes and decom-
position methods). In particular, with uncertainty assumptions, it is usually the case
that the problem is decomposed into a master problem and several sub-problems.
In the initial optimisation model, complicating components such as the gas network
(i.e., constraints (17–21) and (23)) and N-k security criterion may be ignored. Once
the optimisation model is solved, its solution can be used for a Monte-Carlo simula-
tion model with more details (including the gas network and N-k security criterion)
compared with the initial optimisation model. The aforementioned complicating
components do not directly affect the objective function i.e., there is no term for
these components in the objective function. It is worth noting that the focus here is to
solve a GTEP in which the impact of the gas network is also considered, and we do
not intend to optimise the gas network as well. Therefore, these components affect the
feasibility of the problem. If the solution satisfies all requirements, then the optimal
solution is generated. It is worth mentioning that the notion of optimal solution here
may be challenged. Otherwise, constraints that enforce the violated requirements
22 M. Noorizadegan and A. Shokri

Fig. 2 The diagram for a

simulation-based
optimisation method

are built and added to the optimisation problem. Procedures to generate feedbacking
constraints that impose the conditions of the eliminated components are difficult to
develop and highly depend on the assumptions made on these eliminated compo-
nents. However, there are two advantages of using a simulation-based optimisation
in solving GTEP. First, constraints associated with these components are complex
and maybe non-linear and non-convex if gas network is included. Even if these
constraints have no impact on the optimal solution, they make it very difficult for
solvers to find even a feasible solution. Thus, it would be good to eliminate them
in the initial problem. Second, if these constraints change the optimal solution, the
initial GTEP without these constraints will produce good working solutions for the
complex components. That will help us to define feedback constraints around these
solutions and apply neighbourhood search methods, instead of searching the entire
feasible region of GTEP.
Optimization Methods on Electricity Generation and Transmission … 23

5 Summary

We discussed the key features of a power generation and transmission expansion

planning problem as an optimisation problem. We provided a general framework for
modelling an integrated GTEP and gas network with reasonable features. An approx-
imation of gas network was used to manage the effect of natural gas consumed by
gas-fired units which can be considerably affected by uncertainty of power generation
by renewable sources. We applied a conservative approach (Interval Optimization) to
deal with uncertainty of electricity load and power generation by renewable sources.
This approach is particularly useful as (1) the resulting model is significantly smaller
than models produced by other approaches, and (2) it ensures that in the worst cases
and duck curve, net load will be safely managed. We also discussed equipment fail-
ures as another source of uncertainty and suggested a cutting plane-based method for
N-k security criterion. Finally, solution methods were reviewed. Benders’ decom-
position is by far the most commonly used method. To incorporate more technical
details of GTEP for practical purposes, we suggested a simulation-based optimisation
framework.

References

Allahdadi Mehrabadi, R., Parsa Moghaddam, M., & Sheikh-El-Eslami, M. K. (2020). Generation
expansion planning in multi electricity markets considering environmental impacts. Journal of
Cleaner Production, 243, 118611.
Alolo, M., Azevedo, A., & Kalak, I. E. (2020). The effect of the feed-in-system policy on renewable
energy investments: Evidence from the EU countries. Energy Economics, 92, 104998.
Ban, M., Yu, J., Shahidehpour, M., & Yao, Y. (2017). Integration of power-to-hydrogen in day- head
security-constrained unit commitment with high wind penetration. Journal of Modern Power
Systems and Clean Energy, 5(3), 337–349.
Caunhye, A. M., & Cardin, M. A. (2018). Towards more resilient integrated power grid capacity
expansion: A robust optimization approach with operational flexibility. Energy Economics, 72,
20–34.
Chen, Q., & McCalley, J. D. (2005). Identifying high risk n-k contingencies for online security
assessment. IEEE Transactions on Power Systems, 20(2), 823–834.
Chen, S., Liu, P., & Li, Z. (2019). Multi-regional power generation expansion planning with air
pollutants emission constraints. Renewable and Sustainable Energy Reviews, 112, 382–394.
Coester, A., Hofkes, M. W., & Papyrakis, E. (2018). Economics of renewable energy expansion
and security of supply: A dynamic simulation of the German electricity market. Applied Energy,
231, 1268–1284.
Conejo, B. A. J., Chen, S., & Constante, G. E. (2020). Operations and long-term expansion planning
of natural-gas and power systems: A market perspective. Proceedings of the IEEE, 108(9), 1541–
1557.
Conejo, A.J., Castillo, E., Minguez, R., & Garcia-Bertrand, R. (2006). Decomposition techniques
in mathematical programming, 1st edn. Springer.
Cramton, P., Ockenfels, A., & Stoft, S. (2013). Capacity market fundamentals. Economics of Energy
& Environmental Policy, 2(2), 27–45.
24 M. Noorizadegan and A. Shokri

Denholm, P., O’Connell, M., Brinkman, G., & Jorgenson, J. (2015). Overgeneration from solar
energy in California: A field guide to the duck chart. Technical report, National Renewable
Energy Laboratory, U.S.
Ding, T., Hu, Y., & Bie, Z. (2018). Multi-stage stochastic programming with nonanticipativity
constraints for expansion of combined power and natural gas systems. IEEE Transactions on
Power Systems, 33(1), 317–328.
Esmaili, E., Ghamsari-Yazdel, M., Amjady, N., Chung, C. Y., & Conejo, A. J. (2020). Transmission
expansion planning including TCSCs and SFCLs: A MINLP approach. IEEE Transactions on
Power Systems, 35(6), 4396–4407.
Fallahi, F., & Maghouli, P. (2020a). Integrated unit commitment and natural gas network operational
planning under renewable generation uncertainty. International Journal of Electrical Power &
Energy Systems, 117, 105647.
Fallahi F., & Maghouli, P. (2020b). An efficient solution method for integrated unit commitment
and natural gas network operational scheduling under “duck curve”. International Transactions
on Electrical Energy Systems.
Feldman, D., & Margolis. R. (2019). Q4 2018/q1 2019 solar industry update. Technical report,
National Renewable Energy Laboratory, U.S.
Flores-Quiroz, A., Palma-Behnke, R., Zakeri, G., & Moreno, R. (2016). A column generation
approach for solving generation expansion planning problems with high renewable energy
penetration. Electric Power Systems Research, 136, 232–241.
Fu, R., Feldman, D., & Margolis, R. (2018). U.S. solar photovoltaic system cost benchmark: Q1
2018. Technical report, National Renewable Energy Laboratory, U.S.
Guzelsoy, M., & Ralphs, T. K. (2006). Integer programming duality. Technical report, Lehigh
University Industrial and Systems Engineering.
IEA. (2019). World energy outlook. Technical report. https://www.iea.org/reports/world-energy-
outlook-2019/electricity.
Jabr, R. A. (2013). Robust transmission network expansion planning with uncertain renewable
generation and loads. IEEE Transactions on Power Systems, 28(4), 4558–4567.
Levin, T., Kwon, J., & Botterud, A. (2019). The long-term impacts of carbon and variable renewable
energy policies on electricity markets. Energy Policy, 131, 53–71.
Li, L., & Taeihagh, A. (2020). An in-depth analysis of the evolution of the policy mix for the
sustainable energy transition in China from 1981 to 2020. Applied Energy, 263, 114611.
Lohmann, T., & Rebennack, S. (2017). Tailored benders decomposition for a long-term power
expansion model with short-term demand response. Management Science, 63(6), 2027–2048.
Makarov, Y., Guttromson, R., Huang, Z., Subbarao, K., Etingov, P., Chakrabarti, B., & Ma, J.
(2010). Incorporating wind generation and load forecast uncertainties into power grid operations.
Technical report, Pacific Northwest National Laboratory.
Micheli, G., Vespucci, M.T., Stabile, M., Puglisi, C., & Ramos, A. (2020). A two-stage stochastic
MILP model for generation and transmission expansion planning with high shares of renewables.
Energy Systems.
Moradi Sepahvand, M., Amraee, T. (2020). Hybrid AC/DC transmission expansion planning consid-
ering HVAC to HVDC conversion under renewable penetration. IEEE Transactions on Power
Systems. https://doi.org/10.1109/TPWRS.2020.2988195.
Morales, J. M., Conejo, A. J., & Pérez-Ruiz, J. (2009). Economic valuation of reserves in power
systems with high penetration of wind power. IEEE Transactions on Power Systems, 24(2),
900–910.
Nemhauser, G., & Wolsey, L. (1988). Integer and combinatorial optimization. Wiley.
Newbery, D. M. (2016). Towards a green energy economy? the EU energy union’s transition to
a lowcarbon zero subsidy electricity system—lessons from the UK electricity market reform.
Applied Energy, 179, 1321–1330.
De Oliveira, L.E., Freitas, F.D., da Silva, I.C., & Gomes, P.V. (2017). Dynamic and static transmis-
sion network expansion planning via harmony search and branch & bound on a hybrid algorithm.
Optimization Methods on Electricity Generation and Transmission … 25

In E. Oliveira, J. Gama, Z. Vale, & H. Lopes Cardoso (eds.) Progress in Artificial Intelligence
(pp. 271–282). Cham, Springer International Publishing.
Pourahmadi, F., Hosseini, S. H., Dehghanian, P., Shittu, E., & Fotuhi-Firuzabad, M. (2020).
Uncertainty cost of stochastic producers: Metrics and impacts on power grid flexibility. IEEE
Transactions on Engineering Management, 1–12.
Rodgers, M. D., Coit, D. W., Felder, F. A., & Carlton, A. (2018). Generation expansion planning
considering health and societal damages: A simulation-based optimization approach. Energy,
164, 951–963.
Şen, G., Nil, M., Mamur, H., Doğan, H., Karamolla, M., Karaçor, M., Kuyucuoğlu, F., Yörükeren,
N., & Bhuiyan, M. R. A. (2018). The effect of ambient temperature on electric power generation
in natural gas combined cycle power plant—A case study. Energy Reports, 4, 682–690.
Simo, A., Barbulescu, C., Kilyeni, S., Renghea, S., & Popescu, V. (2015). Dynamic transmission
network expansion planning. case study for 50 buses test power system. In 50th international
universities power engineering conference (UPEC), 1–6.
Singh, K. J., Philpott, A. B., & Wood, R. K. (2009). Dantzig-Wolfe decomposition for solving
multistage stochastic capacity-planning problems. Operations Research, 57(5), 1271–1286.
Sundar, K., Coffrin, C., Nagarajan, H., & Bent, R. (2018). Probabilistic n-k failure-identification
for power systems. Networks, 71(3), 302–321.
Trapp, A. C., Prokopyev, O. A., & Schaefer, A. J. (2013). On a level-set characterization of the
value function of an integer program and its application to stochastic programming. Operations
Research, 61(2), 498–511.
Wu, L., Shahidehpour, M., & Li, Z. (2012). Competition of scenario-based and interval optimization
approaches to stochastic scuc. IEEE Transactions on Power Systems, 27(2), 913–921.
Wu, Z., Zeng, P., & Zhang, X. (2016). Two-stage stochastic dual dynamic programming for transmis-
sion expansion planning with significant renewable generation and n-k criterion. CSEE Journal
of Power and Energy Systems, 2(1), 3–10.
Demand-Driven Electricity Supply
Options of Electric Vehicles: Modelling,
Simulation, and Management Strategy
of Public Charging Stations

Elvin Coban and Gokturk Poyrazoglu

1 Introduction

Electric vehicles (EVs) have gained more attention in recent years as growing global
warming awareness and search for clean energy sources increase (Xiang et al., 2016).
Some countries have already adapted to the increased usage of EVs, for example,
EV sales in the first half of 2020 make up 48% of the total market sales in Norway
(Zachary, 2020). The major automakers state that R&D efforts for gasoline and diesel
vehicles will be ceased by 2025 (Coren, 2018) so structural changes are expected to
be achieved by using EVs (Ballentin et al., 2011). Especially, a detailed simulation
showed the reduction of CO2 emissions due to substituting the conventional vehi-
cles with EVs (Teixeira & Sodré, 2016). However, this transition from conventional
vehicles to EVs raises many new questions, and the most critical among them is
about how to handle this demand-driven electricity supply efficiently, specifically,
charging of EVs.
There are three main types of EVs: battery EVs (BEVs), plug-in hybrid EVs
(PHEVs), and hybrid EVs (HEVs) (Types of Electric Vehicles, 2020). BEVs
comprising 67% of the world’s electric car fleet in 2019 are fully-electric vehicles
with rechargeable batteries and no gasoline engine, such as Tesla Model 3, whereas,
PHEVs can recharge the battery through both regenerative braking and plugging
in to an external electrical power source, such as Mercedes C350e. Compared to
HEVs, PHEVs have approximately 20 times longer traveling distance before their
gas engines start providing assistance. Lastly, HEVs are powered via both gasoline
and electricity where the car’s own braking system provides energy to recharge the

E. Coban (B)
Industrial Engineering, Ozyegin University, Istanbul, Turkey
e-mail: [email protected]
G. Poyrazoglu
Electrical & Electronics Engineering, Ozyegin University, Istanbul, Turkey
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 27

A. B. Dorsman et al. (eds.), Applied Operations Research and Financial Modelling
in Energy, https://doi.org/10.1007/978-3-030-84981-8_3
28 E. Coban and G. Poyrazoglu

battery, such as Toyota Prius Hybrid. 2.1 million new electric cars were sold in 2019
as reported in the International Energy Agency (IEA) Global EV Outlook 2020 since
consumers choose a variety of climate-conscious, energy-efficient EVs (IEA, 2020).
However, there still remains an important problem as all EVs have a relatively short
driving range limited by insufficient battery technologies. Additionally, few charging
bases are unevenly distributed (Poch et al., 2010). Thus, the adoption of EVs faces
low availability of charging infrastructure (Wiederer & Philip, 2010).
There are three major categories of charging units and different categories deter-
mine both the charging times and the car types that can be charged (Office of Energy
Efficiency & Renewable Energy, 2020). Thus, investment supporting the charging
infrastructure, especially computing the location of CSs and the types of charging
units, is very critical. The total number of chargers were approximately 7.3 million
worldwide in 2019, 2.1 million more than in 2018. Even if slow chargers seem to
dominate the market, the fast chargers are slowly increasing in popularity (The Global
Electric Vehicle Market in, 2020: Statistics Forecasts, 2020) as DC fast charger adds
60 to 80 miles of range per 20 min of charging (Office of Energy Efficiency &
Renewable Energy, 2020).
The unavailability of large-scale charging infrastructure delays the adoption of
EVs. As there are few EVs on the road, charging service providers delay investing
heavily in constructing the infrastructure. Additionally, efficiency and innovation in
EV charging operations are vital (Shen et al., 2019). It is estimated that a total of
300–350 million EVs will be on the road in 2040 and 350 million CSs will be needed
to support this system (Outlook, 2019). For example, 10 million EVs are expected
to be on the roads by 2040, and 10 million CSs are required to support these EVs in
Turkey. Based on a preliminary analysis, public CSs will comprise 60% due to the
absence of parking lots suitable for home use and stronger electrical infrastructure
at public CSs in Turkey (Shura Energy Transition Center, 2019).
In addition to the infrastructure of CSs, electricity pricing policies adopted
by countries also affect the number of EVs. For example, China and the U.S.
reduced purchase subsidies substantially in 2019, whereas, several countries, such as
Germany, Italy, and France, introduced new and/or strengthened incentives. China
has extended its subsidy scheme to 2022 at a reduced pace and has also urged cities
to loosen the quotas for car permits in order to stimulate the auto industry (Covid-19
Response Will Influence the Pace of Transition to Electric Vehicles, 2020). Addi-
tionally, EVs have an initial higher purchase-price than internal combustion engine
vehicles, however, they have lower running and maintenance costs. The average price
in the U.S. to charge an EV is computed as $1.20 per gallon on an equivalent basis in
2018 and even if the gasoline prices stay low, the sales of EVs have been increasing
since 2015 (IEA, 2020). All of these policies and/or incentives affect the pace of EV
adoption.
Motivated by the possibility of designing the charging infrastructure in advance,
we study public CSs with some of the key investment decisions related to CSs: where
to locate the CSs and how to compute the required number and type of charging
units once the CS is located. We explain an existing mathematical model to locate
the CSs and summarize our findings for the proposed backup coverage extension.
Demand-Driven Electricity Supply Options of Electric Vehicles … 29

Once the optimal locations are computed via the mathematical model, we explain a
simulation framework to compute the required number and type of charging units
satisfying some performance metrics, such as utilization and profit ratio. We capture
the stochasticity of EV arrivals by utilizing Google Popular Times (Google, 2020). To
the best of our knowledge, this chapter is the first to summarize some of the strategic,
tactical, and operational problems of public CSs via both mathematical modelling and
simulation, with an additional discussion of pricing policies. The relevant literature
is summarized in Sect. 2, whereas, we explain a CS location mathematical model
and its possible extensions in Sect. 3. In Sect. 4, simulation developed to compute
the number and the types of charging units for a given location is introduced with its
numerical analysis. Then, we discuss different pricing policies and open problems
in Sect. 5. Finally, conclusion is in Sect. 6.

2 Related Work

We summarize the related work under two sections. First, we explain the current
condition of EV adoption in five countries with the highest cumulative EV sales
in Sect. 2.1. Then, we summarize the relevant location and simulation literature in
Sect. 2.2.

2.1 Worldwide Current Condition

The popularity of EVs keeps growing around the world due to decreasing production
costs, increasing deployed charging infrastructure, and the recognition of the urgency
to reduce the transport sector’s climate emissions. China has the largest EV market
as Europe has the second-largest EV market, ahead of the U.S. EV progress varies
a lot within Europe and understanding this variation combined with the countries’
effective policies will increase the pace of adoption of EVs.
We summarize the current EV condition of China, the U.S., Norway, Germany,
and the United Kingdom representing the best performers in terms of cumulative
sales.
China: EV sales of China are reported as 1.15 million in 2019. Stringent restric-
tions, such as Beijing issuing only 10,000 permits for the registration of internal
combustion engine vehicles per month, are applied to accelerate switching to EVs.
Tax exemptions are also applied while purchasing an EV. Additional exemptions are
also introduced whenever the market slows down, such as a 10% service tax waiver
for EV purchase was applied in April 2020 to boost the EV market during the Covid-
19 outbreak (Mordor Intelligence, 2020). However, China has also added additional
criteria for the subsidies, for example, an EV purchased by a commercial fleet user
would be given subsidies only if its aggregated driving range exceeded 20,000 km
per annum (Mobility Foresights, 2020). In 2019, over 516,000 public EV charging
30 E. Coban and G. Poyrazoglu

poles are reported, an increase from around 387,000 stations in the previous year
(Total Number of Public Electric Vehicle Charging Stations in China, 2015).
U.S.: 1.18 million EVs were on the road in 2019 and more than 18 million EVs are
expected in 2025. To supply adequate charging, more than 25,000 CSs are operating
in 2020 having more than 68,800 Level 2 and DC fast charging units (Number
of Public Electric Vehicle Charging Stations & Charging Outlets in the U.S. as of
March 3, 2020). Approximately 10,860 units were DC fast chargers making long-
distance travel more practical for EVs (Office of Energy Efficiency & Renewable
Energy, 2020). There are various types of incentives provided to EV users, such as
income tax credit, sales tax exemption, purchase rebate, and high-occupancy vehicle
lane (Zhang et al., 2014). Similar to China, the U.S. also reduced these incentives
due to increasing sales of EVs. States also differentiate a lot with respect to the
additional laws and incentives provided. For example, California provides rebates,
rates, charging infrastructure, and vehicle-grid integration technologies for EV users
to achieve their zero-emission vehicle adoption (C. P. U. Commission, 2020).
Norway: 58% of 2019 eV sales share in Europe belongs to Norway aiming its
goal of all zero-emission vehicle sales by 2025. The cumulative EV sales of Norway
are reported as 334,000 while the number of public regular and DC fast charge points
is 10,337 and 3,426, respectively (Hall et al., 2019). According to EV registrations,
BEVs represent 76% of 2019 sales. There is 25% VAT and import taxes’ exemption
for BEVs, whereas, BEVs are also allowed to travel in bus lanes and offered reduced
or free tunnel and ferry tolls.
Germany: 3% of 2019 eV sales share in Europe belongs to Germany, Europe’s
largest vehicle market. The cumulative EV sales of Germany are reported as 292,000
while the number of public regular and DC fast charge points is 47,832 and 4,169,
respectively (Hall et al., 2019). Similar to the U.S., cheap gasoline inflates the payback
period for EVs in Germany. With additional government incentives, the Germany EV
market managed to comprise a 3% EV sales share in 2019 (IEA, 2020). A financial
subsidy up to 3,000 is provided by the federal government for an EV in addition
to exemption from the annual circulation tax for BEVs and fuel cell vehicles, and
tax benefits for electric company cars. The government decided to extend the bonus
through 2025, with the bonus doubled until January 2022 as a response to the Covid-
19 outbreak (Hall et al., 2019). According to EV registrations, BEVs represents 58%
of 2019 sales.
United Kingdom: 3.2% of 2019 EV sales share in Europe belongs to the United
Kingdom, Europe’s third largest EV market by sales volume. The cumulative EV
sales of the United Kingdom is reported as 270,700 while the number of public
regular and DC fast charge points is 22,359 and 4,735, respectively (Hall et al., 2019).
Plug-in car grants up to 3,000 and discounts on ownership taxes is provided by the
government to support the EV adoption in addition to incentives for the buildout of
public and private CSs. BEVs and PHEVs evenly share the market.
To sum up, China remains the world’s largest EV market, followed by Europe
and the U.S. (The Global Electric Vehicle Market in, 2020: Statistics Forecasts,
2020). The following table summarizes the cumulative EV sales, the total number
of charging units, and the ratio of cumulative sales to the total number of charging
Demand-Driven Electricity Supply Options of Electric Vehicles … 31

Table 1 Summary of the EV

Country Cumulative sales Total # of Ratio
condition of China, the U.S.,
charging units
Norway, Germany, and the
United Kingdom (Hall et al., China 3,500,000 516,000 6.8
2019) U.S 1,400,000 68,800 20.3
Norway 334,000 13,763 24.3
Germany 292,000 52,001 5.6
United Kingdom 270,700 27,094 10.0

units representing the availability of the infrastructure. Even if charging units are not
differentiated as fast or slow in Table 1, the computed ratio reflects the investment in
charging infrastructure. Additionally, this ratio may indicate each country’s future
tendency in adopting EVs, such as Norway has a goal of all-zero emission vehicle
sales by 2025 (Hall et al., 2019).
To achieve quick and sustainable adoption of EVs, both financial and nonfinancial
incentives are critical (Zhang et al., 2014). As EV sales increase, investment in CSs
should increase. Especially fast chargers are important as they can increase the range
an EV can travel. Even if the total number of fast chargers is much lower than the
slow chargers, the number of fast chargers is expected to increase more than the
slow chargers. Additionally, vehicle-to-grid (V2G) technology enabling electricity
to be returned to the grid from EV batteries, the same way stationary storages are
connected to the grid, will increase the popularity of EVs. For example, the European
standard for V2G charging will be ready in late 2020 (The Global Electric Vehicle
Market in, 2020: Statistics Forecasts, 2020), whereas there are successful adoptions
in the U.S. to balance the grid of the energy markets via V2G, such as V2G company
Nuvve’s participation at California (Nuvve, 2020).

2.2 Location and Simulation of CSs

Location decisions should be made carefully when a logistics system is started from
scratch. In other words, a new type of service, charging EVs, will be provided
through the network of CSs. Location decisions may be strategic or tactical resulting
in sizeable investments. Changing sites and equipment is usually unlikely in the
short or medium term if the facility is purchased or built specifically for the newly
offered service. However, if space and equipment are rented, location decisions can
be updated in the medium term. It is also important to note that location decisions
may affect demand. For example, if a CS is located far away from the popular demand
locations, EVs may need to travel extra distance to reach the CS that will decrease
the range of the EV coverage (Ghiani et al., 2004).
In order to locate a set of CSs, the demand should be estimated. GPS trajectory
data collected from the electric taxi fleet and electric buses are used in Shahraki et al.
(2015), Sun et al., (2020). However, this is not possible for private vehicles due to
32 E. Coban and G. Poyrazoglu

their limited private vehicles GPS data. Thus, traffic demand data between origin–
destination pairs are used to generate the charging demand. It is important to note that
determining the whole trip paths and charging plans for EV drivers simultaneously
results in time-saving compared to computing EV trip paths and charging plans
separately (He et al., 2018).
A CS location problem is studied via a mathematical model minimizing the total
cost in Zhu et al. (2016). However, the assumption of enough power for the EVs
ensuring that drivers can reach their destination-regions (or the CSs) from their
origin-regions (e.g., residences) is limiting. Moreover, another CS location problem
is studied as a maximal covering problem in Lisbon forecasting the charging demand
by regression using census data (such as age) (Frade et al., 2011). Similarly, locations
of scooter recharging stations are computed via an integer program (Wang, 2007),
whereas the objective of total access cost as a function of walk distance between
zones is minimized to compute the locations of CSs in Seattle, Washington (Chen
et al., 2013). However, time is not considered during the proposed mathematical
model developed in Chen et al. (2013). A similar CS location problem is solved in
Arayici and Poyrazoglu (2019) minimizing the total cost comprising distance-based
costs and fixed costs. The authors also consider time and remaining charge levels,
however, they also did not differentiate between EVs and types of charging sockets.
Once the CS location is computed, the next step is to manage the tactical and/or
operational levels decisions, such as deciding the number and the types of charging
units and the queuing discipline of EVs at the CS. Immediate EV charging without
any waiting time and/or fully charged EVs when leaving the CS are studied more
often in the literature. For example, commercial car-sharing is studied via a control
software for fully charging EV fleet in a smart grid architect (Lützenberger et al.,
2014). However, it is very unlikely that an EV can always start to get charged as
soon as it arrives at the CS. To capture this waiting time, an M/M/1 queue is modeled
with one charging unit (Keskin et al., 2019). Simulation of a large vehicle fleet is
developed on the basis of individual driving profiles to analyze user mobility and
grid support (Teixeira & Sodré, 2016). The authors show that coordinating charging
loads provide flexibility to shift loads without limiting mobility, hence avoiding
load fluctuations. Another simulation study is developed in Luo et al. (2020) to
schedule charging for three types of EVs considering the convenience of drivers, the
performance of the transport system, and the distribution network. Only one type of
charging (DC fast) is studied in addition to quick battery swaps. Since the strategy
is to reduce the longest waiting time and the number of EVs in the queues, they
distribute EVs evenly to charging units. A mathematical model is proposed to satisfy
the EV charging demand in a highway CS based on a fluid dynamic traffic model
and the M/M/s queuing theory (Bae & Kwasinski, 2011).
One of the key concerns about the studies in the literature is long historical EV
usage data unavailability. Additionally, how to generate such data is also ambiguous
(Arayici & Poyrazoglu, 2019). There are studies that consider the arrivals and depar-
tures as random and utilize a probability density function to model vehicle mobility
(Li & Zhang, 2015; Pflaum et al., 2018; Rezaee et al., 2013).
Demand-Driven Electricity Supply Options of Electric Vehicles … 33

The purpose of this chapter is to provide guidance for investment decisions about
public CSs: where to locate the CSs, how many charging units are required, and
which types of charging units are needed. To answer these questions, we first explain
an existing mathematical model to locate the CSs and discuss possible extensions.
Once the locations of CSs are known, then a simulation framework of a public CS
with multiple charging units, random EV arrivals and departures, and random initial
state-of-charges (SoC) is explained. Due to a lack of real data, we utilize Google
Popular Times (Google, 2020) to generate random arrivals and departures. In other
words, this chapter provides some of the strategic, tactical, and operational problems
of public CSs with additional discussion of pricing policies.

3 Locating Charging Stations

Locating CSs properly is critical for the sustainability of EV usage. This strategic
and/or tactical level decision forces charging companies to locate their CSs with
optimization tools. We select a location model that was introduced in Arayici and
Poyrazoglu (2019) as our base model and discuss further possible extensions in this
section. The nomenclature is represented in Table 2.
The mathematical model proposed in Arayici and Poyrazoglu (2019) is as follows.
 
min dv,t, j xv,t, j + C j yj (1)
v,t, j j

s.t. xv,t, j ≤ 1 ∀v ∈ V, t ∈ T (2)
j

xv,t, j ≥ 1 ∀v ∈ V (3)
t, j

xv,t, j ≤ y j ∀v ∈ V, t ∈ T, j ∈ J (4)


xv,t, j ≤ Q j y j ∀ j ∈ J, t ∈ T (5)
v

yj ≥ p (6)
j
 
Rv,t − Nv,t + av,t − E v, j,t xv,t, j ≥ 0 ∀v ∈ V, t ∈ T (7)
j

j (E v, j,t x v,t, j )
Rv,t + av,t − ≤ Rmax ∀v ∈ V, t ∈ T (8)
2
34 E. Coban and G. Poyrazoglu

Table 2 Nomenclature
Indices
v Electric vehicle (EV), v ∈ V
t Time, t ∈ T
j Charging location, j ∈ J
Sets
V Set of EVs, {1, 2, …, V }
T Set of time, {morning, noon, afternoon}
J Set of possible charging locations, {1, 2, … J }
Parameters
dv,t, j Distance between EV v and location j at time t
Qj Charging capacity of CS at location j
Cj Fixed cost of deploying a CS at location j
p Lower bound on the number of deployed CSs
Rinitial Initial charge of EVs
Rmax Maximum capacity of EV charge
E v, j,t Additional total distance between the CS at location j and EV v’s location at
time t (distance from EV location to CS location and from CS location to
EV location)
Nv,t Distance between the current location of EV v and the next location at time t
Decision variables
xv,t, j 1 if EV v is charged at CS at location j during time t; 0 otherwise
yj 1 if a CS is deployed at location j
Rv,t EV v’s state of charge at time t
av,t Charging amount of EV v at time t

av,t ≥ 0 ∀v ∈ V, t ∈ T (9)


av,t ≤ xv,t, j Rmax ∀v ∈ V, t ∈ T (10)
j

Rv,t+1 = Rv,t − Nv,t + av,t − E v, j,t xv,t, j ∀v ∈ V, t ∈ T : t ≤ |T | − 1 (11)
j

Rv,1 = Rinitial ∀v ∈ V (12)

Rv,t ≥ dv, j,t xv,t, j ∀v ∈ V, t ∈ T, j ∈ J (13)

xv,t, j ∈ {0, 1} ∀v ∈ V, t ∈ T, j ∈ J (14)

Demand-Driven Electricity Supply Options of Electric Vehicles … 35

y j ∈ {0, 1} ∀ j ∈ J (15)

The objective function is to minimize the total cost due to the total distance
traveled by all EVs and the fixed costs of locating CSs. Constraint (2) forces an EV
to be in at most one of the CSs for all time periods, whereas Constraint (3) states
each EV should be charged at least once during a day. Constraint (4) links x and
y binary variables by stating that an EV can only be charged at location j if the
corresponding CS is deployed at location j. The capacity of each operational CS
is satisfied by Constraint (5). The lower bound on the number of deployed CSs is
at least p by Constraint (6). Constraint (7) states that an EV should go to its next
destination with the current amount of charge, and if it is required to go to a CS that
is not on its way, the remaining distance should be considered, whereas the state of
charge being greater than the maximum charge capacity is prevented by Constraint
(8). Constraint (9) defines the nonnegativity of decision variable a. Constraint (10)
prevents each EV to be charged more than the maximum capacity. The balance of
charging between consecutive times is defined by Constraint (11) that is formulated
similar to inventory balance equations: the state of charge or the remaining range
must be at least enough for the distance to the next destination of the EV. Initial
charges of EVs are set by Constraint (12), whereas Constraint (13) states that the
EV range should not be smaller than the distance to a CS in order to prevent any
empty charge situation. Constraints (14) and (15) define x and y decision variables
as binary variables.
This model is solved for a small example in Arayici and Poyrazoglu (2019) to
compute which CSs would be used to charge each EV during a day. We introduce
backup coverage for CSs as an extension to this model considering that a backup
coverage may be necessary if the assigned CS cannot operate, such as in case of
maintenance. One possible way to incorporate a backup CS is as follows.

yj ≤ w j, j  ∀ j ∈ J : j = j  (16)
j  ∈J j

w j, j  ≤ y j  ∀ j, j  ∈ J : j = j  (17)

where wj,j’ is a binary decision variable that equals 1 if a CS that is located at j’ is

backup CS for the CS that is located at j; 0 otherwise, and J j is the set of locations
whose distance from the CS that is located at j is less than a threshold distance. We
force at least one backup coverage (an additional CS coverage) with Constraint (16)
and Constraint (17) states that backup coverage (j’) is possible only a CS is located at
location j’. We solve the same instance introduced in Arayici and Poyrazoglu (2019)
with varying threshold distance values. Table 3 shows the distances between each
candidate CS (indexed by j in the mathematical model). When the threshold value is
60 km (or higher) for the backup coverage, two CSs are deployed at locations 4 and
5 at the optimal solution, whereas, three CSs are deployed at locations 3, 4, and 5
when the threshold for backup coverage is 50 km. Further decreasing the threshold
 36

Table 3 Distance between candidate CSs where distance values in bold show that a backup coverage of a CS at the row is satisfied via a CS that is deployed at
the column
When the threshold distance is 60 km When the threshold distance is 50 km
j 1 2 3 4 5 j 1 2 3 4 5
1 0 20 35 60 85 1 0 20 35 60 85
2 20 0 15 40 65 2 20 0 15 40 65
3 35 15 0 35 50 3 35 15 0 35 50
4 60 40 35 0 60 4 60 40 35 0 60
5 85 65 50 60 0 5 85 65 50 60 0
E. Coban and G. Poyrazoglu
Demand-Driven Electricity Supply Options of Electric Vehicles … 37

value results in infeasibility as it becomes impossible to assign a backup CS for all

candidate CSs. Additionally, the computation time is still less than a second for the
backup coverage extension.
There are other possible extensions that can be studied for this model. The first
limiting assumption is EVs’ locations that are fixed at each time interval as the
authors are motivated by the traffic congestion data. However, in real life drivers
can deviate from their first planned paths if they realize congestion. Hence, one can
introduce more nodes of which some are defined as origins and destinations and a
routing problem can be solved while computing the locations of CSs. Additionally,
a day can be divided into more time intervals than the ones studied in the model.
More vehicle types and charging types (including the times to charge etc.) can be
introduced to mimic a real setting. Note that different charging types are necessary
due to varying technical features of EVs and these technical features also determine
the range that an EV can travel. Initial charge levels can be introduced for each vehicle
separately, whereas, the weights used in the objective function comprising two types
of costs, both distance-based and fixed, should be carefully computed to capture the
relative importance of these cost types. Since deploying a CS is a strategic and/or
tactical level decision, all CSs should be carefully planned in advance.
One may also enforce a lower bound on the number of backup coverages via a new
constraint or reward additional backup coverages via an updated objective function.
Even the backup coverage constraints can be modified such that backup coverage
constraints are only valid for the deployed CSs. Lastly, the objective function can
also be updated based on the target(s) of the charging business. For instance, one can
also minimize the maximum distance covered as in p-center mathematical model
(Daskin, 2011).

4 Simulation Modeling

After locating the CSs, we develop a simulation framework for each CS computing the
types and numbers of charging units via analyzing various performance metrics, such
as the total cost, utilization, and average waiting time of EVs (Poyrazoglu & Coban,
2021). There are six main modules in the simulation framework as represented in the
flowchart in Fig. 1: (i) Event generator, (ii) Vehicle features generator, (iii) Incoming
vehicle mobility, (iv) Outbound vehicle mobility, (v) Queue tracking module, and
(vi) Valet service (Optional).
Our first module, the event generator, generates an event belonging to one of the
following types: the arrival of a vehicle, the departure of a vehicle from a charging
unit, and reneging (i.e., leaving queue after joining) based on the number of vehicles
in the system and the dynamically changing probability density function. The simu-
lation framework tracks the time of each event and the current number of vehicles in
the system. For instance, if there is no vehicle in the queue, then the reneging event
cannot be generated as the next event. The event generation module keeps generating
38 E. Coban and G. Poyrazoglu

Fig. 1 Flowchart of the simulation platform

new events until the closing time of the shopping mall. Records, such as the arrival
and departure times of EVs are held with each event.
If the next event is the new arrival of an EV, the vehicle features generator module
creates the brand and the current state-of-charge (SoC) of the vehicle at the time of
arrival. Battery SoC is a value between 0 and 100%. Since vehicles arrive at its
destination after mobility, the probability to have nearly 100% SoC is very low
compared to other values. Therefore, the vehicle’s SoC is determined by a beta
Demand-Driven Electricity Supply Options of Electric Vehicles … 39

distribution. Additionally, there are three types of charging units: (i) AC charger, (ii)
regular DC charger (24 kW DC), and (iii) fast DC charger (50 kW DC). The priority
list of chargers for each vehicle sorts these three types from the most preferred to the
least preferred. The brand of an EV determines energy transfer limits due to varying
technical characteristics. The duration of stay for charging is not only dependent on
the SoC but also on the technical limitations of on-board converter of the EV and the
type of the charging unit.
The incoming vehicle mobility module is called after the vehicle features generator
module. First, the charging decision is made by a coin flip. If the decision is not
charging the vehicle, the corresponding vehicle is directed to the regular parking
lot. If the decision is to charge the vehicle, then we first check whether there is any
idle charging unit. If all charging units are busy, then the vehicle is directed either
to queue or to the regular parking lot again based on a coin flip. In case of any idle
charging units, the charging priority list of the vehicle will be checked and the most
preferred available charging unit will be picked to start charging the EV.
If the next generated event is a departure of a vehicle from a charging unit, the
outbound vehicle mobility module is called. One of the EVs will depart based on the
type of the departure scenario: (i) random selection, (ii) first-come-first-leave, (iii)
select from a beta distribution representing the longer the parking time of an EV,
the higher the chance of its departure. Since a charging unit becomes available, we
check whether there is any vehicle in the queue that can be charged at the newly
idle charging unit. We define different scenarios to choose an EV in case of multiple
EVs in the queue: (i) selecting an EV randomly, (ii) selecting the EV that waits for
the longest to be charged, and (iii) selecting an EV based on the beta distribution
assigning a higher probability on the EVs waiting longer than the others.
In the case of a reneging event generated as the next event, an EV leaves the
queue without being charged. We study three cases for reneging: (i) the EV is chosen
randomly, (ii) when the vehicle that waits the longest reneges, and (iii) when the
EV is chosen based on a beta distribution that assigns a higher probability to longer
waiting times.
Additionally, a valet service module is an optional module in the proposed simu-
lation framework. An EV whose charging is finished as its maximum SoC is reached
may not be removed from a charging unit if its departure is not generated. In other
words, the driver of an EV may not return back to his/her car immediately when the
charging is over. We assume the price of charging unit occupancy to be the same
as the charging cost per minute. If the EV is not removed even after its charging
is finished, it will be charged an additional amount for keeping the corresponding
parking lot busy. Thus, we propose a valet service module by which such EVs will
be removed from the charging units by a valet and another EV (that is either in the
queue or newly arrived) can start charging.
A technical feature about maximum charging percentage is also considered as
e-mobility professionals claim that 80% SoC is indeed the full capacity of an EV.
Hence, charging up to 100% is not necessary. In the valet service module, the valet
can control the maximum level of charging defined by two scenarios: (i) up to 80%
SoC and (ii) up to 100% SoC.
40 E. Coban and G. Poyrazoglu

The charging units comprise of three charging sockets that are categorized by
their input electric current as AC or DC, and by their speed of power transfer as slow
or fast. Even if the AC charging sockets may vary based on the maximum capacity,
off-the-shelf 22 kW AC chargers are available and their $500–1000 price range is
acceptable for most EV users. Hence, 22 kW charging sockets are considered as AC
charging sockets in this study. The DC charging sockets have a wider range of costs
due to their power capacity. The general standard of power levels is 24 and 50 kW in
DC charging sockets, but up to 350 kW charging socket is also possible. In this study,
we will consider only 24 and 50 kW DC charging sockets in addition to the 22 kW AC
charging sockets. The electrical energy consumption per vehicle and the daily energy
consumption in the charging units are computed based on the charging starting time
and the charging socket type. The electrical energy consumption E i,j is defined for
EV j and charging socket i where maximum power capacity of the charging socket
in kW, (Pi max ), AC/DC converter capacity in the vehicle in kW, (C j max ), charge start
time, (t a j ), vehicle departure time, (t d j ) (note that the vehicle departure time may
differ from the charging end time), SoC of the battery at the start of occupancy in
kWh, (S j 0 ), and the vehicle maximum battery capacity in kWh, (B max j ) are defined.
If the charging socket is AC, E i,j is computed by Eq. 18.
⎧ ⎫
⎨ min Pimax , C max
j t dj − t aj ⎬
E i, j = min , B max − S 0j (18)
⎩ 60 j
⎭

If the charging socket is DC, E i,j is computed by Eq. 19.

Pimax t dj − t aj
E i, j = min{ , B max
j − S 0j } (19)
60
The daily electricity consumption E tot is the total energy provided to charged EVs
represented in Eq. 20.

E tot = E i, j (20)
i, j

4.1 Numerical Results

To perceive real-life settings, we consider a CS located at a shopping mall operating

between 10 am and 10 pm in Turkey. All parked vehicles are enforced to leave the
shopping mall at 10 pm. The random arrivals and departures are generated according
to Google Popular Times to mimic the trend of people’s presence at that location as
Google aggregates and anonymizes its users’ Google Location History data (Google,
Demand-Driven Electricity Supply Options of Electric Vehicles … 41

Table 4 The technical details of the selected EVs

Brand & version Battery capacity Charge protocol Internal AC/DC Charging socket
(kWh) inverter capacity priority (highest
(kW) to lowest)
Renault ZOE V1 22 AC 22 kW 22 AC
Nissan LEAF V2 39 DC CHAdeMO 6.6 22DC-50DC-AC
BMW i3 V2 33 DC Combo 11 AC-22DC-50DC
VW eGolf V2 36 DC Combo 7.2 AC-22DC-50DC
Hyundai Ioniq 28 DC Combo 3.3 AC-22DC-50DC
KIA Soul V1/V2 30 DC CHAdeMO 3.7 AC-22DC-50DC
Jaguar I-Pace 90 DC Combo 7.4 50DC-22DC-AC
TOGG SUVa 90 DC Combo 11 50DC-22DC-AC
a Assumptions based on related news

2020). We consider eight types of EVs with their corresponding technical features
represented in Table 4.
We assume Poisson distribution for vehicle arrivals. Based on the shopping mall’s
Google Popular Times’ data, the rate of arrivals per hour and the number of arrivals
per day are varying. For example, the number of arrivals during the weekends is
much higher than the number of arrivals during weekdays. The event generator
module uses the shopping mall’s Google Popular Times’ data for Sundays and fits a
Poisson distribution for every time period (i.e., an hour). Then, the calculated rate of
arrival per hour (λ) is used as an indicator to create the possibility of an arrival event.
Moreover, for each event generation, the possibility of a departure is also calculated
by following the current number of vehicles in the parking lot. Beta distribution (when
α = 2 and β = 5) is utilized during the incoming and outbound vehicle mobility
modules (i.e., during queuing, departure, reneging) assigning higher probabilities to
the EVs staying longer according to the event. Additionally, each vehicle’s SoC is
also determined by a beta distribution function of α = 3.7 and β = 5. We simulate
the case of a valet service module during which an up-to 100% maximum charging
scenario is selected. We simulate the number of charging units between 1 and 7 in the
increments of 1 and consider each charging unit has 3 sockets: (i) one AC, (ii) one
DC fast, and (iii) one DC regular. In other words, our simulation results are for the
number of sockets between 3 and 21 in the increments of 3. The statistics of the total
number of EV arrivals are shown for varying number of charging sockets in Table
5 when the number of replications is 100. The shopping mall realizes between 332
and 439 eV arrivals with an average of 384.4 EVs over all simulation replications.
Our performance metrics are average waiting time in queue, utilization of the
charging units, the average utilization of CS, profit ratio, happy customers, hourly
energy consumption, and instantaneous power requirement. Our goal is to choose the
reasonable and acceptable number and the types of charging units while capturing
the uncertainty involved in the EV arrivals and departures.
42 E. Coban and G. Poyrazoglu

Table 5 The statistics of the total number of EV arrivals with respect to varying number of sockets
when the number of replications is 100
Charging sockets Ave St. dev Max Min
3 387.0 18.0 434 339
6 384.5 19.2 438 335
9 389.5 18.0 435 338
12 382.9 19.0 439 336
15 382.1 19.9 432 334
18 383.0 20.6 426 332
21 381.8 19.4 429 342

The average waiting time decreases nonlinearly as the number of charging sockets
increases. The simulation results show that the median waiting time for a customer
is 33 min if only one charging unit (3 sockets) is deployed with the upper adjacent
value (UAV) of 107 min. The median waiting time reduces by 81% to 9.27 min
if two charging units (6 sockets) are deployed. It reduces to 5.31, 4.21, and 3.51
for three, four, and five charging units, respectively. The UAV reduces by 70% to
33.97 min if two charging units (6 sockets) are deployed. If we increase the number of
charging units to three, four, and five, the UAVs equal to 20.11, 14.50, and 12.07 min,
respectively. Similar to the average waiting time, the average utilization of the CS
decreases as the number of sockets increases. For example, the average utilization
equals 94.8% for 3 sockets but reduces to 77.9% for 15 sockets. Note that each EV
type has a preference list for the charging sockets. The average utilization equals
90.8% when there is one charging unit. As the number of charging units increases
from two to seven with increments of one, the average utilizations are 83%, 76%,
70%, 65%, 59%, and 54%, respectively.
A charging business generates revenue by setting a price for charging, such as price
per minute, and has a liability to pay the cost of electricity based on a tariff agreement
with the electricity retailer. Even there is no fixed cost, such as the investment in our
study, we can still compute a ratio of profit margin by dividing the revenue by the
cost of electricity and subtracting one from the resultant ratio. The simulation results
show that the profit ratio varies between 173 and 207% and increases as the number of
sockets increases as represented in Fig. 2a. The missed opportunity for the business
appears when there is not enough socket to satisfy the customers at the time of arrival,
but that does not necessarily mean that every new socket has the same financial impact
as the previous sockets.
Another important metric is happy customers ratio measuring the percentage of
customers who were willing to charge their vehicle and succeeded represented in
Fig. 2b. As the number of sockets increases, the happy customers ratio increases as
expected. If the number of charging units equals 1, the median of the happy customers
ratio is 34.4%, whereas the ratio increases over 85% when five charging units are
deployed as in Fig. 2b. Additionally, the happy customers ratio increases nonlinearly
as the number of sockets increases. The ratio is over 95% when the number of
Demand-Driven Electricity Supply Options of Electric Vehicles … 43

(a) Profit ratio vs. the number of deployed (b) Happy customers ratio vs. the number
sockets. of deployed sockets.

Fig. 2 Analysis of profit and happy customers ratios

charging sockets is 21. In other words, almost all customers succeeded charging
their EVs when there are 21 charging sockets. Note that the fixed investment cost of
charging sockets is not included in our study and the average utilization of charging
units is computed as 54% when there are 21 sockets (i.e., the charging sockets remain
idle on average for almost half of the operating times of the shopping mall). Hence,
one may consider investing in a lower number of charging sockets than 21 sockets
even if the happy customers ratio is the highest among other cases.
The electricity is monetized per kWh, hence, understanding the energy consump-
tion helps to understand the expected cost of this business and the decision-making
process for the investors. The hourly energy consumption representing the hourly
cost for a charging unit operator varies based on the utilization of the charging sockets
during an hour of the day. The arrivals are time-varying. For example, there are peak
hours during which long queues for charging are more likely to be observed. As
mentioned before, an EV user may not remove his/her EV from the charging socket
whenever charging is finished. Thus, the valet service may lower such inefficiencies
by switching a charged vehicle with another one from the queue that may result in
higher hourly energy consumption. Figure 3 represents the average hourly energy
consumption (with respect to operational hours of the shopping mall represented in
the x-axis) varying the number of sockets. As expected, the higher the number of
sockets the higher the hourly consumption. The profile of the energy consumption
curve is similar to the arrival curve of the shopping mall. The shopping mall starts
operating at 10 am and the average number of EV arrivals is the lowest compared
to the other operating hours of the day. Thus, the hourly average consumption is the

Fig. 3 Average hourly energy consumption vs. the number of sockets (3 to 21 from left to right)
44 E. Coban and G. Poyrazoglu

lowest for the initial operating hours. For the first couple of hours, the arrivals to the
shopping mall are increasing leading to an increase in the occupancy of the charging
sockets. Then, the hourly consumption keeps increasing nonlinearly until it reaches
its peak at the 8th operating hour and starts decreasing till the closing time of the
shopping mall as shown in Fig. 3.
When the CS has only a single charging unit, almost 90% of the average utilization
of the sockets throughout a day is computed by 100 runs of the simulation. Hence, the
hourly consumption curve behaves almost like a line at around 80kWh median. As
we increase the number of charging units, for example, seven charging units, some
sockets become idle, but the energy consumption curve due to the synchronous
occupancy follows a trend similar to the curve of hourly popularity of the shopping
mall. The peak is reached at the 8th hour, corresponding to the time period between
5 and 6 pm, with a median consumption of 390 kW.
In addition to average hourly energy consumption, tracking the instantaneous
power requirement is vital to assure the stable and secure operations of CSs. The
worst-case scenario (the maximum) of instantaneous power requirement is used for
electrical design projects to make sure the CS would be kept active during almost
all possible events and the smart charging studies show that there might be business
and operational opportunities by analyzing the instantaneous demand. Although our
focus is not on the electrical infrastructure of a CS, it is important to note that
unless the utilization of charging sockets hits 100%, there might be no need for over-
designing the electrical infrastructure with expensive investment. Figure 4 represents
the instantaneous power requirement (when the time increment is minutes) for seven
different deployment scenarios. The instantaneous power requirement for the cases
where there are more than 12 sockets deviates more than the instantaneous power
requirement of the remaining cases. We show that although the maximum power
capacity of seven charging units (i.e., 21 sockets) in this simulation is 693 kW, the
median of the peak power consumption is 440 kW. However, for a single charging
unit or two charging units, the peak power is reached within 3 h after the shopping
mall opening time, and the median peak power is kept constant for the next several
hours of operation as all charging units are almost constantly utilized (preventing
larger deviations over time).
To sum up, while designing a CS, the first utilization of charging sockets should
be checked to control whether the CS may realize any problems due to the design
of the electrical infrastructure. Then, the charging business should decide on the

Fig. 4 Instantaneous power consumption vs. the number of sockets (3 to 21 from left to right)
Demand-Driven Electricity Supply Options of Electric Vehicles … 45

service level guaranteed to its customers in addition to the profit ratio. The charging
business may also segment its customers by separating a group of charging units
reserved for premium customers or giving higher priority to premium customers
while providing service. The number of EV users will keep increasing with various
government incentives and global warming awareness. Even if utilizing real data
is impossible due to the dynamic nature of the problem, we provide a simulation
framework that can be easily updated with respect to data. For example, arrival rates
or types of charging units can be updated easily. Hence, charging companies can use
our simulation framework while designing their CSs.

5 Pricing Policies and Open Problems

Pricing EV charging is important as low pricing can speed up the adoption of EVs.
There are different pricing options for charging. For example, EVgo offers two types
of charging plans, pay-as-you-go and membership rates (Types of electric vehicles,
2020), and companies like General Motors helps EVgo triple its fast charger network
in the U.S. (O’Kane, 2020). Another example of pricing charging is ZES (Charge
station points, 2020) that charges depending on the type of the socket and average
power (i.e., the ratio of the energy used by the total time spent at CS). However,
these prices are valid for members, if you are not a member, an additional fixed cost
is charged. ZES also accepts a reservation, but a blockage fee is charged in case of a
missed reservation. Even if details of pricing policies vary, such as a special license
required to sell electricity in some countries prevents tracking of an EV’s real power
consumption, every company keeps track of the type of the sockets (AC, DC regular,
DC fast) and the energy consumption while charging (EŞARJ Membership & Tariffs,
2020).
Equivalent cost analysis between oil (petroleum) and electricity is vital to under-
stand the pace of EV adoption. One should note that comparison to oil prices is
very critical since some EV users may start switching back to oil or diesel cars.
For example, Norway’s new EV charging prices are expected to force drivers into
subscription agreements with carmakers in order to access cheaper charging rates
which increases concerns about fairness for owners of electric vehicles that charge
at slower rates (Schmidt, 2020).
In addition, electricity demand can be tailored as EV drivers can be promoted
through special pricing offers, such as charging an EV within 3 h at a specific CS will
cost less compared to charging the EV at a CS that is located at a popular destination.
Companies can charge membership fees based on their customers’ profiles (such as
the average distance traveled, preference about charging times, and their EV types).
Customer segmentation can be studied to offer different membership packages, such
as having priority over other EVs at all CSs or a guaranteed reservation system with
a premium membership. These packages should be designed carefully as well to
motivate the EV users to manage both the demand and the cost as much as possible.
46 E. Coban and G. Poyrazoglu

Another topic that may require further investigation is the Vehicle-to-Grid (V2G)
infrastructure, which may eventually integrate the EV batteries with the electricity
grid infrastructure to help providing power when it is needed. The batteries may be
used not only to provide power for the mobility of the vehicle, but they may also help
the grid operators to support voltage and frequency for secure and stable operations.

6 Conclusion

In this chapter, we discuss the challenges and research opportunities for public CSs.
The total number of EVs on the road keeps increasing due to global warming aware-
ness and governments’ policies and/or incentives. However, designing the public
charging network remains an important concern, even slowing the pace of widespread
use of EVs. Thus, the demand-driven electricity supply options of EVs at public CSs
should be carefully analyzed.
We first summarize the worldwide current EV condition via the current EV adop-
tions in five countries with the highest cumulative EV sales. The policies and/or
incentives of these countries are also mentioned briefly while explaining their EV
adoptions. A ratio of EV cumulative sales to the total number of charging units is
used as a performance metric to indicate the availability of the infrastructure. Then,
we study the location problem of CSs. We first explain an existing mathematical
model computing the location of public CSs and summarize the results of a possible
extension of that model: backup CS coverage. We find that the threshold distance
used to identify candidate backup CSs is very important as this extension may result
in infeasibility since all candidate CSs are forced to have a backup CS that is deployed
at optimality. We also summarize other possible extensions for the location model.
Given the locations of CSs, the next decision is to compute the number and the
type of charging units. We describe a simulation framework to model a public CS
(with/without valet service) varying the number and type of chargers. Due to a lack
of real historical data, Google Popular Times (Google, 2020) are utilized to generate
the EV arrivals to a shopping mall. We report several performance metrics, such as
average waiting time, energy consumption. We conclude that one may choose the
type and the number of charging units based on their performance metrics, such as
lower number of charging units will be preferred to achieve higher utilization on
all charging units, whereas a higher number of charging units will be preferred to
achieve lower average waiting time.
In addition, we discuss different pricing models and remaining open problems on
how to best design public CSs. To sum up, this chapter covers some of the strategic,
tactical, and operational decisions about public CSs showing which data are required
and what type of decisions can be made. We should always keep in mind that without
a proper charging infrastructure, the adoption of EVs will take longer than planned.
Demand-Driven Electricity Supply Options of Electric Vehicles … 47

References

Arayici, K. C., & Poyrazoglu, G. (2019). The optimal placement model for electric vehicle charging
stations. In 2019 International Conference on Power Generation Systems and Renewable Energy
Technologies (PGSRET) (pp. 1–4). IEEE.
B. E. Outlook (2019). 2019 edition, London, United Kingdom.
Bae, S., & Kwasinski, A. (2011). Spatial and temporal model of electric vehicle charging demand.
IEEE Transactions on Smart Grid, 3(1), 394–403.
Ballentin, R., Hartmann, D., Kopp, D., Loewel, T., Sund, M., & Templ, W. (2011). E-mobility in
the context of electric energy distribution grids. Bell Labs Technical Journal, 16(3), 47–60.
Charge Station Points (2020). https://zes.net/en/pricing.html. Accessed 1 Nov 2020.
Chen, T. D., Kockelman, K. M., & Khan, M. (2013). Locating electric vehicle charging stations:
Parking-based assignment method for Seattle. Washington, Transportation Research Record,
2385(1), 28–36.
C. P. U. Commission (2020). Zero-emission vehicles. www.cpuc.ca.gov/zev/. Accessed on 1 Aug
2020.
Coren, M. J. (2018). Nine countries say they’ll ban internal combustion engines, So far, it’s just
words. Quartz.
Covid-19 Response Will Influence the Pace of Transition to Electric Vehicles (2020). www.fuelsa
ndlubes.com/fli-article/covid-19-response-influence-pace-transition-electric-vehicles/?lang=en.
Accessed 7 Nov 2020.
Daskin, M. S. (2011). Service science. Wiley.
EŞARJ Membership and Tariffs (2020). https://esarj.com/en/membership. Accessed 1 Nov 2020.
Frade, I., Ribeiro, A., Gonçalves, G., & Antunes, A. P. (2011). Optimal location of charging stations
for electric vehicles in a neighborhood in Lisbon. Portugal, Transportation Research Record,
2252(1), 91–98.
Ghiani, G., Laporte, G., & Musmanno, R. (2004). Introduction to logistics systems planning and
control. Wiley.
Google. (2020). Popular times, wait times and visit duration. https://support.google.com/business/
answer/6263531?hl=en. Accessed on 26 May 2020.
Hall, D., Wappelhorst, S., Mock, P., & Lutsey, N. (2020). European electric vehicle factbook
2019/2020. Technical report (2020). https://theicct.org/sites/default/files/publications/EV-EU-
Factbook-2020.pdf
He, J., Yang, H., Tang, T.-Q., & Huang, H.-J. (2018). An optimal charging station location model with
the consideration of electric vehicles driving range. Transportation Research Part c: Emerging
Technologies, 86, 641–654.
IEA. (2020). Global EV outlook 2020-analysis. Technical report. www.iea.org/reports/global-ev-
outlook-2020
Keskin, M., Akhavan-Tabatabaei, R., & Çatay, B. (2019). Electric vehicle routing problem with
time windows and stochastic waiting times at recharging stations. In 2019 Winter Simulation
Conference (WSC) (pp. 1649–1659). IEEE.
Li, Y., & Zhang, J. (2015). Research into probabilistic representation of electric vehicle’s charging
load and its effect to the load characteristics of the network. In 2015 5th International Conference
on Electric Utility Deregulation and Restructuring and Power Technologies (DRPT) (pp. 426–
430). IEEE.
Luo, Y., Feng, G., Wan, S., Zhang, S., Li, V., Kong, W. (2020). Charging scheduling strategy for
different electric vehicles with optimization for convenience of drivers, performance of transport
system and distribution network. Energy, 194, 116807.
Lützenberger, M., Küster, T., & Albayrak, S. (2014). Integrating electric vehicles into smart grid
infrastructures a simulation-based approach that became reality. In Proceedings of the winter
simulation conference 2014 (pp. 1061–1072). IEEE.
Mobility Foresights (2020). China electric vehicles market 2020–2025. https://mobilityforesights.
com/product/china-electric-vehicle-market/. Accessed on 1 July 2020.
48 E. Coban and G. Poyrazoglu

Mordor Intelligence. (2020). China electric vehicles market-growth, trends, and forecasts (2020–
2025). Technical report. www.mordorintelligence.com/industry-reports/china-electric-vehicles-
ev-market-outlook
Number of Public Electric Vehicle Charging Stations and Charging Outlets in the U.S. as of March
3 (2020). www.statista.com/statistics/416750/number-of-electric-vehicle-charging-stations-out
lets-united-states. Accessed 1 July 2020.
Nuvve (2020). We make electric vehicles greener. https://nuvve.com/. Accessed on 1 Aug 2020.
Office of Energy Efficiency and Renewable Energy (2020). Vehicle charging. www.energy.gov/
eere/electricvehicles/vehicle-charging. Accessed 1 July 2020.
O’Kane (2020). General Motors will help EVgo triple its fast charger network in the
US. www.theverge.com/2020/7/31/21349614/general-motors-evgo-fast-charging-network-inv
estment. Accessed 31 Aug 2020.
Pflaum, P., Alamir, M., & Lamoudi, M. Y. (2018). Probabilistic energy management strategy
for EV charging stations using randomized algorithms. IEEE Transactions on Control Systems
Technology, 16(3), 1099–1106.
Poch, L., Mahalik, M., Wang, J., & Vyas, A. (2010). Impacts of plug-in hybrid electric vehicles on
the electric power system in the western united states. In IEEE PES general meeting (pp. 1–7).
IEEE.
Poyrazoglu, G., & Coban, E. (2021). A stochastic value estimation tool for electric vehicle charging
points. Energy, 227, 120335.
Rezaee, S., Farjah, E., & Khorramdel, B. (2013). Probabilistic analysis of plug-in electric vehicles
impact on electrical grid through homes and parking lots. IEEE Transactions on Sustainable
Energy, 4(4), 1024–1033.
Schmidt, B. (2020). Norway horrified as new rates make EV charging prices higher than petrol.
https://thedriven.io/2020/01/20. Accessed 1 Mar 2020.
Shahraki, N., Cai, H., Turkay, M., & Xu, M. (2015). Optimal locations of electric public charging
stations using real world vehicle travel patterns. Transportation Research Part d: Transport and
Environment, 41, 165–176.
Shen, Z.-J.M., Feng, B., Mao, C., & Ran, L. (2019). Optimization models for electric vehicle service
operations: A literature review. Transportation Research Part b: Methodological, 128, 462–477.
Shura Energy Transition Center. (2019). Transport sector transformation: Integrating electric
vehicles into Turkey’s distribution grids. Technical report. https://www.shura.org.tr/wp-content/
uploads/2019/12/Transport-sector-transformation.Integrating-electric-vehicles-into-Turkey%
E2%80%99s-distribution-grids.pdf
Sun, Z., Gao, W., Li, B., Wang, L. (2020). Locating charging stations for electric vehicles. Transport
Policy.
Teixeira, A. C. R., & Sodré, J. R. (2016). Simulation of the impacts on carbon dioxide emissions from
replacement of a conventional Brazilian taxi fleet by electric vehicles. Energy, 115, 1617–1622.
The Global Electric Vehicle Market in 2020: Statistics & Forecasts (2020). Technical report. www.
virta.global/global-electric-vehicle-market
Total Number of Public Electric Vehicle Charging Stations in China from 2015 to 2019 (2020). www.
statista.com/statistics/993121/china-public-electric-vehicle-charging-station-number/. Accessed
1 July 2020.
Types of Electric Vehicles (2020). www.evgo.com/why-evs/types-of-electric-vehicles/. Accessed
10 July 2020.
Wang, Y.-W. (2007). An optimal location choice model for recreation-oriented scooter recharge
stations. Transportation Research Part d: Transport and Environment, 12(3), 231–237.
Wiederer, A., & Philip, R. (2010). Policy options for electric vehicle charging infrastructure in C40
cities.
Xiang, Y., Liu, J., Li, R., Li, F., Gu, C., & Tang, S. (2016). Economic planning of electric vehicle
charging stations considering traffic constraints and load profile templates. Applied Energy, 178,
647–659.
Demand-Driven Electricity Supply Options of Electric Vehicles … 49

Zachary, S. (2020). 69% of autos sold in Norway in 2020 have a plug. https://cleantechnica.com/
2020/07/09/69-of-autos-sold-in-norway-in-2020-have-a-plug/. Accessed 10 July 2020.
Zhang, X., Xie, J., Rao, R., & Liang, Y. (2014). Policy incentives for the adoption of electric vehicles
across countries. Sustainability, 6(11), 8056–8078.
Zhu, Z.-H., Gao, Z.-Y., Zheng, J.-F., & Du, H.-M. (2016). Charging station location problem of
plug-in electric vehicles. Journal of Transport Geography, 52, 11–22.
A Review on Smart Energy Management
Systems in Microgrids Based on Power
Generating and Environmental Costs

Özgür İcan and Taha Buğra Çelik

Nomenclature

AC Absorption chiller
BESS Battery energy storage system
BMG Biomass generator
CC Compression chiller
CHPG Combined heat and power generator
CSS Cooling storage system
DE Differential evolution
DG Diesel generator
EC Electric chiller
EMS Energy management system
FC Fuel cell
GB Gas boiler
GE Gas engine
GT Gas turbine
HEG Hydro electric generator
HSS Heating storage system
MT Micro turbine
NG Natural gas
NPCE Net present cost of electricity
PAFC Phosphoric acid fuel cell
PV Photovoltaics

Ö. İcan (B) · T. B. Çelik

Faculty of Economics and Administrative Sciences, Ondokuz Mayıs University, Kurupelit
Kampüsü, İİBF, 55139 Atakum/Samsun, Turkey
e-mail: [email protected]
T. B. Çelik
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 51

A. B. Dorsman et al. (eds.), Applied Operations Research and Financial Modelling
in Energy, https://doi.org/10.1007/978-3-030-84981-8_4
52 Ö. İcan and T. B. Çelik

RES Renewable energy sources

RPL Real power loss
SG Synchronous generator
TCE Total cost of electricity
TE Total emission
TEC Total emission cost
TESS Thermal energy storage system
TPC Total production cost
UG Utility grid
VSI Voltage stability Index
WT Wind turbine
ABC Artificial bee colony
ACO Ant colony optimization
AMPSO Adaptive mutation particle swarm optimization algorithm
BA Bat algorithm
CSA Cuckoo search algorithm
CSOA Chicken swarm optimization algorithm
DEGL Differential evolution with global and local neighborhoods
DEMPSO Differential evolutionary (De) and modified Pso
DENGMS Differential evolution and the niche guided mating selection
strategy
DPLP Dynamic programming combined with the linear programming
GA Genetic algorithms
GAMS General algebraic modeling system
GOAPSNN Particle swarm optimization aided artificial neural network and
grasshopper optimization algorithm
GWO Grey wolf optimization
HBB-BC Hybrid big bang-big crunch
HPEMG-HGPO Heuristic-based programmable energy management controller
with hybrid genetic particle swarm optimization
IBA Improved bat algorithm
ICA Imperialist competitive algorithm
ISA Interior search algorithm
JAYAGM Jaya algorithm with the gradient method (Gm)
SFS Stochastic fractal search algorithm
LOHAWEC Lexicographic optimization and hybrid augmented-weighted
epsilon-constrain
MHS Modified harmony search
MILP Mixed-integer linear programming
MINLP Mixed-integer nonlinear programming
MOALO Multi-objective ant lion optimizer
MOFEPSO Multi-objective feasibility enhanced particle swarm optimization
MOPSO Multi-objective particle swarm optimization
NSGA Non-dominated sorting algorithms Ga
PAFC Phosphoric acid fuel cell
A Review on Smart Energy Management Systems … 53

Law Dome AntarcƟca Ice Core Data

340

330

CO2 Mixing Ratio (PPM)

320

310

300

290

280
1830 1880 1930 1980
Air Age (Year)

Fig. 1 Ice cap data showing CO2 mixing levels1

PSO Particle swarm optimization

PSOFMN Particle swarm optimization with fuzzy max–min technique
SOGSNN Squirrel optimization with gravitational search–aided neural
network
TSKFS-RBFNN Takagi–Sugeno–Kang fuzzy system under radial basis function
neural network

1 Introduction

As a result of the fossil-fuels backed industrialization within the last two centuries,
the inevitable greenhouse effect of the exhaust gases—mainly CO2, has made the
single biggest impact on the climate change. According to publicly available ice core
data provided by Etheridge et al. (1996) visualized in Fig. 1, CO2 mixing levels were
between 270 and 275 ppm ranges during pre-industrialization era while it is stated
that it is over 400 ppm according most recent studies.

1Source CDIAC. Carbon Dioxide Information Analysis Center. Historical CO2 Records from the
Law Dome. DE08, DE08 2, and DSS Ice Cores. Available online: https://cdiac.ess-dive.lbl.gov/tre
nds/co2/lawdome.html. (Accessed on 10 October 2020).
54 Ö. İcan and T. B. Çelik

Median Temperature Anomaly From 1960-1990 Average Temp

1

0.8

0.6

0.4

0.2
Celsius

0
1850 1870 1890 1910 1930 1950 1970 1990 2010
-0.2

-0.4

-0.6

-0.8
Year

Fig. 2 Temperature anomalies within the last century

Today, scientists have arrived at a consensus that the temperature anomalies are
the result of human activities (Cook et al., 2016). In Fig. 2 the temperature data taken
from Morice et al. (2012)’s the HadCRUT4 dataset2 on global climate change it can
be clearly seen that the anomaly has been accelerated alarmingly in the last decades.
The world leadership has arrived to the point that deepening climate crisis deserves
the top priority. Thus in 2015 in Paris there have been arrived at the first legally
binding international agreement concerning climate change. The agreement aims to
limit temperature increase well below 2 °C degrees above pre-industrialization levels
in this century.
Albeit the situation is worsening continuously, the necessity of employing conven-
tional power generating systems has still been considered as a valid option according
to certain policy makers because of the growing energy demand. However, the green-
house effect is well beyond the tolerable levels according to the consensus of the most
prominent researchers of the world and there’s no sustainable way for fossil based
energy being an important part of the power grids throughout world in the future.
Today, renewable energy systems are on an accelerated rise in the overall power
production. According to REN21,3 a think-tank focusing on renewable energy, on its
“2019 Global Status Report” it has been stated that, renewables correspond to %26 of
all produced electrical energy as of 2018 end by doubling their corporate resourcing
compared to previous year. The biggest penetration to the grid comes mainly from
hydro, wind power and solar systems. As the advances in technology of solar panels
are being witnessed, in particular pure solar systems or hybrid systems coupled with

2Source https://crudata.uea.ac.uk/cru/data/temperature/. (Accessed on 6 December 2020).

3REN21.Renewables 2019 Global Status Report; REN21 Secretariat: Paris, France, 2019; ISBN
978–3-9,818,911–7 1. Available online: http://www.ren21.net/gsr-2019/ (accessed on 6 September
2020).
A Review on Smart Energy Management Systems … 55

an array of other elements besides solar, like wind turbines are widely accepted as
the future of renewable power production.
Because of the nature of the resources of renewable energy generation, most
promising systems tend to be the ones aiming local demand. These decentralized
systems with considerably small capacity are considered as microgrids. Because
of the volatility of renewable energy resources, a system depending on a single
source cannot be considered as stable. Thus microgrids are systems which usually
consist of several subsystems harvesting energy from various renewable energy
resources simultaneously as well as employing conventional power generating
systems (gas/diesel generators or even utility grid) in case of need. Moreover these
systems also bear the ability to store the excess power by utilizing batteries and
battery management systems.
Considering the multi-component structure of these systems along with the
adherent complexity, reducing power production costs and controlling emission
values below target levels need system optimization. In this study, a systematic
review of papers on energy management systems and smart energy management
systems which aim to improve energy management by experimenting with various
computational optimization techniques has been conducted. Selection criteria and
other details are explained in Sect. 3.

2 Renewable Energy and Applications

Renewable energy or sometimes referred as clean energy is the type of energy derived
from natural resources which are capable of being naturally replenished without any
human efforts. The fact that it is being naturally replenished is a total outcome of
natural events like the sun light, wind, potential or kinetic energy of water masses or
biomass hence the term renewable has been coined for this type of energy sources.
Less than two centuries ago, almost all of the energy consumption of the world
population was coming from wood which can be technically considered as the most
naive form of renewable energy source, although not a carbon–neutral one. After
the industrial revolution, from the mid nineteenth century, coal and later petroleum
started to replace this primitive source of energy and nearly a hundred years later,
around mid twentieth century, with the addition of natural gas almost all of the
energy consumption was coming from fossil fuels. The most recent data derived by
US Energy Information Administration clearly shows that other than hydroelectric,
renewable sources have barely entered to grid shortly before 2000s.4 The same data
shows a fascinating trend in non-hydro resources that within two decades they have
managed to climb to a total power generating supply level in the whole US grid
which multiplying hydro recourses and nearly as much as nuclear sources. However,
the dominant source of energy is still fossil fuels which should be changed in the
favor of carbon–neutral energy sources immediately.

4 U.S. Energy Information Administration, Monthly Energy Review, Appendix D.1, April 2020.
56 Ö. İcan and T. B. Çelik

2.1 Hydroelectric Power

Hydroelectric power is generated from the fast-running or falling water like rivers.
Since ancient times, humans tried to explore the potential of water in generating the
necessary motion for their watermills for processing grains, for watering the crop
lands or for different necessities like cutting timber. During centuries, besides these
traditional use of hydro power naturally evolved into ideas for generating electricity.
In modern hydro power generating facilities, a running water or falling water from
a reservoir is directed to turbines by using tubes and hence generating electricity by
propelling the turbine blades which are internally connected to generators.

2.2 Solar Power

Solar energy is the type of energy derived from the sun itself. The sun is the main
source of energy of the life on the Earth and it has been used in primitive forms
throughout the human history for different purposes like heating water or drying
foods. Today solar power means mostly generating electricity by photovoltaics (PV)
or concentrated solar power (CSP).
In PV sunlight is turned int o electricity using solar panels which exploit photo-
voltaic effect, a chemical phenomenon. In CSP sunlight is concentrated to a small
area in order to heat water and produce electricity via running a steam turbine. These
two approaches can be used in an hybrid way for producing both electricity and heat,
although in today’s renewable energy vision especially PV has a very bright future
as the solar panel technology advances as nearly efficient as harvesting nearly %50
of solar energy it takes.

2.3 Wind Power

Sunlight causes, different parts of lands and oceans, to have different temperature,
pressure and humility profiles. As a result of this divergence between these different
regions of Earth, they warm up and cool down differently compared to each other
hence the air currents between these regions occur. As the warm air elevates in the
atmosphere another cooler mass of air fills in to the region. This movement of mass
of air is called what basically known as winds.
Wind power is usually generated by the help of wind turbines which are basically
blades connected to rotor of a generator, similar to the principles of harnessing
kinetical energy of water in hydroelectrical units.
A Review on Smart Energy Management Systems … 57

2.4 Other Renewables and Contemporary Trends

Among others renewable energy sources, tidal (generating electricity from ocean
waves’ kinetic energy), biomass (generating electricity from burning gases which
arise as the result of the biochemical decomposition of organic material or simply
burning the solid waste and using a steam generator) and geothermal (generating
electricity or heating an area by using naturally available geothermal water resources)
can be listed as the most promising technologies.
However, the main trend in RES is using solar energy based systems especially
PVs with solar panels having the most recent advanced infrastructures in order to
harness maximum energy from the direct sunlight in addition to the wind turbines.
These two type of RES seem to be prevailing as the main generating components in
most of the MG architectures in near future.

3 Microgrids and Smart Energy Management Systems

(SEMS)

MGs are systems in which various electricity generation technologies are operated in
harmony in order to meet a relatively small electrical energy demand. In line with the
needs and constraints, MGs can operate as connected to the main utility grid (UG)
or can operate as islanded mode. The aim of micro grids connected to the utility
grid is to reduce production costs by utilizing renewable energy resources. On the
other hand, in islanded MGs, in addition to reducing consumption costs, it can be an
alternative solution if the UG cannot be reached. In micro grid systems, in addition to
renewable energy sources such as PV, WT and HEG, traditional components such as
DG, MT and FC can be included in the system to secure the electricity distribution.
A representative MG is illustrated in the figure below.
As is seen in Fig. 3, energy storage systems can also be included in MGs. Thus,
electricity that can be produced at low cost for a certain period of time is stored and
it can be used for time periods where production is more costly. At this stage, the
issue of the distribution of electricity produced by various sources and technologies
within MGs gains great importance. The importance of this issue arises from the
consequences of the operation of the system in terms of both economic and envi-
ronmental costs. As a matter of fact, if the generated energy can be stored, there are
constraints and costs of storing energy, on the other hand, if the ESS is included in
the system, the lowest cost electricity that can be produced or supplied for a certain
period of time (for example via UG) should be preferred to meet the demand. At this
point, the problem of determining the most appropriate decisions regarding the oper-
ational and distribution issues of the electricity generated by MGs is encountered. In
order to overcome this problem, smart energy management systems (SEMS) have
come to the fore recently. In the next section SEMS are introduced and related issues
are discussed.
58 Ö. İcan and T. B. Çelik

Fig. 3 Microgrid power generating system

3.1 Smart Energy Management Systems

EMSs are systems in which energy production and distribution are optimized through
hardware and software components in order to meet economic and environmental
requirements by controlling and monitoring the system. The rapid development in
computer hardware and software technologies in recent years has begun to show its
effect in this area as well. Especially the advancement of computational optimiza-
tion techniques such as machine learning and evolutionary algorithms has led to
significant improvements in “smart” decision support systems. In short, employing
advanced computational optimization techniques is the key factor which turns an
EMS into a SEMS.
SEMS applications in the literature show that computational optimization tech-
niques can be applied for different purposes. For instance, in some studies in the
literature (see Saiprasad et al., 2019; Nurunnabi et al., 2019; Boqtob et al., 2019)
primarily the size and capacities of the whole MG or its various components are
optimized in line with needs and constraints. On the other hand, energy production
and distribution can be optimized for economic and environmental purposes through
EMS of an already installed MG. In this study, such researches are included in the
scope of the study. In order to ensure that the performance of the findings obtained
as a result of optimizing energy production and distribution through computational
optimization techniques on SEMS is testable, the improvement made is desired to be
measurable. For this purpose, studies involving comparisons with the performance
A Review on Smart Energy Management Systems … 59

of the system before optimization or with other optimization techniques are included
in the scope of the review. There are many more studies in the literature optimizing
SEMS than included in this paper but since studies involving comparisons with the
performance of the system before proposed optimization technique is applied, are
desired, others are excluded.
The review is summarized in the Tables 1 and 2 below. The Tables contains
seven columns for summary information of reviewed studies which are authors
and year, power generating components, optimization methods, total reduced cost
(%), reference point for cost, reduced emission (%) and reference point for emis-
sion, respectively. In the power generating components column, power generating

Table 1 Review summary (reference point (cost): base case)

Authors and Power Optimization Total Reference Reduced Reference
year generating methods reduced point (cost) emission point
components cost (%) (%) (emission)
Fazlhashemi NG, PV, WT, HBB-BC 30.50 Base case 4.59% Base case
et al. (2020) BES, UG, Algorithm
CHP
Elkadeem PV, WT, FC, Homer Pro® 0.46 Base case 0.48% Base case
et al. (2020) BES, DG optimization
software
Imran et al. PV, UG, BPSO 17.64 Base case 14.66% Base case
(2020) BES
Muqeet and PV, DG, MILP 32.00 Base case N/A N/A
Ahmad BES, UG
(2020)
Maulik and PV; WT, PSOFMN 4.26 Base case 13.91% Base case
Das (2019) DG, GT
Saberi et al. PV, WT, GT, CPLEX 3.97 Base case 2.26% Default
(2019) GB, EC, AC, Solver
BES (GAMS
software)
Ansari et al. PV, WT, UG, GAMS 50.26 Base case 21.03% Base case
(2019) DG
Jin et al. PV, WT, DG, DENGMS 26.41 Base case 71.93% Base case
(2019) MT, FC
Et-Taoussi PV, GT, DPLP 38.81 Base case 21.00% Base case
et al. (2019) BES,
Pooranian PV, UG, MILP 51.62 Base case N/A N/A
et al. (2018) BES, HEG,
CHP, BMG
Boulal et al. PV, UG, CPLEX 3.97 Base case 2.26% Base case
(2018) BES, GT Solver
(GAMS
software)
Table 2 Review summary (reference point (cost): other techniques)
60

Authors Power generating Optimization methods Total Reference point (cost) Reduced Reference
components reduced emission point
cost (%) (%) (emission)
Elsakaan et al. (2019) MT, FC, PV, WT, BES, UG EMFO 4.93 CPSO-L 1.97% AMPSO-T
Murty and Kumar (2020) PV, WT, FC, MT, DG, BES MILP 1.34 Tabu Search PSO 0.30% ABC
Alomoush (2019) WT, PV, DG, GT, FC, MT, NSGA-III 8.91 MOFEPSO 11.59% NSGA-II
UG
Dey and Bhattacharyya (2019) GT, FC, PV, WT, DG, BES DEGL 2.46 DE N/A N/A
Zhou et al. (2019) PV, WT, MGT, BES, UG MPC 0.30 TOMLAB N/A N/A
Ghanbari-Mobarakeh and Moradian PV, WT, MT, BES CSOA 6.50 TOMLAB 1.14% TOMLAB
(2019)
Sedighizadeh et al. (2019) PV, WT, PAFC, MT, BES DEMPSO 7.80 MINLP 18.03% MINLP
Roy (2019) PV, MT, WT, BES GOAPSNN 27.92 ANFASO N/A N/A
Trivedi et al. (2018) PV, WT, HEG, BMG, NG, ISA 3.0 CSA 4.00% CSA
DG, BES
Elattar (2018) PV, WT, SG, CHPG MHS 1.99 ISA N/A N/A
Hosseini et al. (2017) FC, MT, PV, WT, BES MOALO 0.21 PSO 0.02% GA
Kamboj and Chanana (2016) PV, WT, FC, BES, UG MINLP 24.44 AMPSO 3.00% AMPSO
Bhoye et al. (2016) UG, PV, WT, DG, BES, NG, JAYAGM 0.35 PSO 1.88% PSO
Aghajani et al. (2015) PV, WT, BES, UG, MT, MOPSO 8.83 NSGA-II 19.26% NSGA-II
PAFC
Rezvani et al. (2015) PV, FC, BES, WT, MT LOHAWEC 5.55 PSO 26.60% PSO
(continued)
Ö. İcan and T. B. Çelik
Table 2 (continued)
Authors Power generating Optimization methods Total Reference point (cost) Reduced Reference
components reduced emission point
cost (%) (%) (emission)
Vasanthakumar et al. (2015) PV, WT, FC, DE, GT, BES CSA 2.90 PSO 0.26% PSO
Jiajie et al. (2015) MT, DG, PV, WT Improved ICA 0.64 ICA N/A N/A
Trivedi et al., (2015) PV, WT, CHP, Synchronous ACO 5.55 Gradient Method 8.70% Gradient
generators (2) Method
Roy and Mandal (2014) WT, PV, FC, MT, DG, BES Hybrid ABC 5.25 ABC N/A N/A
A Review on Smart Energy Management Systems …
61
62 Ö. İcan and T. B. Çelik

components of the MG system for each study is listed. Computational optimiza-

tion techniques adopted to optimize energy generation and distribution in SEMS is
included in optimization methods column. The fourth and sixth columns include the
percentage reduced energy production cost and percentage reduced emission after
the optimization techniques are applied in the SEMS.
As is seen in the Tables 1 and 2, huge improvements are made in terms of energy
and emission cost reduction in some studies. Several details should be taken into
account in order to avoid the misinterpretation of the information summarized in
Tables. Reference point columns for “Total Reduced cost (%)” and “Reduced emis-
sion (%)” are added to prevent misinterpretation here since reference points for
improvements in SEMS are different. It also would not be wise to compare perfor-
mance of optimization techniques between studies since each SEMS which is utilized
to conduct study has different conditions, components, environmental factors etc.
from each other.
Reviewed studies here are divided into two categories in terms of reference
point of total cost reduction. The first category include studies, where authors try
to compare the concerning technique’s improvements over a base-case scenario, i.e.
the case with no computational optimization techniques are employed. The second
category consists of the studies where the researchers follow a strategy based on
comparing a proposed technique with other techniques. These categories and the
related information are summarized in Tables 1 and 2 respectively.
In order to evaluate the information summarized in the Tables properly, which
components are included in SEMS should be considered. For instance, some of
the studies in Tables 1 and 2 include BES in microgrids. Energy storage systems
provide more room for optimization of energy production costs and distribution
stability. In Table 1, studies including BES in microgrid starts with (Fazlhashemi
et al., 2020) and they used HBB-BC algorithm to optimize the operation cost, real
power loss, the voltage stability index (VSI), and the greenhouse gas emissions of the
MG. (Elkadeem et al., 2020) proposes a systematic and integrative decision-making
approach for efficient planning and assessment of hybrid renewable energy-based
MG. (Imran et al., 2020) proposes a heuristic-based programmable energy manage-
ment controller to manage the energy consumption in residential buildings to mini-
mize electricity bills, reduce carbon emissions, maximize user comfort and reduce the
peak-to-average ratio. Muqeet and Ahmad (2020), considering a real-time university
campus that is currently feeding its load from the national grid only, propose an EMS
strategy for an institutional MG to reduce its operational cost and increase its self-
consumption from green RES. Saberi et al. (2019) proposed a multi-objective model
for reducing carbon emission and operation cost in the presence of real time demand
response program. Boulal et al. (2018) aims the problem of optimal management of
the energy flows within a multi-source system of electricity production. Et-Taoussi
et al. (2019) optimize an EMS is based on 24 h ahead forecast data of the residential
loads consumption, the photovoltaic power and the electricity tariffs for minimizing
the production cost of the active/reactive power, and reducing the CO2 equivalent
emissions. Pooranian et al. (2018), address the problem of minimizing the total daily
energy cost in a smart residential building composed of multiple smart homes with
A Review on Smart Energy Management Systems … 63

the aim of reducing the cost of energy bills and the greenhouse gas emissions under
different system constraints and user preferences.
Remaining studies in Table 1 are the ones which has no BES included in microgrid
system, such as (Maulik & Das, 2019) which introduces an optimal power dispatch
strategy for simultaneous reduction of cost and emission from generation activities in
an AC–DC hybrid MG under load and generation uncertainties. Ansari et al. (2019)
proposed a framework that minimizes operation cost, and GHG emissions. Jin et al.
(2019) presents a multi-objective energy management to optimize the renewable MG
operation while satisfying a demand response and various operation constraints.
On the other hand, there are such studies where several computational optimization
techniques are compared in terms of total reduced cost and reduced emission and such
studies are listed in Table 2. In some of the studies listed here, several computational
optimization techniques are compared with proposed techniques but in order to keep
it brief, only best technique compared to proposed technique is displayed.
Number of studies including BES in Table 2 is also dominating same as Table
1. Here, (Elsakaan et al., 2019) aims at finding the optimal scheduling of renewable
energy resources in isolated and grid-connected MGs and the problem is formulated
as a non-linear constrained multi-objective optimization problem. Murty and Kumar
(2020) established an optimal energy dispatch strategy for grid connected and stan-
dalone MG. Dey and Bhattacharyya (2019) uses three soft computing techniques
which are particle swarm optimization, differential evolution, and differential evolu-
tion with local global neighborhoods to perform a novel dynamic cost analysis of
a MG and minimize its overall cost which includes fuel, emission, operation and
maintenance cost, installation, and depreciation costs. Zhou et al (2019) designed
a novel model that has predictive control with feedback correction to optimize the
energy dispatch and minimize the operation costs of a MG. Ghanbari-Mobarakeh
and Moradian (2019), based on demand response program, proposed a new energy
management scheme in order to obtain the optimized performance of the MG.
Sedighizadeh et al. (2019) the stochastic operation scheduling of a MG consisting of
non-dispatchable and dispatchable resources to minimize operation cost and emis-
sions. Roy (2019) presents a hybrid technique for optimal power flow management
and production cost minimization of MGs-connected system with RES. resources.
Trivedi et al. (2018) schedule the generating units within their bounds together with
minimizing the fuel cost and emission Values with Interior Search Algorithm (ISA).
Hosseini et al. (2017) proposes a sustainable simulation method for managing energy
resources from the point of view of virtual power players operating in a smart grid.
Kamboj and Chanana (2016) utilized MINLP to obtain the optimal power schedule
of a day. Bhoye et al. (2016) solve the problem of Combined Economic Emission
Dispatch and it is optimized by meta-heuristic techniques. Aghajani et al. (2015)
proposed a multi-objective energy management system in order to optimize MG
performance in a short-term in the presence of RESs. Rezvani et al. (2015) proposes
a multi-objective framework for the optimal scheduling of a MG in order to concur-
rently minimize the total operation cost and minimize the emission caused by gener-
ating units. In the work of (Vasanthakumar et al., 2015), using CSA, generation cost
and emission cost of the MG are minimized while satisfying system hourly demand
64 Ö. İcan and T. B. Çelik

and system constraints. Roy and Mandal (2014) takes into consideration the optimal
configuration of the MG at a minimum fuel cost, operation and maintenance costs
as well as emissions reduction with ABC algorithm.
Remaining studies in Table 2 are the ones that BES is not included in the micro-
grid. Alomoush (2019) deals with the multi-objective economic-emission dispatch
problem which aims at operating costs, emission level, emission tax, and cost
of power purchase from the main external grid. Elattar (2018) adopted modified
harmony search (MHS) algorithm to solve the combined economic emission dispatch
(CEED) problem of the MG taking into account the solar and wind power cost func-
tions. Jiajie et al. (2015), aiming at minimizing MG total operation costs, set up a
mathematical model of MG dynamic economic dispatch based on chance constrained
programming. Trivedi et al. (2015) proposes multiple environment dispatch problem
solution in MGs using ACO technique to solve the generation dispatch problem.

4 Conclusion

Under the light of contemporary literature on renewable energy, it appears that the
main goal in the future of energy markets would be the rapid transformation of
power generation into a carbon–neutral state in a very agile manner. This can be
achieved only by effectively employing renewable energy resources which is also
the consensus of the major stakeholders.
Electricity generation units, which are combination of power generating and
storing components that harness renewable energy resources effectively are often
qualified as MGs. MGs have the ability to meet the local demand in place, without
resorting to the main grid, hence they ease the objective of abandoning fossil fuels
and reducing carbon footprint of generated energy per unit. Satisfying local demand
without creating any additional burden on the main grid and incurring minimal power
losses are also the driving factors that would keep the MGs architecture favorable
in the future. However, as today’s RES technologies have not yet reached sufficient
capacity, they’re often connected to the main grids.
Another handicap is that these energy sources can show serious fluctuations in
electricity production as they are affected by unpredictable or very difficult to predict
natural factors like weather or wind characteristics. When also the variability in the
electricity demand of consumers, especially for household/private consumption, is
considered, energy management i.e. power generation and dispatch in MGs can
become quite complex. In this respect, it is beneficial to use prediction techniques
along with the optimization of EMS in order to reduce the uncertainty in natural
conditions.
When the trend in electricity markets is obviously pro-liberal and as long as the
auction based pricing schemes are valid, demand response programs are fundamental
for managing the load on the grid. Moreover, grid participants need to plan their
operations and dispatch schemes according to their costs and market prices. These
costs are broadly power generating costs and environmental costs. Although the cost
A Review on Smart Energy Management Systems … 65

of generating electricity has been studied well and always considered as the main
objective so far, environmental costs are no-more a secondary concern but a primary
one and should has to be managed very carefully along with operating costs. In
this sense, it seems beneficial to adopt a broader approach that would put demand
response programs in the center of EMS optimization in order to minimize both the
economic and environmental costs of fluctuations in electrical load.

References

Aghajani, G. R., Shayanfar, H. A., & Shayeghi, H. (2015). Presenting a multi-objective generation
scheduling model for pricing demand response rate in micro-grid energy management. Energy
Conversion and Management, 106, 308–321.
Alomoush, M. I. (2019). Microgrid combined power-heat economic-emission dispatch considering
stochastic renewable energy resources, power purchase and emission tax. Energy Conversion and
Management, 200, 112090.
Ansari, M., Ansari, M., & Asrari, A. (2019). A framework for simultaneous management of green-
house gas emission and substation transformer congestion via cooperative microgrids. In: 2019
North American power symposium (NAPS) (pp. 1–6). IEEE.
Bhoye, M., Pandya, M. H., Valvi, S., Trivedi, I. N., Jangir, P., & Parmar, S. A. (2016). An emis-
sion constraint economic load dispatch problem solution with microgrid using JAYA algorithm.
In: 2016 international conference on energy efficient technologies for sustainability (ICEETS)
(pp. 497–502). IEEE.
Boqtob, O., El Moussaoui, H., El Markhi, H., & Lamhamdi, T. (2019). Optimal sizing of grid
connected microgrid in Morocco using Homer Pro. In: 2019 international conference on wireless
technologies, embedded and intelligent systems (WITS) (pp. 1–6). IEEE.
Boulal, A., Chakir, H. E., Drissi, M. H., Griguer, H., & Ouadi, H. (2018). Optimal management
of energy flows in a multi-source grid. In: 2018 renewable energies, power systems & green
inclusive economy (REPS-GIE) (pp. 1–6). IEEE.
CDIAC. (2020). Carbon dioxide information analysis center. Historical CO2 records from the Law
Dome DE08, DE08–2, and DSS Ice Cores. Available online: https://cdiac.ess-dive.lbl.gov/tre
nds/co2/lawdome. Accessed on 10 Oct 2020.
Cook, J., Oreskes, N., Doran, P. T., Anderegg, W. R., Verheggen, B., Maibach, E. W., & Rice, K.
(2016). Consensus on consensus: a synthesis of consensus estimates on human-caused global
warming. Environmental Research Letters, 11(4), 048002.
Dey, B., & Bhattacharyya, B. (2019). Dynamic cost analysis of a grid connected microgrid using
neighborhood based differential evolution technique. International Transactions on Electrical
Energy Systems, 29(1), e2665.
Elattar, E. E. (2018). Modified harmony search algorithm for combined economic emission dispatch
of microgrid incorporating renewable sources. Energy, 159, 496–507.
Elkadeem, M. R., Wang, S., Azmy, A. M., Atiya, E. G., Ullah, Z., & Sharshir, S. W. (2020). A
systematic decision-making approach for planning and assessment of hybrid renewable energy-
based microgrid with techno-economic optimization: A case study on an urban community in
Egypt. Sustainable Cities and Society, 54, 102013.
Elsakaan, A. A., El-Sehiemy, R. A., Kaddah, S. S., & Elsaid, M. I. (2019). Optimal economic–
emission power scheduling of RERs in MGs with uncertainty. IET Generation, Transmission &
Distribution, 14(1), 37–52.
Etheridge, D. M., Steele, L. P., Langenfelds, R. L., Francey, R. J., Barnola, J. M., & Morgan, V. I.
(1996). Natural and anthropogenic changes in atmospheric CO2 over the last 1000 years from air
in Antarctic ice and firn. Journal of Geophysical Research: Atmospheres, 101(D2), 4115–4128.
66 Ö. İcan and T. B. Çelik

Et-Taoussi, M., Ouadi, H., & Chakir, H. E. (2019). Hybrid optimal management of active and reactive
power flow in a smart microgrid with photovoltaic generation. Microsystem Technologies, 25(11),
4077–4090.
Fazlhashemi, S. S., Sedighizadeh, M., & Khodayar, M. E. (2020). Day-ahead energy management
and feeder reconfiguration for microgrids with CCHP and energy storage systems. Journal of
Energy Storage, 29, 101301.
Ghanbari-Mobarakeh, P., & Moradian, M. (2019). A new paradigm for distributed generation
management considering the renewable energy uncertainties and demand response resources.
International Journal of Renewable Energy Research (IJRER), 9(1), 215–225.
Hosseini, K., Araghi, S., Ahmadian, M. B., & Asadian, V. (2017). Multi-objective optimal
scheduling of a micro-grid consisted of renewable energies using multi-objective ant lion
optimizer. In: 2017 Smart Grid Conference (SGC) (pp. 1–8). IEEE.
Imran, A., Hafeez, G., Khan, I., Usman, M., Shafiq, Z., Qazi, A. B., & Thoben, K. D. (2020).
Heuristic-based programable controller for efficient energy management under renewable energy
sources and energy storage system in smart grid. IEEE Access, 8, 139587–139608.
Jiajie, W., Birong, X., & Shulei, D. (2015). Dynamic economic dispatch of MicroGrid using
improved imperialist competitive algorithm. In: 2015 8th international conference on intelligent
computation technology and automation (ICICTA) (pp. 397–401). IEEE.
Jin, S., Mao, Z., Li, H., & Qi, W. (2019). An improved decomposition based multi-objective evolu-
tionary algorithm for the operation management of a renewable micro-grid. Journal of Renewable
and Sustainable Energy, 11(1), 015303.
Kamboj, A., & Chanana, S. (2016). Optimization of cost and emission in a Renewable Energy
micro-grid. In: 2016 IEEE 1st international conference on power electronics, intelligent control
and energy systems (ICPEICES) (pp. 1–6). IEEE.
Maulik, A., & Das, D. (2019). Optimal power dispatch considering load and renewable generation
uncertainties in an AC–DC hybrid microgrid. IET Generation, Transmission & Distribution,
13(7), 1164–1176.
Morice, C. P., Kennedy, J. J., Rayner, N. A., & Jones, P. D. (2012). Quantifying uncertainties
in global and regional temperature change using an ensemble of observational estimates: The
HadCRUT4 data set. Journal of Geophysical Research: Atmospheres, 117(D8).
Muqeet, H. A. U., & Ahmad, A. (2020). Optimal scheduling for campus prosumer microgrid
considering price based demand response. IEEE Access, 8, 71378–71394.
Murty, V. V. S. N., & Kumar, A. (2020). Multi-objective energy management in microgrids with
hybrid energy sources and battery energy storage systems. Protection and Control of Modern
Power Systems, 5(1), 1–20.
Nurunnabi, M., Roy, N. K., Hossain, E., & Pota, H. R. (2019). Size optimization and sensitivity
analysis of hybrid wind/PV micro-grids-a case study for Bangladesh. IEEE Access, 7, 150120–
150140.
Pooranian, Z., Abawajy, J. H., & Conti, M. (2018). Scheduling distributed energy resource operation
and daily power consumption for a smart building to optimize economic and environmental
parameters. Energies, 11(6), 1348.
Rezvani, A., Gandomkar, M., Izadbakhsh, M., & Ahmadi, A. (2015). Environmental/economic
scheduling of a micro-grid with renewable energy resources. Journal of Cleaner Production, 87,
216–226.
Roy, K., & Mandal, K. K. (2014). Hybrid optimization algorithm for modeling and management of
micro grid connected system. Frontiers in Energy, 8(3), 305–314.
Roy, K. (2019). Analysis of power management and cost minimization in MG—A hybrid
GOAPSNN technique. International Journal of Numerical Modelling: Electronic Networks,
Devices and Fields, 32(5), e2624.
Saberi, K., Pashaei-Didani, H., Nourollahi, R., Zare, K., & Nojavan, S. (2019). Optimal performance
of CCHP based microgrid considering environmental issue in the presence of real time demand
response. Sustainable Cities and Society, 45, 596–606.
A Review on Smart Energy Management Systems … 67

Saiprasad, N., Kalam, A., & Zayegh, A. (2019). Triple bottom line analysis and optimum sizing of
renewable energy using improved hybrid optimization employing the genetic algorithm: A case
study from India. Energies, 12(3), 349.
Sedighizadeh, M., Esmaili, M., Jamshidi, A., & Ghaderi, M. H. (2019). Stochastic multi-objective
economic-environmental energy and reserve scheduling of microgrids considering battery energy
storage system. International Journal of Electrical Power & Energy Systems, 106, 1–16.
Trivedi, I. N., Jangir, P., Bhoye, M., & Jangir, N. (2018). An economic load dispatch and multiple
environmental dispatch problem solution with microgrids using interior search algorithm. Neural
Computing and Applications, 30(7), 2173–2189.
Trivedi, I. N., Thesiya, D. K., Esmat, A., & Jangir, P. (2015). A multiple environment dispatch
problem solution using ant colony optimization for micro-grids. In: 2015 international conference
on power and advanced control engineering (ICPACE) (pp. 109–115). IEEE.
Vasanthakumar, S., Kumarappan, N., Arulraj, R., & Vigneysh, T. (2015). Cuckoo search algorithm
based environmental economic dispatch of microgrid system with distributed generation. In: 2015
international conference on smart technologies and management for computing, communication,
controls, energy and materials (ICSTM) (pp. 575–580). IEEE.
Zhou, X., Yan, H., Zhang, H., & Peng, C. (2019). Model predictive control with feedback correc-
tion for optimal energy dispatch of a networked microgrid. Transactions of the Institute of
Measurement and Control, 41(6), 1540–1552.
Measuring Efficiency and Productivity
Change in the Turkish Electricity
Distribution Sector

Yetkin Cinar and Tekiner Kaya

1 Introduction

The electric power industry is the pillar of national economic development and
the support of people’s daily lives (Zhang et al., 2017). Chen et al. (2017) assert
that all economies around the world are electricity dependent. However, environ-
mental consciousness is becoming an increasingly important variable in the elec-
tricity industry. This transformation urges electricity distribution companies (EDCs)
to redesign their networks, reduce losses and utilise fewer pollutant resources, which
means increasing efficiency.
Efficiency in electricity distribution is generally determined by technical, manage-
rial, technological, and environmental factors. A considerable body of empirical liter-
ature has grown over the years, investigating the role of ownership, environmental-
structural and geographical factors and internal dynamics as a determinant of firm
performance and productivity. Data Envelopment Analysis (DEA), which is a non-
parametric efficiency measurement method for operations with multiple inputs and
outputs is one of the most commonly used mathematical programming models for
performance evaluation (Liu et al., 2013). DEA has been widely utilised in bench-
marking the performance of electricity distribution and production companies, as in
other fields.
large empirical literature has grown over the years investigating the role of
ownership as.

Y. Cinar (B)
Faculty of Political Sciences, Ankara University, Cemal Gursel Str., 06590 Cebeci/Ankara, Turkey
e-mail: [email protected]
T. Kaya
Faculty of Economics and Administrative Sciences, Nevsehir Haci Bektas Veli University,
Zubeyde Hanim Str., 50100 Nevsehir, Turkey

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 69

A. B. Dorsman et al. (eds.), Applied Operations Research and Financial Modelling
in Energy, https://doi.org/10.1007/978-3-030-84981-8_5
70 Y. Cinar and T. Kaya

Electricity distribution in Turkey is conducted by 21 regional monopolies.

Successful completion of the distribution sector liberalisation process in 2013 trans-
ferred 21 distribution regions to the private sector with US$13 billion in privatisation
revenue in total. The primary motivation of this chapter is to present how the EDCs
have changed after the privatization in terms of efficiency. Furthermore, the effects
of environmental and regional factors on EDCs’ efficiency are investigated. To our
knowledge, no studies have yet assessed the efficiency of the electricity distribution
sector since liberalisation in Turkey. Therefore, our study would be a significant
contribution to the literature in the relevant field. Additionally, since Turkey has
regions that have considerably different local dynamics, EDCs’ efficiencies may
differ based on these environmental factors. Understanding these dynamics may
enable policymakers to enact better laws and contribute to the use of resources more
efficiently.
The structure of this chapter is as follows. In Sect. 2, empirical studies analysing
the performance of the electricity distribution sector are reviewed. Section 3 presents
a brief overview of the Turkish electricity market. Methodology, variable selection,
and data are presented in Sect. 4, including the structure of the DEA process and
the efficiency measurements of Turkish EDCs. Section 5 discusses the results, and
conclusions are drawn in Sect. 6.

2 Literature Review

The relative efficiency of EDCs has attracted considerable interest worldwide in the
past decade due to the restructuring of the electricity energy sector, particularly with
the introduction of regulation, privatisation and trade liberalisation in several coun-
tries (Santos et al., 2011). Many papers have been published that compile the studies
of efficiency measurements of EDCs, utilising DEA for different dimensions, such
as decision making units (DMUs), period of time, inputs or outputs utilised in the
studies, methodology and country information (Petridis et al., 2019; Pérez-Reyes &
Tovar, 2009; Fallahi et al., 2019; Tavassoli et al., 2020; Santos et al., 2011; Jamasb
and Pollitt, 2000). Another complete overview of the relevant studies is provided by
Jamasb & Pollitt (2001), including single and cross-country benchmarking break-
downs. Zhou et al. (2008) additionally list 100 studies, including the details of type
of study, country or region, methodological aspect and application scheme (covering
electricity distribution).
In total, 65 studies that utilise DEA and DEA-based multi-model approaches
are reviewed in this section. Of these, 30 utilise financial inputs or outputs in their
models. The majority of studies (23 of 30) were conducted before 2010. Bodbe &
Tanaka (2018) examine the efficiency of EDCs by utilising DEA in India with data
from 2005–2012. They utilise total assets as input to represent the length of line and
transformer capacity. Recently, Şirin (2017) has applied panel data methodology to
understand the factors affecting the costs of the EDCs from 2011–2014. The research
utilises the length of line, the number of customers, energy demand, energy losses
Measuring Efficiency and Productivity Change … 71

(%), ownership, retail unbundling, and real exchange rates as explanatory variables,
whereas real total expenditures (capital expenditures and operating expenditures)
without total expenditures and with financing expenses are dependent variables.
Borghi et al. (2016) additionally utilise financial variables to analyse the effect of
ownership on EDC-level productivity via fixed-effects Ordinary Least Square (OLS).
They utilise tangible fixed assets, the number of employees, and material cost as
inputs and operating revenues as output.
The remaining 35 studies measure the efficiency of EDCs by utilising technical
and operational inputs and outputs, as in our study, and they are presented in Table
1.
Regarding Turkey, there are a number of studies (Table 1) measuring EDCs’
performance. The first study measuring EDCs’ distribution performance in Turkey
is from Bagdadioglu et al. (1996). Results of the study indicate that high-performance,
state-owned EDCs have been determined to be privatised. Bagdadioglu et al. (2007)
have analysed the gains of the mergers among EDCs in Turkey by utilising panel data
between 1999 and 2003. Çelen (2013) has examined panel data for the power distri-
bution sector in Turkey and provides empirical evidence that private ownership has a
positive effect on EDC efficiency. Çelen & Yalçın (2012) have proposed a combined
multi-criteria methodology of fuzzy Analytical Hierarchy Process (AHP)/Technique
for Order Performance by Similarity to Ideal Solution (TOPSIS)/DEA methods and
have applied it to the Turkish electricity distribution market to compute efficiency
performances of the EDCs by generating a quality-of-service variable via TOPSIS.
Şirin (2017) has conducted a panel data analysis to understand the factors affecting
the cost of Turkish EDCs. Petridis et al. (2019) have proposed a two-stage approach
(network DEA model and directional distance function integration) to evaluate the
profit efficiency of EDCs.

3 An Overview of the Turkish Electricity Market

There are two main entities in the Turkish electricity market, namely Turkish Elec-
tricity Generation and Transmission Co. (TEAS) and Turkish Electricity Distribution
Co. (TEDAS). TEAS is responsible for the generation, transmission and wholesale of
electricity while TEDAS undertook the distribution activities (Karahan and Toptas,
2013). TEDAS was separated into regional distribution companies over time. The
liberalisation of the electricity distribution sector began in Turkey in 2004 in line
with the harmonisation with EU rules, and the government transferred its operations
to 21 private EDCs. As Bagdadioglu & Odyakmaz (2009) remark, liberalisation aims
to ensure the development of a financially sound and transparent electricity sector
that operates in a competitive environment under provisions of civil law to ensure
delivery of sufficient, high quality, low-cost and environmentally friendly electricity
that is subject to autonomous regulation and market supervision. Currently, 21 private
EDCs’ tariffs are regulated by Energy Market Regulatory Authority (EMRA). With
approximately 45 million customers and 229.59 billion kWh of electricity sales across
Table 1 Example research on EDCs’ performance
72

Author Sample from Country/ies Model Inputs/Outputs

(Period)
Hjalmarsson and Veiderpass (1992) Sweden (1970–1986) DEA Inputs (3): Hours worked by all
employees, low voltage power lines,
high voltage power lines, transformer
capacity
Outputs (4): Low/high voltage
electricity received by customers, # of
low/high voltage electricity customers
Milliotis (1992) Greece DEA Mixed Input and Output (8): Length of
line, transformer capacity, general
expenses, administrative # of
employee, technical # of labor, # of
customer, electricity delivered, total
area
Pollitt (1995) 14 countries (Asian, Australia DEA – OLS Inputs (3): # of employees, network
USA, EU) (1990) extension, transformer capacity
Outputs (5): # of customer, residential
sales, non-residential sales, service
area, max. demand
Bagdadioglu et al. (1996) Turkey (1991) DEA Input (5): # of employees, transformer
capacity, length of line, general
expenses, energy losses (%)
Outputs (4): # of customer, electricity
delivered, max. demand, service area
Yunos and Hawdon, (1997) Malaysia (1975–1990) DEA (Malmquist Productivity Index) Inputs (4): Transformer capacity, # of
employee, energy loss(%), public
generation capacity factor (%)
Outputs (1): Electricity delivered
Y. Cinar and T. Kaya

(continued)
Table 1 (continued)
Author Sample from Country/ies Model Inputs/Outputs
(Period)
Kumbhakar and Hjalmarsson (1998) Sweden (1997–1990) DEA Inputs Model 1 (4): # of employees,
transformer capacity, kilometers of
high/low voltage lines. Model 2 (2): #
of employees, transformer capacity
Outputs Model 1 (2): Low/high voltage
electricity received by the customers.
Model 2 (5): Low/high voltage
electricity received by the customers,
kilometres of high/low voltage lines,
transformer capacity
Zhang and Bartels (1998) Australia, New Zealand and DEA–Monte Carlo Simulation Inputs (3): # of employees, length of
Sweden lines, transformer capacity
Outputs (1): # of customers
Measuring Efficiency and Productivity Change …

Pacudan and Guzman(2002) Philippines (1996–2001) DEA Inputs (3): # of employees, energy loss
(%), length of line
Outputs (3): # of customers, service
area, sales to customer
Resende (2002) Brazil (1997–1998) DEA Inputs (3): # of employees,
transformers capacity, network
extension
Outputs (4): Concession area, # of
customer, industrial sales,
non-industrial sales
(continued)
73
Table 1 (continued)
74

Author Sample from Country/ies Model Inputs/Outputs

(Period)
Jamasb and Pollitt (2003) Italy, Norway, Portugal, Spain, DEA–COLS – SFA Inputs Model 1 (1): Total cost. Model 2
Netherlands (1997–1999) (3): Operating cost, length of line,
energy loss (%)
Outputs Model 1 (3): Electricity
delivered, # of customers, length of
line. Model 2 (2): Electricity delivered,
# of customers
Estache et al (2004) South America (1994–2001) DEA – SFA Inputs (3): # of employees, transformer
capacity, length of line
Outputs (3): # of customers, electricity
delivered, service area
Ghaderi et al. (2006) Iran (2000–2004) DEA-COLS-PCA Input (3): Length of line, transformers
capacity, # of employees
Output (2): # of customers, electricity
delivered
Hirschhausen et al. (2006) Germany (2001) DEA – SFA Inputs (3): # of employees, capital,
peak load capacity
Outputs (2): Electricity delivered, # of
customers
(continued)
Y. Cinar and T. Kaya
Table 1 (continued)
Author Sample from Country/ies Model Inputs/Outputs
(Period)
Von Hirschhausen et al. (2006) Germany (2001) DEA-SFA Inputs Models 1, 2, 7 (3): # of
employee, length of line
Model 3 (3): # of employee, length of
line, peak load capacity
Model 4 (3): # of employee, length of
line, turnover
Model 5 (3): # of employee, length of
line, energy loss
Model 6 (2): # of employee, length of
line
Output Model 1 (2): Sales to customer,
# of customers
Models 2, 3, 4, 5 and 6 (3): Sales to
Measuring Efficiency and Productivity Change …

customer, # of customers, inverse

density index
Model 7 (3): Sales to customer, # of
industrial customers, # of households
customers
Hess and Cullmann (2007) East and West German (2005) DEA – SFA Input (2): # of employees, length of
line
Outputs (3): Electricity delivered, # of
customers, inverse density index
Cullmann and Hirschhausen, 2008a, b) Eastern Europe (2002) DEA – FDH Inputs (2): # of employees, length of
line (capital)
Outputs (2): Electricity delivered, # of
customers
75
76 Y. Cinar and T. Kaya

Turkey in 2019, EDCs together form one of the largest organisations in the country
with 56,903 employees (for more details of the reform program and the background
of the Turkish electricity sector, see EMRA Electricity Market Development Report,
(2019).
Turkey has to import almost %70 of its primary energy. In 2019, electricity imports
reached 2,211,506 MWh while exports are 2,788,667 MWh. Industrial and commer-
cial consumption comprised 41.14% and 28.38%, respectively, of total electricity
consumption, whereas the household consumption proportion was 24.56%.

4 Methodology

Several numerical methods have been applied to assess the distribution performance
of EDCs in different countries. In these studies, parametric and non-parametric
methods, such as DEA, OLS, Malmquist Productivity Index (MPI), Stochastic
Frontier Analysis (SFA), and panel data analysis are generally utilised.

4.1 Efficiency Analysis and DEA

Data Envelopment Analysis (DEA) is a multi-factor efficiency analysis method for

measuring the relative efficiencies of a homogenous set of decision-making units
(DMUs, here EDCs). DEA was first introduced by Charnes et al. (1978).
Formally, consider that n DMUs {1, …,n} and each decision-making unit j for
j = 1, 2, . . . , n uses m different inputs, xi j , for i = 1, 2, . . . , m and produces s
different outputs yr j for r = 1, 2, . . . , s. The optimal efficiency score (φ) for oth
DMU can be obtained by the following linear programming model (1) and its dual
form (2):

Maxφ0
s.t.

n
λ j xi j ≤ xio
j=1 (1)

n
λ j yr j ≥ φo yr o
j=1

λj ≥ 0
Measuring Efficiency and Productivity Change … 77

Minφ0
s.t.

n
λ j yr j ≥ yr o
j=1 (2)

n
λ j xi j ≤ φo xio
j=1

λj ≥ 0

Here λj is a vector of weights assigned to each DMUs (EDCs). The assumptions

made on this vector determine the shape of the efficient frontier (envelopment) and
the production return to scale. Both are valid under the constant return to scale
(CRS) assumption. Coelli et al. (2005) has modified the CRS linear programming
to account for variable return to scale (VRS) by adding convexity constraint. The
convexity constraint ensures that an inefficient firm is only “benchmarked “ against
firms of a similar size.
The above problem is run n times to compute the relative efficiency scores for each
DMU in the sample. Then, the obtained value of φ represents the efficiency score
of j-th DMU. This unit is considered efficient if φ = 1; otherwise, it is considered
inefficient.
The CRS efficiency score represents the overall efficiency, which measures ineffi-
ciencies due to the input and output configuration as well as the size of the operations.
The VRS efficiency score represents pure technical efficiency—a measure of effi-
ciency without scale efficiency (Avkiran, 1999). The ratio of CRS and VRS provides
a measurement of the scale efficiency of DMUs (Fare et al., 1985). It is possible
that a firm is both technically and allocatively efficient but scale of operation of the
firm may not be optimal (Coelli et al. 2005). Hence, it shows if the DMUs are scale
efficient or not.

4.2 Malmquist Productivity Index

One of the advantages of DEA is that it consists of methodologies for utilising

techniques in calculating the MPI for cross-period productivity analysis, assuming
that the underlying technology is CRS. Caves et al. (1982) demonstrate that the
efficiency change of two observations can be measured as the ratio of the distance
of these observations to a common technology via MPI. Fare et al. (1992, 1994,
1997) calculate MPI (distance from observation at q period to the technology of
p period) based on inputs and outputs in between the referenced p period and the
following q period. Thus, the measure of the change in efficiency between p and
q periods and the amount of frontier shift (a measure of technological change) are
measured simultaneously. Calculation of four different functions that present the
78 Y. Cinar and T. Kaya

distances between observations and technologies at p and q periods is required.

Generally, mathematical models by Fare et al. (1997), who have redefined classical
DEA formulations for cross periods, are utilised while defining these functions.
The model that presents the efficiency of oth DMU at time of q against an efficient
frontier (technology) at period p is as follows:

Minφop,q
s.t.

n
q q
λ j yr j ≥ yrpo
j=1 (3)

n
q q p
λ j xi j ≤ φop xio
j=1
q
λj ≥ 0

Let’s assume that p and q stand for time t and time t + 1, respectively in model
(3). In this model, q substituting p = q = t in the optimisation problem yields an
efficiency value for firms in period t relative to the period t frontier,  t,t . The relative
efficiency of a DMU in period t + 1 compared with the t + 1 frontier, φjt+1,t+1 , is
calculated by setting p = q = t + 1. This permits the calculation of the efficiency
of a firm in period t relative to the period t + 1 frontier: φjt,t+1 . Finally, when p = t
+ 1 and q = t, we can find the efficiency of a DMU in t + 1 relative to the frontier
at t: φjt+1,t (Yunos & Hawdon, 1997). It is worth noting that, to compute the results
of defined functions for all periods and observations, n(3t - 2) linear programming
models should be set and solved (n and t stand for the number of observations and
number of periods, respectively).
In Fig. 1, if K  is observed under CRS technology in period p, and K is observed
under CRSt in the following period q, then the efficiency change, in this case, is
represented as follows:

Fig. 1 Efficient frontiers

and efficiency levels
determined by DEA relative
to different returns to scale in
different periods
Measuring Efficiency and Productivity Change … 79

q,q
φo y K /y E
E f f iciency Change = p, p = (4)
φo y K  /y E 

In Eq. (4), yo stands for the level of output for DMUo . Additionally, Eq. (5)
measures the technological change level.
 p,q p, p 1/2  1/2
φo φo y K /y E  y K  /y E 
T echnological Change = q,q = q, p = x (5)
φo φo y K /y E y K  /y E

MPI is expressed as follows:

M P I = E f f iciency Change × T echnological Change (6)

4.3 Tobit Analysis

To explain efficiency results and determine the factors which have the potential to
affect efficiency, Tobit models are utilised. Although regression models are utilised
to predict the behaviour of dependent variables, least square regression results have
deviations in case the dependent variable is censored by nature. Hence, the Tobit
model presented by Tobin (1958) is frequently utilised to explain efficiency scores
in DEA literature for cases that include limited dependent variables. In both the
electricity production and electricity distribution sectors, the drivers of the DMUs’
efficiencies are determined by Tobit regression models (Meibodi, 1998; Olatubi &
Dismukes, 2000; Murillo-Zamorano & Vega-Cervera, 2001; Domah, 2004; Barros
& Peypoch, 2008; Yang & Pollitt, 2009; Çelen, 2013; Pérez-Reyes & Tovar, 2009).
The Tobit model is explained based on Ogunyinka and Ajibefun’s (2004) study,
as follows:
⎧ n
⎨  β X + u , if L <  β X + u < U
n
j j j j j j j j
T I E j = j=1 j=1 (7)
⎩
0 , otherwise

TIE j is the technical inefficiency measure for each EDC and is computed by 1
– the CRS efficiency score of the related EDC. X j is a vector of an environmental
factor for the j-th EDC; βj is a degree of change in the outcome variable; uj represents
residuals distributed normally and independently with a mean of 0 and a common
variance. Finally, L j and U j are the distribution’s lower and upper censoring points,
respectively. Green (1995) remarks that since TIE scores approach 0, Eq. (7), which is
between 0 and 1, represents how variables affect the inefficiency levels of technically
inefficient EDCs statistically.
80 Y. Cinar and T. Kaya

Simar and Wilson (2007) expose that Tobit models have some limitations. Corre-
lations between efficiency scores and environmental factors lead to the inconsis-
tency problem of estimators. To overcome this inconsistency problem, correlation
coefficients are checked and scores are minimised.

5 Application

5.1 Data and Sample

In the literature, there are a variety of variables utilised as inputs and outputs, which
are determined by considering the international experience and data availability.
Although the priority of EDCs is to make a profit, the inclusion of only process-
oriented inputs and outputs rather than financial variables makes this study more
practical and process-oriented. The electricity distribution data set consists of 21
Turkish EDCs over a five-year (2015–2019) period. These EDCs are; ADM, Akdeniz
(AD), Akedas (AK), Aras (AR), Baskent (BS), Bogazici (BI), Camlıbel (CB), Coruh
(CR), Dicle (DC), Firat (FR), GDZ, IstAnadolu (IA), Kayseri (KC), Meram (MC),
Osmangazi (OA), Sakarya (SK), Toroslar (TO), Trakya (TR), Uludag (UL), Vangolu
(VG) and Yesilirmak (YS). As for input and output variables, data were collected
from EMRA, and this recent data set covers 100% of the total number of utilities.
In total, 735 observations were obtained. Summary statistics of the data are
presented in Table 2.

5.1.1 Input and Output Variables

The input and output selection process is an important factor in efficiency measure-
ment. As Jamasb & Pollitt (2001) note, DEA results can be sensitive to model inputs
and outputs. In our literature review (Table 1), while the most utilised inputs are
number of employees, quality, capital, total and operational expenditures, energy
losses, length of line and transformer capacity (as an indicator of investment); outputs
are electricity delivered, number of customers, service area, industrial and non-
industrial sales and quality measures. Jamasb & Pollitt (2001) additionally outline
the most widely utilised inputs and outputs for measuring the efficiencies of EDCs.
Based on their study, while network length, transformer capacity and the number of
employees are among the most commonly utilised inputs, the most widely utilised
outputs are determined to be the number of customers and electricity delivered.
Considering the aim of our study and data availability, we utilised four inputs in
our analysis: namely, number of employees, length of line, transformer capacity and
quality variables. Transformers’ capacity and length of line variables are considered
to be representative of investment.
Table 2 DEA models used to assess EDCs performance (Continue)
Author Sample from Country/ies (Period) Model Inputs/Outputs
Cullmann & Hirschhausen Poland (1997–2002) DEA-FDH-SFA Inputs (2): # of employees, length of
(2008b) line,
Outputs (3): # of customers, electricity
delivered, inverse density index
Sadjadi & Omrani (2008) Iran (2004) Robust DEA Inputs (3): # of employees, length of
line, transformer capacity
Outputs (2): Sales to customer, # of
customers
Azadeh et al (2009) Iran (2000–2004) DEA–COLS–SFA–PCA–NT Inputs (3): Length of line,
transformers capacity, # of employees
Outputs (2): # of customers, sales to
customer
Perez-Reyes & Tovar (2009) Peru (1996–2006) Malmquist Prod. Index, Tobit Inputs Model 1 (4): # of employees,
Measuring Efficiency and Productivity Change …

length of line, energy loss (%), # of

substations. Model 2 (3): # of
employees, energy loss (%), capital
Outputs (2): Sales to customer, # of
customer
Growitsch et al (2009) 8 EU Countries and UK (2002) SFA Inputs (Model 1): Total expenditures
(TOTEX). Model 2: Customer
minutes lost (CML)
Outputs (2): # of customers, number
of energy units supplied
(continued)
81
Table 2 (continued)
82

Author Sample from Country/ies (Period) Model Inputs/Outputs

Ramos-Real et al. (2009) Brazil (1998–2005) Malmquist Prod. Index Inputs (3): # of employees, length of
line, energy loss (%)
Output (2): Electricity delivered, # of
customers
Khodabakhshi (2010) Iran (2000) DEA (Super-efficiency) Input (3): Length of line, transformer
capacity, # of employee
Output (2): Units delivery and service
area
Santos et al. (2011) Portuguese (2002–2006) Malmquist Prod. Index Inputs (3): Operational expenditure,
length of line, transformers cap
Outputs (2): Electricity delivered,
quality (interruption time)
Çelen & Yalçın (2012) Turkey (2002–2009) FAHP-TOPSIS-DEA Inputs (3): # of employees, length of
the line, transformer capacity
Outputs (3): Quality, electricity
delivered, # of customers
Growitsch et al. (2012) Norway (2001–2004) Factor Analysis – SFA Input(1): Social cost
Outputs (2): # of customers, electricity
delivered
Other Variables: 40 weather variables
and 38 geographic variables
(continued)
Y. Cinar and T. Kaya
Table 2 (continued)
Author Sample from Country/ies (Period) Model Inputs/Outputs
Celen (2013) Turkey (2002–2009) DEA-Tobit Input Model 1 (4): # of employees,
length of line, transformer capacity,
quality
Input (Model 2) (5): # of employees,
length of line, transformer capacity,
quality, energy loss
Output (2): Electricity delivered and #
of customers
Kuosmanen et al. (2013) Finland (2012) DEA-SFA-StoNED-Monte Carlo Inputs (2): Length of line, total cost,
proportion of underground cables
Outputs (3): Electricity delivered,
length of line, # of customers
Mullarkey et al. (2015) Ireland (2008) DEA Input (4): # of employee, length of
Measuring Efficiency and Productivity Change …

line, transformer capacity, location

Output (6): Gross/net energy
consumption, # of customer,
diagnostic parameter, service area, line
density
Xavier et al. (2015) Brazil (2006–2007) Two-Stage DEA Inputs (3): # of employees, length of
line, transformer capacity
Outputs (2): # of customers, electricity
delivered
Mirza et al. (2017) Pakistan (2006–2013) Malmquist Prod. Index Inputs (3): Energy losses, peak load,
length of line
Outputs (2): Electricity delivered,
growth in # of customers
(continued)
83
Table 2 (continued)
84

Author Sample from Country/ies (Period) Model Inputs/Outputs

Bongo et al. (2018) Philippines (2015) DEA- Super Efficiency Inputs (2): Purchased electricity,
length of lines
Outputs (3): # of consumers,
electricity delivered and energy loss
(%)
Fallahi et al. (2019) Iran (2005–2014) DEA Inputs (4): Transformer capacity, # of
employees, length of line, energy
losses (%)
Output (3): Electricity delivered, # of
customers, service area
Petridis et al. (2019) Turkey (2011) Network DEA Inputs (5): # of employees, Net
Consumption, # of transformers,
length of line, transformer capacity
Intermediate Outputs (1): Electricity
delivered
Undesirable Output (2): Quality,
energy losses
Outputs (2): # of customers, number
of Towns/Villages
Tavassoli et al. (2020) Iran (2017) Network DEA Inputs (4): # of employee, fuel oil,
natural gas, gas oil
Outputs (7): Electricity delivered,
export, sale to neighbour, sale to big
industry, energy loss (%), service area,
sales to customer
Y. Cinar and T. Kaya
Measuring Efficiency and Productivity Change … 85

In our literature survey, the number of employees is found to be an input variable

in 40 of the 65 studies. Length of line is utilised in 35, transformer capacity is
utilised in 27 and quality is utilised in four. Energy losses are widely considered
to be an input (18 of the 65 studies used these as a variable) in literature in recent
years (Fallahi et al., 2019; Şirin, 2017; Simab & Hadgifam, 2010; Tavassoli et al.,
2015; Tavassoli, 2020; Mirza et al., 2017). Dashti et al. (2013) state that the ratio
of energy losses has a significant impact on distribution system efficiency, whereas
Bongo et al. (2018) express energy loss as the most common factor that leads to the
inefficiency of DMUs. However, Fallahi et al. (2019) reveal that dropping the losses
from the calculation is less impactful to the results, as indicated by two different
models’ 99% correlation coefficient. Additionally, Tavassoli et al. (2020) have found
that there are no major efficiency ranking changes observed in EDCs after removing
the energy loss variable from the model. Çelen (2013), who considers energy loss
to be an environmental factor, exposes that EDCs operating in Turkey with higher
energy loss ratios have unexpectedly higher efficiency performances.
In line with these observations, we used the energy loss variable as an environ-
mental factor. Electricity delivered is utilised as the first output since the purpose
of the EDCs is to supply electricity. The second output utilised in this study is the
number of customers who are provided with power over a year. Based on our litera-
ture survey, electricity delivered and the number of customers are the most common
output variables (used in 37 and 49 of 65 studies, respectively) used in EDC produc-
tivity measures. Incorporating this variable provides a magnitude of the number of
towns and villages, as it reflects the total users in both villages and cities (Petridis
et al., 2019). Although a strong correlation (87% on average) is exposed between the
number of customers and the electricity delivered for the five-year period, moderate
correlations are also observed in the analysed period. Furthermore, EDCs with few
customers per km2 are more efficient. Hence, as a variable of the customer density
function, the number of customers is considered as output.

5.1.2 Environmental Factors

We additionally consider the effect of environmental factors on EDCs’ efficiency

in this study. The environmental factor data set has been collected from various
databases. Urbanisation levels of EDC regions are determined by the “Measuring
Urbanisation Level in Turkish Districts by Population Density and Urban Functions”
(2019) report from the Ministry of Industry and Technology of Turkey. The districts’
urbanisation is our first environmental factor and categorised into four levels: Levels
1 and 2 represent urban cities, whereas 3 and 4 signify rural areas. Urbanisation levels
of EDC regions were specified by weighting the urbanisation levels of cities with the
number of customers served in each city. Based on this calculation methodology, four
of 21 districts are considered urban. Korhonen & Syrjänen (2003) utilise urbanisation
as an environmental factor, and the results of their study indicate that the proportion
of efficient units and the average efficiency of the units is on the same level in urban
86 Y. Cinar and T. Kaya

companies as in the whole population. As with urbanisation, the effect of districts’

development status on EDCs’ efficiencies is considered an environmental factor.
The development index of EDC regions is the second environmental factor and
calculated by utilising the “Research of Development Grading Research of Cities
and Districts—SEGE” (2013). This report presents the development scores of each
city. As it is known which city is served by which EDC, development indices of EDC
regions are computed by weighting the city indices with the number of customers
served in related cities. The third and fourth environmental factors are commercial
and industrial electricity delivered (MWh) in total and network loss (MWh), respec-
tively. The data set for both variables covers annual data from EMRA “Electricity
Market Development Reports” (2015–2019). Von Hirschhausen et al. (2006) have
utilised the number of household customers as an output in their model, and they
imply that larger utilities generally have more industrial customers compared to
household customers. Since industrial customers obtain lower prices and thus lead
to a lower average per-unit revenue, this has an adverse effect on the efficiency score.
Therefore, in this study, commercial and industrial electricity delivered is considered
an important factor in EDCs’ efficiency.
The final environmental factor is the service area of EDCs in km2 . The geograph-
ical dispersion of customers among different EDCs may be an important factor
impacting efficiency in the context of Turkey, but it is non-controllable. A moderate
correlation between length of line and service area (0.66) proves this geographical
impact. The service area is employed as an output variable in the literature to reflect
the difficulty of meeting customer services over a less densely populated area (Bagda-
dioglu et al., 1996; Estache et al., 2004; Fallahi et al., 2019; Khodabakhshi, 2010;
Lins et al., 2007; Mullarkey et al., 2015; Pacudan & Guzman, 2002; Pollitt, 1995;
Tavassoli et al., 2020). Jamasb & Pollitt (2003) additionally suggest that the service
area variable may be potentially useful. However, since EDC districts have varying
types of geographical characteristics in Turkey, and EDCs have no choice to select
or change service areas, it is considered a structural environmental factor.

6 Results

6.1 DEA Results

The DEA analysis was conducted for each year from 2015 to 2019. Hess and Culmann
(2007) emphasise that the electricity sector is subject to the legal duty to serve all
customers in a predefined service territory, and therefore, utilising the input-oriented
approach is appropriate, assuming that a firm is producing a specific output while
minimising the inputs. The efficiency scores were calculated using EMS software
developed by Holger Scheel. Table 3 presents the results of the input-oriented model
in technical CRS, VRS and scale efficiency (SE).
Measuring Efficiency and Productivity Change … 87

Table 3 Descriptive statistics of EDCs

Variable Input/Output/ Mean Standard Min Max
Description Environmental Deviation
Factor
# of Employee Input 2627 1396 873 7098
(Person)
Length of Line 53,864 233,304 19,439 113,814
(Km)
Transformer 7408 4538 2216 17,850
Capacity (MVA)
Quality 1744 1367 139 10,131
(Interruption
Duration per
Customer – min)
Electricity Output 10,425,499 7,315,400 1,606,617 27,722,101
Delivered (MWh)
# of Customers 2,017,406 1,201,738 599,921 5,134,980
Urbanization Environmental 2.5 0.5 1.5 3.4
Level* Factor
Energy Loss 1,050,187 122,700 164,074 5,509,736
(MWh)**
Development 2.7 1.6 0.5 6.5
Status***
Commercial + 7,787,515 6,555,161 937,312 25,560,789
Industrial
Electricity
Delivered (MWh)
Service Area 37,414 20,010 1969 77,273
(Km2 )
* Based on Ministry of Industry and Technology Research Report data (2019). Scores: 1 to 2 means

more urban. After score 2, region is accepted as more rural. **Energy losses per EDC. ***Based on
Ministry of Development, Research of Development Grading Research of Cities and Districts report
data (2013). Low score means undeveloped regions while a high score is presenting developed.
88 Y. Cinar and T. Kaya

The yearly time series demonstrates that average efficiency scores decreased
slightly from 2015 to 2019. In 2018, a recovery period was observed. However, in
line with the economic conditions in Turkey, efficiency scores continued to decline
in 2019 (Fig. 2).
The individual efficiency scores of EDCs across returns to scale measures are
reported in Appendix 1.
Figure 3 presents some implications for the optimal scale of economics, which
indicates that a higher scale means higher efficiency, considering the positive slope
of the linear CRS line. This means that increases in the scale, which is measured and
indicated here by electricity delivered, the efficiency trend level is rising. Figure 3 also
shows gaps between CRS and VRS scores which is the reason for scale inefficiencies.
Here the gaps are higher in lower scales (<53.000). A level of 5.3 million MWh of
electricity delivered seems to be a breakpoint of optimal for scale efficiency for the
electricity distribution sector in Turkey. Seven EDCs are operating under this level,
which means that these companies are operating on an improper scale.

100% CRS VRS Scale Eff.

94.7%
Average Efficiency (%)

95%
92.8%
90.6% 90.7%
90% 88%

85%
82%

80%

75%
2015 2016 2017 2018 2019

Fig. 2 CRS, VRS and scale efficiency changes (2015–2019)

Fig. 3 Relation of total electricity delivered and efficiency (CRS and VRS)
Measuring Efficiency and Productivity Change … 89

Table 4 Distributional characteristics of 5 years average efficiency scores of 21 EDCs

Mean Std. Dev Min Max
Technical Eff. (CRS) 0.85 0.16 0.43 1
VRS 0.92 0.13 0.46 1
SE 0.93 0.09 0.63 1

During the execution of the DEA model, DMUs allow their more advantageous
variables (input and output) to have higher importance or weightings to achieve
higher efficiency scores. Table 4 illustrates how these weights have been distributed
between input factors and output factors by the DMUs’ EDCs while they reach their
efficiency scores by years.
Table 4 additionally presents that the two inputs which are assigned the most
weight are transformer capacity (58.93%) and the number of employees (20.85%)
by mean values, representing two well-known production factors: capital and labour,
respectively. Between these two output factors, the number of customers is nearly 1.5
times more important than electricity delivered in terms of assessing weights while
achieving efficiency scores.
An inefficient EDC can become efficient by taking some efficient EDCs as refer-
ences to itself and changing its’ inputs and/or outputs to reach them. DEA provides
us the reference set information for each inefficient unit, and how to improve perfor-
mance. Appendix 2 illustrates the reference set of inefficient EDCs. The most set
peers are UL (43 times), IA (27 times) and YS (25 times). While UL and IA have
been standing in peer groups more frequently since 2017, IA was the dominant peer
EDC until 2017. Although there are seven scale-efficient EDCs every year, the differ-
entiating attributes of UL and IA are that they have higher customer density, lower
energy loss and higher quality of electricity.

6.2 Tobit Results

To provide overall significance for the Tobit analysis, likelihood ratio test statis-
tics were checked. We utilised the built-in coefficient testing procedures to test the
exclusion of all of the explanatory variables. The Jarque–Bera test was applied if
the residuals were normally distributed. E-Views package program was utilised to
estimate Tobit models. Results expose an appropriate model fit, which means the null
hypothesis of the homoscedasticity and normality are not rejected. Table 5 presents
the descriptive statistics of the Tobit regression model variables. Wide divergences
are observed in EDC inefficiencies and environmental factors.
Table 6 presents the estimation results of these models. Based on the results, we
have two major factors that affect efficiency significantly: energy loss and commer-
cial and industrial electricity delivered. Regarding energy loss, it seems that there is
no consensus on how the energy losses impact the efficiency of EDCs. One of the
90 Y. Cinar and T. Kaya

Table 5 Weights given to the factors by EDCs (in average)

Inputs Outputs
Number of Length of Transformer Quality Electricity Number of
Employee Line (%) Capacity (%) (%) Delivered Customers
(%) (%) (%)
2015 13.71 6.90 63.48 15.90 42.48 57.52
2016 21.24 8.52 56.86 13.33 41.05 58.95
2017 19.90 6.71 63.24 10.14 33.67 66.33
2018 24.05 10.14 55.71 10.10 40.33 59.67
2019 25.33 10.38 55.38 8.90 39.67 60.33
Average 20.85 8.53 58.93 11.68 39.44 60.56
(%)

Table 6 Descriptive statistics of Tobit regression model variables

Variable Technical Urbanization Development Commercial Service Energy
Inefficiency Level Level of and Area Loss
(TEI) (%) Region Industrial (m2 ) (MWh)
Electricity
Delivered
(MWh)
Mean 14.5 2.5 2.68 7,787,515 34,415 1,050,187
Minimum 0 1.5 0.5 937,312 1969 129,955
Maximum 56.8 3.4 6.5 25,560,789 77,273 6,153,617
Std. 15.7 0.5 1.6 6,555,161 20,010 1,227,000
Deviation

reasons for that may be that energy losses are considered input, output, or environ-
mental factors. For instance, Fallahi et al. (2019) and Tavassoli et al. (2020) reveal
that dropping the energy losses from the calculation has no major impact on effi-
ciency scores and rankings. Mirza et al. (2017) have found that electric utilities are
incurring more technical and non-technical losses due to the length of their networks,
which could improve productivity and refers indirectly to the customer density. Çelen
(2013) emphasises that energy loss is a significant problem that needs to be solved,
especially in the eastern part of Turkey. However, the research has found an unex-
pected result that claims that higher energy loss and theft ratios have demonstrated
higher efficiency performances. In line with the literature, our finding represents that
higher energy losses have a significant negative effect on efficiency (Table 7).
As stated previously, industrial customers obtain lower prices with a correspond-
ingly lower average per-unit revenue; furthermore, it is a member of the customer
density function, and they have the potential to affect efficiency. We have found that
higher commercial and industrial electricity delivered in the region means higher effi-
ciency. As Pombo and Taborda (2006) emphasise, the expected sign of this variable is
positive because greater density allows the utility to exploit its network’s economies
Measuring Efficiency and Productivity Change … 91

Table 7 Estimation results of Tobit regression model

Statistically Significant Years
Environmental Factors that 2015 2016 2017 2018 2019 5 Years Average
Affect Inefficiency
Urbanization
Development Status of
Region
Commercial and Industrial (−) (−) (−) (−) (−)
Electricity Delivered (MWh)
Energy Loss (MWh) (+) (+) (+) (+) (+)
Service Area (km2 )
(−): Statistically significant negative effect on inefficiency; (+): Statistically significant positive
effect on inefficiency.

of scale; moreover, profitability encourages new investments that lead to increases in

industrial sales, which cover the utility against the risk of bad debtors within the resi-
dential sector in the regulated market. This result may have similarities with customer
density in that region. When the customer density is high, electricity consumption per
km2 is also high because of less input usage (employee, network length, investment,
etc.). Reyes and Tovar (2009) further exhibit that companies with a high proportion
of low-voltage network residential customers are less efficient than companies with
a preponderance of medium-voltage industrial and commercial users.
β coefficients of commercial and industrial electricity delivered and energy loss are
0.002 and 0.007, respectively. This means that a 100,000 MWh increase in commer-
cial and industrial electricity distribution leads to a 0.2% increase in EDCs’ efficiency
scores, whereas a 100,000 MWh energy loss reduction increases the efficiency by
0.7%.

6.3 Malmquist Productivity Index Results

To analyse the changes in efficiency and to observe the technological change (frontier
shift) over time, MPIs were calculated for the years 2015–2019. Figure 4 represents
a summary of the annual means of MPIs obtained. Here we take 2015 as a base year
and set the index equal to 1.
MPI results indicate that on average, productivity has not changed significantly
over the five years. From 2015 to 2019, the overall technical efficiency of the distri-
bution sector was reduced slightly. In terms of technological change, it fluctuated
and reached over 1.00 again in 2019. Therefore, no crucial change was observed.
Additionally, the distribution of MPIs exposes that there was a slight rise in the
first two years and a reduction in productivity from 2018 to 2019. The reason of
the productivity increase in the first two years (until 2017) is due to the increase on
92 Y. Cinar and T. Kaya

Fig. 4 A summary of
EDC-based average MPI
results is presented in
Appendix 3

technological change. In line with the dramatic reduction on technological change

(Techch) after 2017, MPI started to decrease too although the efficiency goes up.
Consequently, the electricity distribution sector in Turkey may be summarised as a
steady sector after 2015. However, in line with the economic growth in Turkey, we
see that the efficiency of EDCs was affected by the gross domestic product (GDP)
reduction in Turkey in after 2016.

7 Conclusions

In this chapter, the efficiency of the EDCs in Turkey was analysed to assess how the
efficiencies have changed in recent years. The factors which led to the inefficiencies
have also been examined. Based on the DEA results, CRS and VRS efficiencies
of EDCs have been reduced slightly between 2015 and 2019 while MPI has been
stable. In this period, average efficiency rate of EDCs is 85.5%. Both DEA and MPI
results suggest that the efficiency of the electricity distribution sector aligns with the
economic conditions of the country Over the five-year analysis period. DEA further
demonstrates that transformer capacity and the number of employee variables are
the crucial inputs that are key in reaching efficiency scores.
In terms of environmental factors that affect efficiency, we have found that higher
commercial and industrial electricity delivered in regions means higher efficiency.
There is no compromise on how the energy losses impact the efficiency of EDCs
in literature; however, findings demonstrate that the energy loss has a significant
positive effect on inefficiency. This means that the lower energy loss results in higher
efficiency. This was an expected result since the electricity losses of the regions
remain high (11% on average in 2019) even after the liberalisation of the sector
in 2013. However, we found that the urbanisation level and development status of
regions have no role in EDCs’ efficiency in Turkey. This result emphasises that the
efficiency of EDCs are still affected by structural mechanisms such as energy loss and
commercialization of the regions other than social dynamics such as the urbanisation
of the regions.
Measuring Efficiency and Productivity Change … 93

DEA results also exhibit that the most inefficient EDCs have proportionally lower
economies of scale (<5.3 million MWh per year). Thus, the minimum level of
5.3 million MWh electricity delivered seems to be a breakpoint of optimal scale.
Decision-makers should be cautious about the decision to divide or combine regions
in the future.

Appendix 1

See Table 8.

Appendix 2

See Table 9.

Appendix 3

See Table 10.

 94

Table 8 Individual efficiency scores and returns to scale measures

EDC 2015 2016 2017 2018 2019
CRS VRS SE CRS VRS SE CRS VRS SE CRS VRS SE CRS VRS SE
ADM 0,88 0,99 0,88 0,72 0,73 0,97 0,79 0,91 0,87 0,69 0,74 0,93 0,67 0,72 0,94
AD 0,80 0,81 0,98 0,84 0,88 0,95 0,80 0,84 0,95 1,00 1,00 1,00 0,84 0,90 0,94
AK 0,86 1,00 0,86 0,76 0,91 0,84 0,70 0,91 0,77 0,70 0,93 0,75 0,68 0,96 0,71
AR 0,79 0,86 0,92 0,74 0,84 0,88 0,75 0,85 0,88 0,76 0,89 0,86 0,75 0,87 0,86
BS 0,90 1,00 0,90 1,00 1,00 1,00 1,00 1,00 1,00 0,85 1,00 0,85 0,78 0,92 0,85
BI 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00
CB 1,00 1,00 1,00 0,94 1,00 0,94 0,88 1,00 0,88 0,87 1,00 0,87 0,86 1,00 0,86
CR 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 0,95 1,00 0,95
DC 0,37 0,39 0,96 0,40 0,41 0,97 0,39 0,43 0,91 0,53 0,57 0,93 0,46 0,48 0,95
FR 1,00 1,00 1,00 0,85 0,94 0,90 0,77 0,92 0,84 0,72 0,85 0,85 0,70 0,83 0,84
GDZ 1,00 1,00 1,00 1,00 1,00 1,00 0,91 0,91 1,00 1,00 1,00 1,00 0,80 0,80 0,99
IA 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00
KC 0,81 1,00 0,81 0,87 1,00 0,87 0,74 1,00 0,74 0,76 1,00 0,76 0,79 1,00 0,79
MR 0,64 0,66 0,96 0,74 0,78 0,95 0,66 0,73 0,91 0,73 0,75 0,98 0,67 0,72 0,93
OG 0,88 0,90 0,98 0,94 0,98 0,96 0,76 0,83 0,91 0,92 0,98 0,94 0,76 0,86 0,89
SK 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00
TO 1,00 1,00 1,00 1,00 1,00 1,00 0,85 1,00 0,85 1,00 1,00 1,00 0,90 1,00 0,90
TR 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00
UL 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00
VG 0,55 0,87 0,63 0,54 0,87 0,63 0,56 0,88 0,64 0,58 0,89 0,65 0,62 0,96 0,65
YS 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00 1,00
Y. Cinar and T. Kaya
Measuring Efficiency and Productivity Change … 95

Table 9 Reference sets for inefficient EDCs

Inefficient EDC Year
2015 2016 2017 2018 2019
ADM IA, GDZ CR, IA, TO BI, UL IA, TO, UL UL, YS
AKDENİZ (AD) IA, TO, UL BI, IA BI, UL IA, UL UL, YS
AKEDAS (AK) SK, TO SK, UL UL UL SK, UL
ARAS (AR) CR, YS CR, IA, YS CR, YS CR, YS YS
BASKENT (BS) FR, IA Eff Eff IA, UL, YS UL, YS
CAMLIBEL (CB) Eff IA, YS CR, YS, UL CR, YS UL, YS
CORUH (CR) Eff Eff Eff Eff YS
DICLE (DC) IA, UL, YS CR, IA, TO, UL CR, UL IA, TO, UL SK, UL
FIRAT (FR) Eff CR, IA, TO CR, UL UL, YS IA, UL, YS
GEDIZ (GDZ) Eff Eff BI, TR, UL Eff BI, SK, UL
KCEDAS (KC) IA, UL, YS CR, IA, TO, UL CR, YS, UL IA, UL, YS UL, YS
MERAM (MR) IA, TR, UL BI, IA BI, UL IA, UL UL
OSMANGAZI OG) IA, TR, UL IA, SK, TO, UL UL, YS IA, UL UL
TOROSLAR (TO) Eff Eff UL Eff SK, UL
VANGOLU (VG) IA, YS CR, IA CR, UL UL, YS IA, YS

Table 10 EDC based

EDC EFFch TECHch TFPch
average MPI changes
ADM 0.936 1.015 0.950
AD 1.014 1.034 1.049
AK 0.942 1.017 0.958
AR 0.988 1.008 0.996
BS 0.965 1.020 0.984
BI 1.000 0.988 0.988
CB 0.963 1.002 0.965
CR 0.988 1.017 1.005
DC 1.052 0.997 1.049
FR 0.915 1.001 0.917
GDZ 0.945 1.008 0.953
IA 1.000 0.961 0.961
KC 0.996 1.020 1.016
(continued)
96 Y. Cinar and T. Kaya

Table 10 (continued)
EDC EFFch TECHch TFPch
MR 1.013 1.030 1.044
OG 0.964 1.049 1.011
SK 1.000 1.040 1.040
TO 0.973 1.007 0.980
TR 1.000 1.018 1.018
UL 1.000 1.049 1.049
VG 1.029 1.002 1.031
YS 1.000 1.019 1.019
Mean 0.984 1.014 0.998

References

Avkiran, N. K. (1999). An application reference for data envelopment analysis in branch banking:
helping the novice researcher. International Journal of Bank Marketing, 17(5), 206–220.
Azadeh, A., Ghaderi, S. F., Omrani, H., & Eivazy, H. (2009). An integrated DEA–COLS–SFA
algorithm for optimization and policy making of electricity distribution units. Energy Policy,
37(7), 2605–2618.
Bagdadioglu, N., Price, C., & Weyman-Jones, T. G. (1996). Efficiency and ownership in electricity
distribution: A non-parametric model of the Turkish experience. Energy Economics, 18(1–2),
1–23.
Bagdadioglu, N., & Odyakmaz, N. (2009). Turkish electricity reform. Utilities Policy, 17, 144–152.
Bagdadioglu, N., Price, C. W., & Weyman-Jones, T. (2007). Measuring potential gains from mergers
among electricity distribution companies in Turkey using a non-parametric model. The Energy
Journal, 28(2).
Barros, C. P & Peypoch, N. (2008). Technical efficiency of thermoelectric power plants. Energy
Economics, 30(6), 3118–3127.
Bobde, S. M., & Tanaka, M. (2018). Efficiency evaluation of electricity distribution utilities in India:
A two-stage DEA with bootstrap estimation. Journal of the Operational Research Society, 69(9),
1423–1434.
Bongo, M. F., Ocampo, İA., Magallano, Y. A. D., Manaban, G. A., & Ramos, E. K. F. (2018). Input–
output performance efficiency measurement of an electricity distribution utility using super-
efficiency data envelopment analysis. Soft Computing, 22(22), 7339–7353.
Borghi, E., Del Bo, C., & Florio, M. (2016). Institutions and firms’ productivity: Evidence from
electricity distribution in the EU. Oxford Bulletin of Economics and Statistics, 78(2), 170–196.
Caves, D., Christensen, L., & Diewert, E. (1982). The economic theory of index numbers and the
measurement of input, output and productivity. Econometrica, 50, 1393–1414.
Çelen, A., & Yalçın, N. (2012). Performance assessment of Turkish electricity distribution utilities:
An application of combined FAHP/TOPSIS/DEA methodology to incorporate quality of service.
Utilities Policy, 23, 59–71.
Celen, A. (2013). Efficiency and productivity (TFP) of the Turkish electricity distribution
companies: An application of two-stage (DEA&Tobit) analysis. Energy Policy, 63, 300–310.
Charnes, A., Cooper, W. W., & Rhodes, E. (1978). Measuring the efficiency of decision making
units. European Journal of Operational Research, 429–444.
Chen, W., Zhou, K., & Yang, S. (2017) Evaluation of China’s electric energy efficiency under
environmental constraints: A DEA cross efficiency model based on game relationship. Journal
of Cleaner Production, 164, 38–44.
Measuring Efficiency and Productivity Change … 97

Coelli, T. J., Rao, D. S. P., O’Donnell, C. J., & Battese, G. E. (2005). An Introduction to Efficiency
and Productivity Analysis. Berlin: Springer.
Cullmann, A., & Von Hirschhausen, C. (2008a). Efficiency analysis of East European electricity
distribution in transition: Legacy of the past? Journal of Productivity Analysis, 29(2), 155.
Cullmann, A., & Von Hirschhausen, C. (2008b). From transition to competition Dynamic efficiency
analysis of Polish electricity distribution companies. Economics of Transition, 16(2), 335–357.
Dashti, R., Yousefi, S., & Moghaddam, M. P. (2013). Comprehensive efficiency evaluation model
for electrical distribution system considering social and urban factors. Energy, 60, 53–61.
Domah, P. (2004). Technical efficiency in electricity generation-the impact of smallness and isolation
of island economies. CMI Working Paper 14, The Cambrigde-MIT Institute.
EMRA (Energy Market Regulatory Authority). (2015, 2016, 2017, 2018, 2019). Electricity market
development report. http://epdk.gov.tr/Detay/Icerik/3-0-0-102/yillik-rapor-elektrik-piyasasi-gel
isim-raporlari.
Estache, A., Rossi, M. A., & Ruzzier, C. A. (2004). The case for international coordination of
electricity regulation: Evidence from the measurement of efficiency in South America. Journal
of Regulatory Economics, 25(3), 271–295.
Fallahi, A., Fallahi, F., Sarhadi, H., Ghaderi, S.F., & Ebrahimi, R. (2019). Application of a robust
data envelopment analysis model for performance evaluation of electricity distribution companies.
International Journal of Energy Sector Management.
Fare, R., Grosskpf, S., Lovell C. A. K. (1985). The Measurement of Efficiency of Production.
Kluwer-Nijhoff, Boston, MA.
Färe, R., Grosskopf, S., Lindgren, B., & Roos, P. (1992). Productivity changes in Swedish phara-
macies 1980–1989: A non-parametric Malmquist approach. Journal of Productivity Analysis,
3(1–2), 85–101.
Fare, R., Grosskopf, S., & Lovell, C. A. K. (1994). Production frontiers. Cambridge University
Press
Färe, R., Grosskopf, S., & Norris, M. (1997). Productivity growth, technical progress, and efficiency
change in industrialized countries: Reply. The American Economic Review, 87(5), 1040–1044.
Ghaderi, S.F., Azadeh, M.A., & Omrani, H. (2006). An Integrated DEA-COLS-PCA model for
performance assessment and optimization of electricity distribution units. In 2006 4th IEEE
International Conference on Industrial Informatics, Singapore, pp. 236–241.
Green, W. H. (1995). LIMDEP Version 7.0 user’s manual. Econometric Software Inc., New York.
Growitsch, C., Jamasb, T., & Pollitt, M. (2009). Quality of service, efficiency and scale in network
industries: An analysis of European electricity distribution. Applied Economics, 41(20), 2555–
2570.
Growitsch, C., Jamasb, T., & Wetzel, H. (2012). Efficiency effects of observed and unobserved
heterogeneity: Evidence from Norwegian electricity distribution networks. Energy Economics,
34(2), 542–548.
Hess, B., & Cullmann, A. (2007). Efficiency analysis of East and West German electricity
distribution companies–Do the “Ossis” really beat the “Wessis”? Utilities Policy, 15(3), 206–214.
Hirschhausen, C. V., Cullmann, A., & Kappeler, A. (2006). Efficiency analysis of German electricity
distribution utilities – non-parametric and parametric tests. Applied Economics, 38(21), 2553–
2566.
Hjalmarsson, L., & Veiderpass, A. (1992). Productivity in Swedish electricity retail distribution.
The Scandinavian Journal of Economics, 193–205.
Jamasb, T., & Pollitt, M. (2001). Benchmarking and regulation of electricity distribution and
transmission utilities: Lessons from international experience in DEA Department of Applied
Economics. Cambridge: University of Cambridge.
Jamasb, T., & Pollitt, M. (2003). International benchmarking and regulation: An application to
European electricity distribution utilities. Energy Policy, 31(15), 1609–1622.
Korhonen, P. J., & Syrjänen, M. L. (2003). Evaluation of cost efficiency in Finnish electricity
distribution. Annals of Operations Research, 121(1–4), 105–122.
98 Y. Cinar and T. Kaya

Khodabakhshi, M. (2010). An output oriented super-efficiency measure in stochastic data envelop-

ment analysis: Considering Iranian electricity distribution companies. Computers & Industrial
Engineering, 58(4), 663–671.
Kumbhakar, S. C., & Hjalmarsson, L. (1998). Relative performance of public and private ownership
under yardstick competition: Electricity retail distribution. European Economic Review, 42(1),
97–122.
Kuosmanen, T., Saastamoinen, A., & Sipiläinen, T. (2013). What is the best practice for benchmark
regulation of electricity distribution? Comparison of DEA, SFA and StoNED methods. Energy
Policy, 61, 740–750.
Lins, M. P. E., Sollero, M. K. V., Calôba, C. M., & Da Silva, A. C. M. (2007). Integrating the
regulatory and utility firm perspectives, when measuring the efficiency of electricity distribution.
European Journal of Operational Research, 181(3), 1413–1424.
Liu, J. S., Lu, L. Y., Lu, W. M., & Lin, B. J. (2013). Data envelopment analysis 1978–2010: A
citation-based literature survey. Omega, 41(1), 3–15.
Meibodi. E. (1998). Efficiency considerations in the electricity supply industry: The case of Iran.
Doctoral dissertation, University of Surrey.
Murillo-Zamorano, L. R., & Vega-Cervera, J. A. (2001). The use of parametric and non-parametric
frontier methods to measure the productive efficiency in the industrial sector: A comparative
study. International Journal of Production Economics, 69(3), 265–275.
Miliotis, P. A. (1992). Data envelopment analysis applied to electricity distribution districts. Journal
of the Operational Research Society, 43(5), 549–555.
Ministry of Development, Research of Development Grading Research of Cities and Districts—
SEGE, Ankara. (2013). https://www.bebka.org.tr/admin/datas/sayfas/89/sege-2017_1581687
211.pdf
Ministry of Industry and Technology Research Report, Measuring Urbanization Level in Turkish
Districts by Population Density and Urban Functions, 1, Ankara. (2019). https://www.mevka.
org.tr/Yukleme/Uploads/DsyMU6cBf913201941250PM.pdf
Mirza, F. M., Mushtaq, I., & Ullah, K. (2017). Assessing the efficiency dynamics of post reforms
electric distribution utilities in Pakistan. Utilities Policy, 47, 18–28.
Mullarkey, S., Caulfield, B., McCormack, S., & Basu, B. (2015). A framework for establishing
the technical efficiency of Electricity Distribution Counties (EDCs) using Data Envelopment
Analysis. Energy Conversion and Management, 94, 112–123.
Ogunyinka, E. O., & Ajibefun, I. A. (2004). Determinants of technical inefficiency on farm produc-
tion: Tobit analysis approach to the NDE farmers in Ondo State [Nigeria]. International Journal
of Agriculture and Biology (pakistan), 6(2), 355–358.
Olatubi, W. O., & Dismukes, D. E. (2000). A data envelopment analysis of the levels and
determinants of coal-fired electric power generation performance. Utilities Policy, 9(2), 47–59.
Pacudan, R., & De Guzman, E. (2002). Impact of energy efficiency policy to productive efficiency
of electricity distribution industry in the Philippines. Energy Economics, 24(1), 41–54.
Pérez-Reyes, R., & Tovar, B. (2009). Measuring efficiency and productivity change (PTF) in the
Peruvian electricity distribution companies after reforms. Energy Policy, 37(6), 2249–2261.
Petridis, K., Ünsal, M. G., Dey, P. K., & Örkcü, H. H. (2019). A novel network data envelopment
analysis model for performance measurement of Turkish electric distribution companies. Energy,
174, 985–998.
Pollitt, M. G. (1995). Ownership and Performance in Electric Utilities. Oxford: Oxford University
Press.
Ramos-Real, F.J., Tovar, B., Lootty, M., De Almeida, E.F, & Pinto Jr, H.Q. (2009). The evolution and
main determinants of productivity in Brazilian electricity distribution 1998–2005: An empirical
analysis. Energy Economics, 31(2), 298–305.
Resende, M. (2002). Relative efficiency measurement and prospects for yardstick competition in
Brazilian electricity distribution. Energy Policy, 30(8), 637–647.
Sadjadi, S. J., & Omrani, H. (2008). Data envelopment analysis with uncertain data: An application
for Iranian electricity distribution companies. Energy Policy, 36(11), 4247–4254.
Measuring Efficiency and Productivity Change … 99

Santos, S. P., Amado, C. A., & Rosado, J. R. (2011). Formative evaluation of electricity distribution
utilities using data envelopment analysis. Journal of the Operational Research Society, 62(7),
1298–1319.
Simab, M., & Haghifam, M. (2010). Using Integrated Model to Assess the Efficiency of Electric
Distribution Companies. IEEE Transactions on Power Systems, 25(4), 1806–1814.
Simar, L., & Wilson, P. W. (2007). Estimation and inference in two-stage, semi-parametric models
of production processes. Journal of Econometrics, 136(1), 31–64.
Şirin, S. M. (2017). A panel data analysis on the costs of Turkish electricity distribution companies.
Energy Strategy Reviews, 18, 250–259.
Tavassoli, M., Ketabi, S., & Ghandehari, M. (2020). Developing a network DEA model for sustain-
ability analysis of Iran’s electricity distribution network. International Journal of Electrical Power
& Energy Systems, 122(106187), 1–13.
Tavassoli, M., Faramarzi, G. R., & Saen, R. F. (2015). Ranking electricity distribution units using
slacks-based measure, strong complementary slackness condition, and discriminant analysis.
International Journal of Electrical Power & Energy Systems, 64, 1214–1220.
Tobin, J. (1958). Estimation of relationships for limited dependent variables. Econometrica: Journal
of the Econometric Society, 24–36.
Xavier, S. S., Lima, J. W. M., & Lima, L. M. M. (2015). How Efficient are the Brazilian Electricity
Distribution Companies? Journal of Control, Automotion and Electrical Systems, 26, 283–296.
Yang, H., & Pollitt, M. (2009). Incorporating both undesirable outputs and uncontrollable variables
into DEA: The performance of Chinese coal-fired power plants. European Journal of Operational
Research, 197(3), 1095–1105.
Yunos, J. M., & Hawdon, D. (1997). The efficiency of the National Electricity Board in Malaysia:
An intercountry comparison using DEA. Energy Economics, 19(2), 255–269.
Zhang, T., & Bartels, R. (1998). The effect of sample size on the mean efficiency in DEA with
an application to electricity distribution in Australia, Sweden and New Zealand. Journal of
Productivity Analysis, 9(3), 187–204.
Zhou, P., Ang, B. W., & Poh, K. L. (2008). A survey of data envelopment analysis in energy and
environmental studies. European Journal of Operational Research, 189(1), 1–18.
Price and Volatility Forecasting
in Electricity with Support Vector
Regression and Random Forest

Mahmut Kara, Kazim Baris Atici, and Aydin Ulucan

1 Introduction

In almost all major economies of the world, the electricity markets have experienced
liberalization and deregulation in the last two decades. Today, the electricity prices are
set dynamically by the competitive market conditions based on bidding mechanisms
similar to other commodities (such as stock or agricultural commodity). On the other
hand, the market structure is unique in nature compared to other commodity markets
since a balance between production and consumption is required due to the fact
that these processes occur simultaneously. As a result, the parties involved in these
competitive markets face plenty of challenges in managing uncertainty. To cope with
this high level of uncertainty in electricity prices and for developing effective bidding
strategies, forecasting based on historical records of the prices is becoming crucial
for all players of the electricity markets. As stated by Brusaferri et al. (2019), the
availability of accurate day-ahead energy price forecasts is very important to achieve
successful participation in liberalized electricity markets.
Forecasting in the electricity markets is a challenging task due to the highly
volatile behavior of the prices and their time-variant dependence on the consumption
changing during the hour of the day, day of the week or time of the year (Amjady &
Farshid, 2009; Lago et al., 2018). A recent shift can be observed in the forecasting
related literature to the application of Machine Learning (ML) methodologies more
predominantly for price and volatility forecasting in various electricity markets all

M. Kara
Ministry of Treasury and Finance, 06490 Cankaya, Ankara, Turkey
e-mail: [email protected]
K. B. Atici (B) · A. Ulucan
Department of Business Administration, Hacettepe University, 06800 Cankaya, Ankara, Turkey
e-mail: [email protected]
A. Ulucan
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 101
A. B. Dorsman et al. (eds.), Applied Operations Research and Financial Modelling
in Energy, https://doi.org/10.1007/978-3-030-84981-8_6
102 M. Kara et al.

over the world (see Sect. 2 for a brief review). This is due to the fact that the nature
of learning algorithms requiring less prior assumptions on the data.
One of the recently liberalized electricity markets in emerging economies is the
Turkish electricity market. The spot electricity markets (day-ahead and intraday
markets) have been introduced in 2011. The hourly price data for day-ahead and
intraday markets are available at the transparency platform of the EPIAS. Along
with other markets, the Turkish electricity market has been the subject of various
forecasting research utilizing machine-learning algorithms. It is observed that neural
network applications are dominant in the research on the Turkish electricity market.
One of the recent examples includes Ugurlu et al. (2018a), in which multi-layer
gated recurrent neural networks are utilized and tested for 2016 prices. In an earlier
research by Kolmek and Navruz (2015), Artificial Neural Network (ANN) model is
applied and its test results, that is 3 weeks in 2010, are compared to Autoregressive
Integrated Moving Average (ARIMA) model. In other research on Turkish markets,
Ozguner et al. (2017) also apply ANN and report similar results with Kolmek and
Navruz (2015). In addition to those, it is possible to come up with recent research
employing econometric, hybrid or multiple stage modeling to the Turkish market
data. Some examples include Hayfavi and Talasli (2014) with a multifactor hybrid
model, Ozozen et al. (2016) presenting a hybrid model combining ANN and ARIMA
models, Ugurlu et al. (2018b) proposing a two-stage model that is based on Anal-
ysis of Variance (ANOVA) followed by Seasonal ARIMA and Avci et al. (2018)’s
ensemble forecasting model.
In the current research, we aim to carry out price and volatility forecasting for
the Turkish day-ahead electricity market with SVR and RF methodologies and
observe the effectiveness of the methods. The motivation behind the selection of these
methods is twofold. First, when compared to econometric methods, these methods
require less assumptions on data and proven effective in different electricity markets.
Second and more importantly, we observe that the efforts of price forecasting in
the Turkish electricity markets mainly focus on neural network applications. As a
response to that, we aim to experiment with two methods that are unconventional
for Turkish electricity market but are widely applied in the literature and therefore,
provide a new methodological perspective. In addition, For our objective, first, using
prices (in Euro) between January-2013 and September-2019, we design a rolling
scheme to forecast hourly prices and apply both methods. The performance metrics
are compared with each other and with those of naive estimations. We also test our
price forecasting model with a reduced number of features and evaluate the sensitivity
of the results, in which short-term features are excluded from the model. Secondly,
the realized volatility data is generated and volatility forecasting is carried out with
SVR and RF. The results are compared with naive estimations.
Considering above objectives, the scope and contribution of the research can
be thought in four aspects. First of all, we implement a framework that is based
on a rolling data scheme covering approximately 6.5 years of forecasting horizon
in a recently liberalized electricity market. Secondly, we assemble two widely-used
learning methodologies to compare their performances within this framework. Third,
this framework is applied separately for price and volatility forecasting that results
Price and Volatility Forecasting in Electricity … 103

in a rich set of solutions with two essential dimensions in electricity markets and
decent accuracy. Finally, we also experiment with a reduced number of features for
further practical considerations.
The chapter is organized as follows. In Sect. 2, we present the previous literature
on both price and volatility forecasting in electricity markets. Section 3 explains the
basics of SVR and RF Methods. Section 4 presents the data, the analysis design
of price forecasting, and volatility forecasting. Section 5 is devoted to the findings
of the analysis in two sub-sections. The first Sect. 5.1 presents the results on price
forecasting and their comparisons. The second Sect. 5.2 is dealing with the findings
of volatility forecasting. Finally, Sect. 6 concludes the chapter.

2 Previous Research

2.1 Price Forecasting with Machine Learning in Electricity

Markets

Electricity price forecasting is one of the primary streams of energy research in which
different methodologies from statistical/econometric methods to machine learning
have been applied (see Debnath & Mourshed, 2018 for a recent review). In the
recent literature, it is possible to notice that the applications of Machine Learning
(ML) algorithms have become predominant. Table 1 presents several examples of
price forecasting research using ML methodologies.
Method-wise, Support Vector Machine type modeling, Neural Network (NN) type
modeling, and hybrid modeling approaches stand out as three main methodologies.
Australian, European, and US electricity markets attract much of the attention by
the ML researchers. The majority of the research is recent, indicating the current
attractiveness of the area.
In price forecasting research with ML, a well-defined flow can be observed in
the majority of the research. A method is (or methods are) proposed, the proposed
method(s) is (are) tested in one (or more) real-world market(s), the forecasting perfor-
mance is measured and the performance is compared either with results of previous
research or with the results of other method(s) (econometric or machine learning).
In general, hourly data are utilized, and day-ahead forecasting is of interest. It is
observed that the research undertake 1–3 years of out-of-sample forecast periods.
The research mainly focuses on obtaining better forecast accuracy by experimenting
with changing methodological considerations or combining different methodologies
in a way to introduce a new hybrid model.
In almost all research, the forecast accuracy is compared with the accuracy of other
methods (econometric or ML). In Table 2, we present the metrics and benchmark
methods of the sample research listed in Table 1. The metrics vary between the
researches; however, Mean Absolute Error (MAE), Mean Absolute Percentage Error
(MAPE), and Mean Square Error (RMSE) are common. The comparisons are based
104 M. Kara et al.

Table 1 Sample electricity price forecasting research with machine learning

Title Market Methodology Data
1 Fan et al. (2007) Australia Hybrid (SVM and Hourly
SOM)
2 Gao et al. (2007) Italy and Australia and SVM Hourly
US and Spain
3 Amjady and Farshid Spain and Australia Hybrid (ANN and FS) Hourly
(2009)
4 Saini et al. (2009) Australia SVM Half-Hourly
5 Aggarwal et al. (2009) Canada (Ontario) SVM Hourly
6 Che and Wang (2010) US (California) Hybrid (SVM and Hourly
ARIMA)
7 Shayeghi and Ghasemi Iran and Canada and Hybrid (Various incl. Hourly
(2013) Spain SVM)
8 Li et al. (2013) US (PJM) PCPF Hourly
9 Mei et al. (2014) US (New York) RF 5 min
10 Keles et al. (2016) Germany ANN Hourly
11 Sandhu et al. (2016) Canada (Ontario) ANN Hourly
12 González et al. (2016) Spain RF Hourly
13 Lago et al. (2018) Belgium and France ANN Hourly
14 Ugurlu et al. (2018a) Turkey ANN Hourly
15 Wang et al. (2019) US (PJM) Hybrid (4-steps) Hourly
16 Brusaferri et al. (2019) Italia and Belgium ANN Daily
17 Cheng et al. (2019) Germany Hybrid (Various incl. Hourly
SVM)
18 Zhang et al. (2019) Australia Hybrid (Various incl. Half-Hourly
SVM)
19 Luo and Weng (2019) US (Texas) Hybrid (2-stage) Hourly
20 Büyükşahin and Ertekin Turkey Hybrid (ANN and Hourly
(2019) ARIMA)
21 Ghayekhloo et al. (2019) US (NYISO) Hybrid (incl. RNN) Hourly
22 Yang et al. (2019) Australia Hybrid (uses MOO) Half-Hourly
23 Marcjasz et al. (2019) Global ANN Hourly
Abbreviations: ANN Artificial Neural Network, FS Feature Selection, MOO Multi-Objective
Optimization, PCPF Panel Cointegration and Particle Filter, PJM Pennsylvania, Jersey, Maryland,
RNN Recurrent Neural Network, SOM Self-Organized Maps, SVM Support Vector Machine, RF
Random Forest
Price and Volatility Forecasting in Electricity … 105

Table 2 Metrics and comparisons of the sample electricity price forecasting research
Title Metrics Benchmark methods
1 Fan et al. (2007) MAE and MAPE –
2 Gao et al. (2007) MAPE and MSE ARIMA
3 Amjady and Farshid (2009) WMAPE ARIMA and ANN
4 Saini et al. (2009) MAPE Linear regression and
other papers
5 Aggarwal et al. (2009) MAPE and RMSE Various ML and
econometric methods
6 Che and Wang (2010) RMSE and MAPE ANN and NARIMA
7 Shayeghi and Ghasemi WMAPE and WMAE ARIMA
(2013)
8 Li et al. (2013) WMAPE ANN
9 Mei et al. (2014) MAPE ARMA and ANN
10 Keles et al. (2016) RMSE and MAD ARIMA
11 Sandhu et al. (2016) MAPE and MAE and RMSE –
12 González et al. (2016) MSE RF
13 Lago et al. (2018) MAPE Own improvements
14 Ugurlu et al. (2018a) MAE Various ML methods
15 Wang et al. (2019) RMSE and MAPE Various ML methods
(including SVM)
16 Brusaferri et al. (2019) RMSE Own improvements
17 Cheng et al. (2019) MAE and RMSE and MAPE Splits of the hybrid
model
18 Zhang et al. (2019) MAPE ARIMA and SVM
19 Luo and Weng (2019) MSE Various ML and
econometric methods
20 Büyükşahin and Ertekin MAE and MSE and MASE Various ML and
(2019) econometric methods
21 Ghayekhloo et al. (2019) MSE and MAPE Various ML and
econometric methods
22 Yang et al. (2019) MAE Various ML and
econometric methods
23 Marcjasz et al. (2019) WMAE ANN
Abbreviations: ANN Artificial Neural Network, ARIMA Auto-Regressive Integrated Moving
Average, MAE Mean Absolute Error, MAPE Mean Absolute Percentage Error, MSE Mean Square
Error, RMSE Root Mean Square Error, SVM Support Vector Machine, WMAE Weighted Mean
Absolute Error, WMAPE Weighted Mean Absolute Percentage Error, RF Random Forest
106 M. Kara et al.

on tests with other econometric or machine learning methodologies. All research

claims that their proposed methodology produces better forecast accuracy than the
benchmark methods.

2.2 Volatility Forecasting in Electricity Markets

In a recent paper, Ghoddusi et al. (2019) review more than 130 articles about machine
learning in energy economics and finance. The review concludes that there exists no
significant body of papers devoted to modeling volatility in energy markets except
for a few examples. Therefore, the research on volatility forecasting of electricity
markets is limited and requires attention. Table 3 presents a sample of research on
volatility forecasting in electricity markets. In general, the domination of econometric
modeling is observed. Below, we briefly summarize the modeling approaches and
findings of the presented research.
In relatively earlier works on the volatility of electricity markets, Haugom and
Ullrich (2012) and Frömmel et al. (2014) both study the realized volatility for Euro-
pean and American electricity markets. Both pieces of research indicate that real-
ized GARCH-type volatility models are, in general, effective tools for forecasting
daily volatility of electricity prices. Tashpulatov (2013) estimates the electricity price
volatilities for England using AR and ARCH modeling. The impact of introduced
institutional changes and regulatory reforms on price and volatility dynamics are
investigated. Results reveal that innovations from the previous week have asym-
metric effects on volatility such that positive innovations from the previous week
have a larger effect on volatility compared to negative ones. Aydogdu (2016) conducts
comprehensive research on the European day-ahead power market, including Turkey,

Table 3 Relevant research on volatility forecasting in electricity markets

Reference Method Market
Haugom and Ulrich (2012) HAR model of realized volatility Pennsylvania and New Jersey
and Maryland
Tashpulatov (2013) AR and ARCH UK
Frömmel et al. (2014) GARCH EPEX power markets
Aydogdu (2016) E-GARCH and T-GARCH European day-ahead power
markets
Qu et al. (2016) Logistic Autoregressive Model Australian New South Wales
Electricity Market
Qu et al. (2018) HAR-GARCH Hybrid Models Australian New South Wales
Electricity Market
Abbreviations: ARCH Autoregressive Conditional Heteroscedasticity, GARCH Generalized
Autoregressive Conditional Heteroscedasticity, EPEX European Power Exchange, HAR
Heterogeneous Autoregression
Price and Volatility Forecasting in Electricity … 107

using econometric modeling. The results reveal that there is not a uniform inverse
leverage effect in electricity prices; that is, price increases are more destabilizing in
some European markets (e.g., Poland, Slovenia, Ireland, Netherlands) than compa-
rable price decreases, but vice versa also holds in some other countries (e.g., Portugal,
France).
In research on the Australian market, Qu et al. (2016) introduce a hybrid model
combining logistic smooth transition structure to the HAR model and show that
this model improves in-sample fit. In recent research, Qu et al. (2018) develop a
hybrid model combining HAR and GARCH models and show that the hybrid model
outperforms benchmark HAR models also in the Australian Electricity Market. They
also incorporate the price jumps in their model structure to improve out-of-sample
forecasting performance.

3 Methods

3.1 Support Vector Regression

Suppose we have k number of features in the training set X with I data points, where
1 ≤ I ≤ k. The training set can be set up as below, in which the corresponding
target set is defined with y ∈ (1, −1). In the given setup, xi = [x1 , x2 , x3 , . . . , xn ]
represents the ith vector in set X .

X = {(xi , yi )}i=1
I
where xi ∈ R n and, y ∈ {1, −1} (1)

In ε-SV regression, the goal is to find a function f(x) that has at most ε deviation
from the actual targets yi for all the training data, and at the same time is as flat
as possible (Vapnik, 1995). The model aims to determine the decision boundary or
hyperplane that differentiates between two classes of input data as good as possible.
The hyperplanes can be formulated with the following kernel functions:

f (x) = w · x + b (2)

In this function, b represents the bias term, and w · x is the inner products of
a new input vector, where w is the weighting vector, and x is the input vector. The
inner product is calculated as follows:


n
w · x = wjxj (3)
j=1

The SVR identifies an optimal separating hyperplane with the maximum margin
by solving the following optimization problem:
108 M. Kara et al.

1
min w T · w subject to yi · (w · xi  + b) − 1 ≥ 0 ∀i (4)
w,b 2

3.2 Random Forest Modeling

RF regression modeling is a supervised machine learning methodology and is devel-

oped using decision tree analysis by Breiman (2001). RF consists of many uncorre-
lated decision trees built upon different samples of the data. The combined prediction
trees are expected to increase the accuracy and stability of the model performances
as compared with those of a single regression model. In general, for RF, the larger
the number of decision trees, the more the convergence of the generalization error to
a specific value; hence the over-fitting is avoided. However, the computation burden
increases as the number of decision trees.
A decision tree is a non-parametric model and does not assume any prior
parameters. The tree grows during the learning process. RF flow is as follows:
• It starts with selecting the random samples from a given dataset with replacement.
• RF constructs a decision tree for every sample. Then it calculates a prediction
result for every decision tree. This step is repeated until it reaches the number of
trees defined in the model.
• Then RF aggregates the prediction results and produces an average prediction
result.
As Breiman (2001) stated, a random vector Θk is generated that is independent
of past random vectors Θ1 , Θ2 , . . . , Θk−1 . However, this generation is done within
the same distribution. As a result, a tree is built using the training set and a classifier
h(x, Θk ) where x is an input vector. For this selection process, Θ consists of a number
of independent random integers between 1 and k. After generating a large number
of trees, the most popular class is voted. This whole process is defined as a random
forest (Breiman, 2001).
Formally, RF is a classifier that consists of a set of tree-structured classi-
fiers {h(x, Θk ), k = 1, . . . } where {Θk } are independent and identically distributed
random vectors. For every tree casts a unit vote for the most popular class at input
vector x.
Price and Volatility Forecasting in Electricity … 109

4 Data and Methodology

4.1 Data Set

The spot electricity markets (day-ahead and intraday markets) of Turkey (introduced
in 2011) are managed and operated by EPIAS Company. The physical transactions are
managed, and the balancing market is operated by TEIAS Company. We obtain the
data from EXIST (Energy Exchange Istanbul) Transparency Platform.1 The hourly
day-ahead electricity price data set dates back to December 2011. For the price
forecasting, we utilize the prices (in Euro) between January 01, 2013, and September
30, 2019. The first two years (i.e., 2011 and 2012) data are not used to leave a period
for the market to be fully established. For volatility forecasting, daily volatility data
are derived using realized volatility calculations (see Sect. 4.3 for details). We deal
with 2,462 prices for each hour that add up to 59,088 data points in total. The
descriptive statistics of the electricity price data set are given in Table 4.
For both price and volatility forecasts, 16 features are used to predict hourly
electricity prices. The list of features is presented in Table 5. The features consist of
commonly used variables in the literature2 as lagged prices, simple moving averages,
electricity load forecasts and day and month of the data points.

4.2 Price Forecasting Model

We begin with the price forecasting using the data set described above. The hourly
prices in the first-week data are used to generate lagged values; therefore, the anal-
ysis period begins on January 08, 2013. For each day, we have 24 h of prices. A
total of 58,967 data points from 59,088 are utilized covering January 08, 2013, to
September 30, 2019, which corresponds to 351 weeks. The forecasts cover hourly
prices in 340 weeks (351 weeks–11 weeks) in a rolling structure. Python is used for
all computations.
In obtaining price forecasts, we follow rolling scheme below (illustrated in Fig. 1):
• Training: 11 weeks from the beginning (January 08, 2013), corresponding to
1,848 data points (24 × 7 × 11), are used as training data (i.e., the data points
from 1 to 1,848 are used for training).
• Forecast: The hourly electricity prices for 1 week, corresponding to 168 data
points, are forecasted (i.e.,the data points from 1,849 to 2,016 are forecasted).
• Rolling and Training: The training data set is moved 168 data points ahead (to
the data points from 169 to 2016).

1 https://seffaflik.epias.com.tr/transparency/index.xhtml.
2 The reader may refer to Shayeghi and Ghasemi (2013); Ugurlu et al. (2018a) and Lago et al.
(2018) for examples.
110 M. Kara et al.

Table 4 Descriptive statistics of the hourly day-ahead electricity prices (Jan. 2013–Sep. 2019)
# of obs Min Max Mean Median Variance Stdev Skewness Kurtosis
Hour 0 2,462 0.0 81.2 46.6 46.4 168.1 13.0 −0.3 0.9
Hour 1 2,462 0.0 81.0 42.7 42.5 154.6 12.4 −0.3 1.3
Hour 2 2,462 0.0 80.8 37.9 37.9 168.8 13.0 −0.3 0.9
Hour 3 2,462 0.0 80.8 33.7 34.5 188.6 13.7 −0.3 0.4
Hour 4 2,462 0.0 80.8 32.3 33.5 194.8 14.0 −0.4 0.2
Hour 5 2,462 0.0 80.8 32.4 33.6 188.7 13.7 −0.3 0.4
Hour 6 2,462 0.0 80.9 33.5 35.6 231.1 15.2 −0.4 0.3
Hour 7 2,462 0.0 81.0 40.2 41.6 211.2 14.5 −0.6 0.9
Hour 8 2,462 0.0 84.0 48.6 49.0 236.8 15.4 −0.6 0.9
Hour 9 2,462 0.0 190.9 53.3 52.0 236.4 15.4 −0.3 3.8
Hour 10 2,462 0.0 273.1 54.6 52.6 245.9 15.7 1.2 20.0
Hour 11 2,462 0.0 316.5 55.9 53.5 248.1 15.7 2.1 34.1
Hour 12 2,462 0.0 273.1 50.7 49.8 247.8 15.7 0.9 16.0
Hour 13 2,462 0.0 218.5 52.2 50.9 250.8 15.8 0.4 6.9
Hour 14 2,462 0.0 518.8 53.9 52.0 323.1 18.0 7.5 185.8
Hour 15 2,462 0.0 409.6 52.4 51.0 275.2 16.6 4.1 88.1
Hour 16 2,462 0.0 232.7 52.1 50.6 243.1 15.6 1.1 13.5
Hour 17 2,462 0.0 320.2 50.6 49.7 269.2 16.4 2.8 39.5
Hour 18 2,462 0.2 212.2 49.5 48.9 187.3 13.7 0.7 8.8
Hour 19 2,462 0.6 136.9 49.7 48.8 138.0 11.7 0.4 2.0
Hour 20 2,462 16.4 126.0 50.1 49.1 120.2 11.0 0.7 1.6
Hour 21 2,462 10.6 81.3 48.7 47.6 115.6 10.8 0.5 0.1
Hour 22 2,462 1.4 82.0 48.4 47.1 175.6 13.3 0.3 −0.3
Hour 23 2,462 0.0 81.2 44.4 43.6 188.6 13.7 0.0 0.2
Overall 59,088 0.0 518.8 46.4 46.3 262.8 16.2 0.9 23.2

• Forecast: A new model is set up, and the hourly electricity prices for 1 week are
forecasted (i.e.,the data points 2,017 to 2,184 are forecasted).
• Rolling and Training and Forecasting: The rolling of the training set and hourly
price forecasting for each week continues until we forecast the hourly prices for
the final week, which is week 351 (i.e.,data points from 58,800 to 58,967).
Furthermore, we select a sample of data corresponding to 20% of the data set for
validation purposes. This is to identify the hyperparameters for the SVR model that
minimize the mean absolute errors.
Price and Volatility Forecasting in Electricity … 111

Table 5 Features
Features Abbreviations Details
Feature 1 LAG1 One hour lagged price
Feature 2 LAG5 Five hours lagged price
Feature 3 LAG24 24 h lagged price
Feature 4 LAG48 48 h lagged price
Feature 5 LAG120 120 h lagged price
Feature 6 LAG168 168 h lagged price
Feature 7 SMA2 Simple moving averages of electricity prices (last 2 h)
Feature 8 SMA5 Simple moving averages of electricity prices (last 5 h)
Feature 9 SMA24 Simple moving averages of electricity prices (last 24 h)
Feature 10 SMA48 Simple moving averages of electricity prices (last 48 h)
Feature 11 SMA120 Simple moving averages of electricity prices (last 120 h)
Feature 12 SMA168 Simple moving averages of electricity prices (last 168 h)
Feature 13 LOAD Electricity load forecast that is given by EPIAS
Feature 14 HOUR The hour of the day (0–23)
Feature 15 DAY The day of the month (1–31)
Feature 16 MONTH The month of the year (1–12)

Fig. 1 Rolling scheme

4.3 Volatility Forecasting Model

For the volatility forecasting of the electricity prices, we use realized volatility as in
Qu et al. (2016), Haugom and Ullrich (2012) and Frömmel et al. (2014). The realized
volatility is calculated as:
112 M. Kara et al.

Table 6 Descriptive statistics of volatility data

# of obs Mean Median Min Max Variance St. dev Skewness Kurtosis
All 2,416 1.14 0.09 0.00 40.84 11.90 3.45 4.83 27.98
Test 1,420 1.47 0.13 0.00 40.84 14.99 3.87 4.21 21.41
period


M
RVt = rt,2 j (5)
j=1

where the price is observed at discrete times j = 1, 2, . . . , M within each day

t = 1, 2, . . . and the jth intraday return of day t is represented as:

rt j = pt j − pt j−1 (6)

pt j is the logarithmic price that is pt j = log(Pt ), where Pt represents the realized

electricity price.
Table 6 presents the descriptive statistics for the volatility data derived using
hourly electricity prices. We consider 10 features (LAG1, LAG5, LAG24, LAG48,
SMA2, SMA5, SMA24, SMA48, DAY, and MONTH). The full data set consists of
2,416 observations out of 2,462 due to the lagging. We also use a rolling scheme for
volatility forecasting with a rolling train data set of 996 observations. The forecast
horizon (test data) consists of 1,420 data points with a mean realized volatility of
1.47. As seen in Table 6, the test period of the data set is a good representation
of the whole data set. Both data sets have similar mean, variance, skewness, and
kurtosis values. The rolling data scheme enables us to increase the size of the data
set. Through the rolling scheme, we test approximately 60% of the whole data set.

4.4 Metrics

We utilize three commonly used metrics to measure the performance of our forecasts
as MAE, MAPE, and RMSE. The formulations of those metrics are given below, where
n represents the number of data that are subject to forecasting, yt is the real value of
the data at time t and ŷt is the predicted value of yt at time t. We use the adjusted
version of the original MAPE, proposed by Armstrong (1985) that is used when there
is the probability that the actual values may be zeros.

1  
n
M AE = yt − ŷt  (7)
n t=1
Price and Volatility Forecasting in Electricity … 113

n  
1  yt − ŷt 
M AP E = (8)
n t=1  (yt + ŷt) /2 

n  2
t=1 yt − ŷt
RMSE = (9)
n

5 Findings

In this section, we present the findings of our analyses in two main parts as price
forecasting and volatility forecasting.

5.1 Price Forecasting

In price forecasting, we first present the results for SVR, followed by RF forecasts
and their comparisons. Then, we also test our price forecasting model with a reduced
number of features and evaluate the sensitivity of the results, in which short-term
features are excluded from the model.

5.1.1 Price Forecasting with SVR

Table 7 presents the average prices and average forecasts for every hour in the test
period together with the values of three performance indicators (see Sect. 4.4 for
formulations) as MAE, MAPE, and RMSE for each hour of the day. The average
price is closest to the average forecasted price at 06:00 and the most distant at noon.
It can be observed that the hourly electricity price forecasts exhibit the lowest
level of errors at hours 11:00, 15:00 and 21:00 of the day in terms of all metrics.
Relatively, the error levels are at their highest during the morning period at hours
06:00, 07:00 and 08:00. Except for these particular hours, forecast performance for
every period are consistent with the average forecast performance. We compare the
SVR performance metrics with those of the naive estimations calculated by taking
the price at t + 1 as the forecasts and the price at t as real. The metrics for both
estimations are presented in Table 8. It is observed that the SVR metrics are superior
to naive estimation metrics.
For the Turkish day-ahead electricity market, one of the most comprehensive and
recent price forecasting research is by Ugurlu et al. (2018a). They use multi-layer
gated recurrent neural networks and test for the year 2016, excluding 9 days due to
very high prices because of natural gas shortage during these days (in our model, we
did not make any exclusion). When compared with their recurrent neural network
114 M. Kara et al.

Table 7 Hourly price forecasting performance with SVR

Hours Average price Average forecasted price MAE MAPE RMSE
00:00 46.57 44.72 4.71 0.122 6.83
01:00 42.67 43.88 3.58 0.112 5.28
02:00 37.85 40.07 4.14 0.139 5.95
03:00 33.72 35.92 4.30 0.164 6.42
04:00 32.31 33.31 3.61 0.187 5.44
05:00 32.41 32.80 4.30 0.221 6.34
06:00 33.47 33.45 5.18 0.287 7.32
07:00 40.20 37.38 5.57 0.204 7.95
08:00 48.56 45.58 5.72 0.167 8.08
09:00 53.27 52.13 3.69 0.095 5.50
10:00 54.63 54.67 2.84 0.073 5.39
11:00 55.93 55.34 2.46 0.060 4.15
12:00 50.73 53.86 3.85 0.089 5.74
13:00 52.15 51.32 2.82 0.073 4.68
14:00 53.94 52.60 2.81 0.060 6.02
15:00 52.40 53.00 2.61 0.064 4.08
16:00 52.10 51.94 2.87 0.071 4.93
17:00 50.61 50.89 3.47 0.088 5.23
18:00 49.49 49.44 3.40 0.086 5.21
19:00 49.66 48.88 3.19 0.074 4.83
20:00 50.07 49.15 3.12 0.067 4.55
21:00 48.70 48.82 2.65 0.058 3.78
22:00 48.45 47.62 3.14 0.074 4.56
23:00 44.36 45.60 3.44 0.098 5.31
Overall 46.43 46.35 3.65 0.114 5.68

Table 8 SVR price

MAE MAPE RMSE
forecasting performance for
the test period SVR 3.65 0.114 5.68
Naive estimation 5.47 0.158 8.22

model, our SVR model hits very close MAE values for the year 2016. The average
MAE values in 2016 are 5.03 and 5.36 in our SVR model and recurrent neural network
model of Ugurlu et al. (2018a), respectively.
Figure 2 plots the price forecasts with the actual prices for the selected weeks of
each season. Note that the model captures the spikes of the electricity prices well for
all seasons. The model appears to capture the actual prices relatively better for the
colder periods of the year (autumn and winter).
Price and Volatility Forecasting in Electricity … 115

Autumn Spring
80 50

30
60
10

40 -10

7-May-19

8-May-19

9-May-19

10-May-19

11-May-19

12-May-19

13-May-19
10-Nov-14
4-Nov-14

5-Nov-14

6-Nov-14

7-Nov-14

8-Nov-14

9-Nov-14
Forecast Price Forecast Price

Winter Summer
80 60
70
60 50
50 40
40
30 30
20
20
10
0 10
29-Dec-15

30-Dec-15

31-Dec-15

1-Jan-16

2-Jan-16

3-Jan-16

4-Jan-16

20-Jun-17

21-Jun-17

22-Jun-17

23-Jun-17

24-Jun-17

25-Jun-17

26-Jun-17
Forecast Price Forecast Price

Fig. 2 The fit of prices and forecasts for each season

5.1.2 Comparisons with RF Forecasts

To provide more insight into Turkish electricity price forecasting and to compare
SVR price forecasting results, we experiment with another learning algorithm as a
common practice in forecasting research. The RF methodology is employed with
the same features as the SVR model. The results are compared in terms of hourly,
monthly, and yearly performance of both modeling approaches. The RF is known as
one of the most accurate learning algorithms available. Therefore, we select RF to
compare our results obtained through the SVR model.
First of all, we present the hourly forecasting performance of RF modeling in
Table 9. Similar to the SVR findings, hours 11:00, 15:00, and 21:00 have relatively
lower error values. Again, the error levels are at their highest during the morning
period at hours 05:00, 06:00, 07:00, and 08:00. When Tables 7 and 9 are compared,
it can be observed that concerning all three metrics (MAE, MAPE, and RMSE), the
accuracy of SVR modeling is superior to RF modeling for every period of the day. The
differences between the metrics are not very substantial between two methodologies
if we look at hourly averages.
The difference between the performance of SVR and RF modeling is more notice-
able when the results are analyzed on a monthly and yearly basis. Tables 10 and 11
present the comparisons on a monthly and yearly averages basis. Especially, MAPE
values are substantially better for SVR in 2019 forecasts.
Overall, in electricity price forecasting, the SVR model consistently produces
better forecasts compared to the RF model on an hourly, monthly, and yearly basis.
116 M. Kara et al.

Table 9 Hourly price forecasting performance with RF

Hours Average price Average forecasted price MAE MAPE RMSE
00:00 46.57 44.74 4.75 0.128 6.76
01:00 42.67 42.56 3.89 0.120 5.63
02:00 37.85 38.38 4.26 0.157 6.20
03:00 33.72 34.45 4.80 0.227 6.98
04:00 32.31 32.16 4.44 0.230 6.40
05:00 32.41 32.09 5.04 0.250 7.07
06:00 33.47 33.21 5.39 0.299 7.49
07:00 40.20 38.40 5.78 0.225 8.00
08:00 48.56 46.19 5.69 0.168 7.92
09:00 53.27 51.47 4.21 0.118 6.20
10:00 54.63 53.75 3.25 0.086 6.08
11:00 55.93 54.44 3.01 0.073 5.47
12:00 50.73 52.54 3.89 0.104 5.61
13:00 52.15 50.67 3.47 0.095 5.47
14:00 53.94 51.91 3.48 0.087 6.84
15:00 52.40 52.01 2.96 0.081 4.97
16:00 52.10 51.12 3.29 0.086 6.21
17:00 50.61 50.27 3.71 0.092 6.51
18:00 49.49 49.00 3.35 0.085 4.85
19:00 49.66 48.64 3.25 0.076 4.77
20:00 50.07 48.84 3.26 0.071 4.65
21:00 48.70 48.28 2.90 0.064 4.13
22:00 48.45 47.26 3.42 0.080 4.78
23:00 44.36 44.80 3.61 0.106 5.46
Overall 46.43 45.72 3.96 0.129 6.11

5.1.3 Feature Reduction in Price Forecasting with SVR

In price forecasting, our SVR model design utilizes the features that are listed in Table
5. Of those features, LAG1, LAG5, SMA2, and SMA5 refer to the lagged prices (1 h
and 5 h) and the simple moving averages (of the prices for the last 2 and 5 h). These
features represent the prices that belong to relatively closer periods to the forecasted
price. To observe the effects of obtaining the price forecasts at least 24 h in advance
(i.e., trade-off between the forecast horizon and the accuracy of forecasts), we also
experiment in the absence of LAG1, LAG5, SMA2, and SMA5 so that the features
related to price are for at least 24 h earlier.
Table 12 presents the hourly price forecasting performance of the SVR model
with 12 features. When compared with Table 7, the overall MAE and RMSE values
Price and Volatility Forecasting in Electricity … 117

Table 10 Performance comparison of SVR and RF by months

MAE MAPE RMSE
Months SVR RF SVR RF SVR RF
1 4.36 4.50 0.141 0.147 6.69 6.89
2 3.66 4.10 0.122 0.150 5.41 5.95
3 4.02 4.28 0.141 0.157 6.03 6.22
4 3.80 4.30 0.124 0.161 5.40 5.83
5 4.11 4.59 0.145 0.182 5.87 6.44
6 4.12 4.41 0.151 0.166 5.95 6.22
7 3.28 3.49 0.103 0.111 5.22 5.38
8 3.13 3.42 0.095 0.105 5.20 5.45
9 2.90 3.41 0.084 0.096 5.35 5.89
10 3.52 3.87 0.091 0.100 5.17 5.34
11 3.09 3.19 0.079 0.080 4.56 4.57
12 3.83 4.03 0.088 0.094 6.96 8.45
Overall 3.65 3.96 0.114 0.129 5.68 6.11

Table 11 Performance comparison of SVR and RF by years

MAE MAPE RMSE
Years SVR RF SVR RF SVR RF
2013 3.64 3.95 0.076 0.085 6.05 7.25
2014 2.81 3.11 0.055 0.060 4.09 4.45
2015 4.09 4.60 0.124 0.140 5.82 6.46
2016 5.03 5.35 0.202 0.226 7.29 7.57
2017 2.99 3.13 0.085 0.088 4.79 4.97
2018 3.04 3.21 0.092 0.098 4.78 4.92
2019 4.00 4.50 0.167 0.220 6.55 6.81
Overall 3.65 3.96 0.114 0.129 5.68 6.11

are higher as expected. In terms of MAPE, we have an interesting result in terms of

forecasts at 03:00. Due to the negative forecasts, especially for the points at which
the prices are zero, the MAPE value is quite high for this period. Overall, MAPE
value is lower than the MAPE of the original model (given as 0.114 in Table 7) due
to the effect of those negative forecasts.
To overcome the effects of negative forecasts on the performance metrics, we
evaluate the results once again after replacing the negative price forecasts with zeros
since negative prices are not technically possible. With this adjustment, we do not
observe substantial differences in the values of the metrics except for the MAPE value
of 03:00 and overall MAPE, which are obtained as 0.272 and 0.156, respectively.
118 M. Kara et al.

Table 12 Hourly price forecasting performance with SVR after feature reduction
Hours Average price Average forecasted price MAE MAPE RMSE
00:00 45.98 46.11 4.84 0.127 7.08
01:00 42.33 42.67 4.92 0.142 7.25
02:00 37.76 37.94 5.72 0.191 8.26
03:00 33.70 33.99 6.43 −1.031 9.19
04:00 32.44 32.63 6.34 0.271 9.06
05:00 32.43 32.56 6.26 0.263 8.88
06:00 33.73 33.56 6.67 0.277 9.27
07:00 40.16 40.78 6.34 0.209 9.13
08:00 48.38 49.38 6.14 0.172 9.06
09:00 52.98 53.46 5.22 0.132 8.10
10:00 54.15 54.61 4.76 0.112 7.75
11:00 55.38 55.53 4.41 0.099 7.06
12:00 50.29 50.85 4.93 0.123 7.05
13:00 51.73 52.20 4.86 0.119 7.23
14:00 53.62 53.56 4.59 0.104 8.13
15:00 52.03 52.28 4.72 0.112 7.04
16:00 51.64 52.05 4.72 0.110 7.53
17:00 49.95 50.35 4.87 0.114 7.52
18:00 48.81 49.26 4.61 0.112 6.52
19:00 49.11 49.40 4.51 0.100 6.27
20:00 49.58 49.74 4.45 0.094 6.08
21:00 48.32 48.50 4.46 0.098 6.17
22:00 47.85 48.06 4.58 0.106 6.40
23:00 43.77 44.30 5.17 0.144 7.56
Overall 46.09 46.41 5.19 0.096 7.72

Therefore, the MAPE value presented in Table 12 as 0.096 is adjusted to 0.156, which
is higher than the MAPE of the original model with 16 features in Table 7 (0.114).

5.2 Volatility Forecasting

Table 13 presents the performance metrics of volatility forecasts with SVR, RF, and
naive estimations. Despite the differences between the average forecasted volatility
values, SVR forecasts have a lower predicting error in terms of all metrics compared
to naive estimations. Both methods predict the realized volatility of the day-ahead
Turkish electricity market with a reasonable forecast error despite its spikes, high
Price and Volatility Forecasting in Electricity … 119

Table 13 Volatility forecast performance of the models

# Average volatility Average forecasted MAE MAPE RMSE
volatility
SVR 1,420 1.47 0.98 0.84 0.64 2.30
RF 1,420 1.47 1.45 0.29 0.22 1.14
Naive estimation 1,420 1.47 1.45 1.13 0.87 2.96

volatility, heavy tails, and highly skewed data structure. Interestingly, in terms of all
metrics, RF is superior to SVR in volatility forecasting.
Figure 3 plots the forecasted volatility values of both models (SVR and RF)
together with the realized volatility for each year of the test period. The figure consists
of five panels corresponding to the five years covered in the test period. In general,
the Turkish electricity market has a volatile structure. As reported by Gayretli et al.
(2019), the spikes are due to natural gas shortage problems. They reveal that 56.9% of
the spikes are due to problems with natural gas-based power plants. Considering the
volatile structure of the market, we observe that both SVR and RF methods present
a good fit, especially for the years 2017 and 2018. RF is superior in capturing the
spikes for all years. Especially, the superior performance of RF is very visible in the
years 2015, 2016, and 2019.

6 Conclusion and Policy Implications

The current research aims to employ price and volatility forecasting for the Turkish
day-ahead electricity market using SVR and RF methodologies. Our initial data
set covers hourly Euro prices January-2013 and September-2019 period. A rolling
data scheme is designed to produce hourly prices for 340 weeks considering 16
features. The volatility forecasting covers realized volatility values comprising more
than 2000 days and 10 features. Three performance metrics are used to compare the
forecasts for our models as MAE, MAPE, and RMSE. We compare the model metrics
for both price and volatility forecast with each other as well as with the metrics of
the naive estimations.
In terms of price forecasting, both models (SVR and RF) present a good fit
compared to naive estimations with MAPE values of 11.4% and 12.9%, respectively.
The MAPE values of our methods are closer to those of neural network applica-
tions to the market, such as Kolmek and Navruz (2015), Ozguner et al. (2017), and
Ugurlu et al. (2018a). Therefore, SVR and RF can also be seen as effective tools
to predict Turkish electricity market prices. The SVR model appears to capture the
actual prices relatively better than the RF model and in terms of seasons, the forecasts
for the colder periods of the year (autumn and winter) are relatively better captured.
In both modeling, we observe that the forecasts exhibit the lowest level of errors at
hours 11:00, 15:00 and 21:00 of the day in terms of all metrics. Relatively, the error
120 M. Kara et al.

Fig. 3 Volatility forecast plots by years

Price and Volatility Forecasting in Electricity … 121

Fig. 3 (continued)

levels are at their highest during the morning periods. We also experiment with the
SVR model using a reduced number of features, in which short term lagged values
and moving averages (LAG1, LAG5, SMA1, SMA5) are dropped. We observe that the
MAPE is 15.6% in that case (note that this is the value when negative forecasts are
adjusted to zero).
In terms of volatility forecasting, RF produces highly better forecasts compared
to both SVR and naïve estimations. When forecasts are plotted together with the
realized volatility values, it is observed that both SVR and RF methods present a
good fit, especially for the years 2017 and 2018. RF is superior to SVR in capturing
the spikes for all years. The superior performance of RF is very visible in the years
2015, 2016 and 2019.
Overall, in the scope of the research, we implement a framework in a rolling
data scheme covering approximately 6.5 years of the forecasting horizon. Within
this framework, we provide two perspectives to forecasting (price and volatility)
with decent accuracy and experiment with different sets of features for providing
122 M. Kara et al.

practical use. We provide a comprehensive set of tests and comparisons to measure

the performances of different models that we set up. In general, the findings reveal
SVR as an effective tool for electricity price forecasting in the Turkish day-ahead
electricity market, whereas RF is superior to SVR in forecasting the volatility. The
results reveal both methods as effective tools to be considered along with other
methodologies for forecasting in the Turkish market based on our tests covering a
long forecasting horizon. In a recently liberalized developing electricity market, the
results reveal that both SVR and RF can be beneficial tools to forecast market prices
for all stakeholders, from traders to the academics working in the area.
The future research directions may include the applications to the intraday market
prices. Also, the research on volatility forecasting seems to be untouched for the
Turkish market. The application of other learning methods such as artificial neural
networks or deep learning methods to estimate the volatility of the market can be of
interest for future applications.

References

Aggarwal, S. K., Saini, L. M., & Kumar, A. (2009). Day-ahead price forecasting in Ontario elec-
tricity market using variable-segmented support vector machine-based model. Electric Power
Components and Systems, 37(5), 495–516.
Amjady, N., & Farshid, K. (2009). Day-ahead price forecasting of electricity markets by a new
feature selection algorithm and cascaded neural network technique. Energy Conversion and
Management, 50, 2976–2982.
Armstrong, J. S. (1985). Long-range forecasting: From crystal ball to computer (2nd Ed.). Wiley.
Aydogdu, E. (2016). Asymmetric volatility in European day-ahead power markets: A comparative
microeconomic analysis. Energy Economics, 56, 398–408.
Avci, E., Ketter, W., & Heck, E. V. (2018). Managing electricity price modeling risk via ensemble
forecasting: The case of Turkey. Energy Policy, 123, 390–403.
Breiman, L. (2001). Random forests. Machine Learning, 45, 5–32.
Brusaferri, A., Matteucci, M., Portolani, P., & Vitali, A. (2019). Bayesian deep learning based
method for probabilistic forecast of day-ahead electricity prices. Applied Energy, 250, 1158–1175.
Büyükşahin, Ü. Ç., & Ertekin, Ş. (2019). Improving forecasting accuracy of time series data using
a new ARIMA-ANN hybrid method and empirical mode decomposition. Neurocomputing, 361,
151–163.
Che, J., & Wang, J. (2010). Short-term electricity prices forecasting based on support vector
regression and auto-regressive integrated moving average modeling. Energy Conversion and
Management, 51, 1911–1917.
Cheng, H., Ding, X., Zhou, W., & Ding, R. (2019). A hybrid electricity price forecasting model
with Bayesian optimization for German energy exchange. Electrical Power and Energy Systems,
110, 653–666.
Debnath, K. B., & Mourshed, M. (2018). Forecasting methods in energy planning models.
Renewable and Sustainable Energy Reviews, 88, 297–325.
Fan, S., Mao, C., & Chen, L. (2007). Next-day electricity-price forecasting using a hybrid network.
IET Generation, Transmission & Distribution, 1(1), 176–182.
Frömmel, M., Han, X., & Kratochvil, S. (2014). Modeling the daily electricity price volatility with
realized measures. Energy Economics, 44, 492–502.
Price and Volatility Forecasting in Electricity … 123

Gao, C., Bompard, E., Napoli, R., & Cheng, H. (2007). Price forecast in the competitive electricity
market by support vector machine. Physica A: Statistical Mechanics and Its Applications, 382(1),
98–113.
Gayretli, G., Yucekaya, A., & Bilge, A. H. (2019). An analysis of price spikes and deviations in the
deregulated Turkish market power market. Energy Strategy Reviews, 26, 100376.
Ghayekhloo, M., Azimi, R., Ghofrani, M., Menhaj, M. B., & Shekari, E. (2019). A combination
approach based on a novel data clustering method and Bayesian recurrent neural network for day-
ahead price forecasting of electricity markets. Electric Power Systems Research, 168, 184–199.
Ghoddusi, H., Creamer, G. G., & Rafizadeh, N. (2019). Machine learning in energy economics and
finance: A review. Energy Economics, 81, 709–727.
González, C., Mira, J. M., & Ojeda, J. A. (2016). Applying multi-output random forest models to
electricity price forecast. Preprints.org, 2016090053.
Haugom, E., & Ullrich, C. J. (2012). Forecasting spot price volatility using the short-term forward
curve. Energy Economics, 34(6), 1826–1833.
Hayfavi, A., & Talasli, I. (2014). Stochastic multifactor modeling of spot electricity prices. Journal
of Computational and Applied Mathematics, 259, 434–442.
Keles, D., Scelle, J., Paraschiv, F., & Fichtner, W. (2016). Extended forecast methods for day-ahead
electricity spot prices applying artificial neural networks. Applied Energy, 162, 218–230.
Kolmek, M. A., & Navruz, I. (2015). Forecasting the day-ahead price in electricity balancing and
settlement market of Turkey by using artificial neural networks. Turkish Journal of Electrical
Engineering & Computer Sciences, 23, 841–852.
Lago, J., De Ridder, F., Vrancx, P., & De Schutter, B. (2018). Forecasting day-ahead electricity prices
in Europe: The importance of considering market integration. Applied Energy, 211, 890–903.
Li, X. R., Yu, C. W., Ren, S. Y., Chiu, C. H., & Meng, K. (2013). Day-ahead electricity price
forecasting based on panel cointegration and particle filter. Electric Power Systems Research, 95,
66–76.
Luo, S., & Weng, Y. (2019). A two-stage supervised learning approach for electricity price
forecasting by leveraging different data sources. Applied Energy, 242, 1497–1512.
Marcjasz, G., Uniejewski, B., & Weron, R. (2019). On the importance of the long-term seasonal
component in day-ahead electricity price forecasting with NARX neural networks. International
Journal of Forecasting, 35(4), 1520–1532.
Mei, J., He, D., Harley, R., Habetler T., & Qu, G. (2014). A random forest method for real-time price
forecasting in New York electricity market. In 2014 IEEE PES General Meeting|Conference &
Exposition, National Harbor, MD (pp. 1–5).
Ozguner, E., Tor, O. B., & Guven, A. N. (2017). Probabilistic day-ahead system marginal price fore-
casting with ANN for the Turkish electricity market. Turkish Journal of Electrical Engineering
& Computer Sciences, 25, 4923–4935.
Ozozen, A., Kayakutlu, G., Ketterer, M., & Kayalica, O. (2016). A combined seasonal ARIMA
and ANN model for improved results in electricity spot price forecasting: Case study in Turkey.
InProceedings of the 2016 Portland International Conference on Management of Engineering
and Technology (PICMET), Honolulu, HI, USA (pp. 2681–2690).
Qu, H., Chen, W., Niu, M., & Li, X. (2016). Forecasting realized volatility in electricity markets
using logistic smooth transition heterogeneous autoregressive models. Energy Economics, 54,
68–76.
Qu, H., Duan, Q., & Niu, M. (2018). Modeling the volatility of realized volatility to improve
volatility forecasts in electricity markets. Energy Economics, 74, 767–776.
Saini, L. M., Aggarwal, S. K., & Kumar, A. (2009). Parameter optimisation using genetic algorithm
for support vector machine-based price-forecasting model in National electricity market. IET
Generation, Transmission & Distribution, 4(1), 36–49.
Sandhu, H. S., Fang, L., & Guan, L. (2016). Forecasting day-ahead price spikes for the Ontario
electricity market. Electric Power Systems Research, 141, 450–459.
124 M. Kara et al.

Shayeghi, H., & Ghasemi, A. (2013). Day-ahead electricity prices forecasting by a modified CGSA
technique and hybrid WT in LSSVM based scheme. Energy Conversion and Management, 74,
482–491.
Tashpulatov, S. N. (2013). Estimating the volatility of electricity prices: The case of the England
and Wales wholesale electricity market. Energy Policy, 60, 81–90.
Ugurlu, U., Oksuz, I., & Tas, O. (2018a). Electricity price forecasting using recurrent neural
networks. Energies, 11(5), 1255.
Ugurlu, U., Tas, O., & Gunduz, U. (2018b). Performance of electricity price forecasting models:
Evidence from Turkey. Emerging Markets Finance and Trade, 54(8), 1720–1739.
Vapnik, V. (1995). The nature of statistical learning theory. Springer.
Wang, F., Li, K., Zhou, L., Ren, H., Contreras, J., Shafie-Khah, M., & Catalão, J. P. (2019).
Daily pattern prediction based classification modeling approach for day-ahead electricity price
forecasting. International Journal of Electrical Power & Energy Systems, 105, 529–540.
Yang, W., Wang, J., Niu, T., & Du, P. (2019). A hybrid forecasting system based on a dual decompo-
sition strategy and multi-objective optimization for electricity price forecasting. Applied Energy,
235, 1205–1225.
Zhang, J. L., Zhang, Y. J., Li, D. Z., Tan, Z. F., & Ji, J. F. (2019). Forecasting day-ahead electricity
prices using a new integrated model. International Journal of Electrical Power & Energy Systems,
105, 541–548.
Forecasting the Hydro Inflow
and Optimization of Virtual Power Plant
Pricing

Sezer Çabuk, Özenç Murat Mert, A. Sevtap Selcuk-Kestel,

and Erkan Kalaycı

1 Introduction

Hydroelectricity power plants (HEPPs) are separated into two main groups, which
are Run-of-River (RoR) HEPPs and HEPPs with Reservoirs in Turkey. RoR HEPPs
is must-run; this means that these power plants are operated if it has inflow being
higher than minimum turbine level. The difference between RoRs and reservoirs is
that the latter can store water and dispatch it in accordance with the power plant
operator’s decision. Therefore, depending on the time and amount of inflow is minor
for reservoirs’ electricity generation. However, RoRs require inflow to generate elec-
tricity, and this implies that seasonality, and weather related events have an effect on
its generation severely.
There is an apparent relationship in the scheme of the trading side, i.e., a nega-
tive correlation between inflow and system price since the inflow affects electricity
generation in both reservoirs and RoRs. Similarly, from the generation side, for
hydroelectricity producers, the aim is to maximize revenue by optimizing available
sources. In other words, dispatch of a hydroelectricity power plant should be sched-
uled by regarding the producing energy at the relatively high priced hours (Bucher,
2011). However, inflow and price are uncertain and have a stochastic role in the
optimization problem. Therefore, inflow and price are needed to be predictable to
optimize the hydro portfolio and to have optimal scheduling (Marco et al., 2014).
The water has an opportunity cost since the water streams to the reservoirs at
no cost, and the variable cost of hydro production is meager, but the amount of

S. Çabuk · Ö. M. Mert · A. S. Selcuk-Kestel (B) · E. Kalaycı

Institute of Applied Mathematics, Middle East Technical University, 06800 Ankara, Turkey
e-mail: [email protected]
S. Çabuk
e-mail: [email protected]
Ö. M. Mert
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 125
A. B. Dorsman et al. (eds.), Applied Operations Research and Financial Modelling
in Energy, https://doi.org/10.1007/978-3-030-84981-8_7
126 S. Çabuk et al.

water available is limited and uncertain. The marginal cost of water (water value)
depends on reservoir volume, the system’s production capacity, supply capacity,
demand expectations, expected spot market prices, and predicted inflow amount to
the reservoir (Seim & Thorsnes, 2007). Hydro inflow forecasting is a critical task in
dam operation and water resource planning and management. Among many studies
on accurate and reliable inflow forecasting, one of the earliest studies that investigate
synthetic streamflow generation to model hydro flow is introduced by Thomas and
Fiering (1962). The dependence between successive observations of hydrological
data captured by time series models ARIMA and Seasonal ARIMA (SARIMA) and
SARIMA with eXogenous variables (SARIMAX) are widely used to analyze the
progress of observed time series and forecast future time series in the hydrological
field (Chov, 1964; Box & Jenkins, 1976; Carlson et al., 1970; Spolia & Chander, 1974;
Tao & Delleur, 1976; Salas et al., 1985; Burlando et al., 1993; Hipel & McLeod, 1994;
Mohan & Vedula, 1995). It is shown that ARIMA and autoregressive Artificial neural
networks (ANN) are suitable to forecast hydro inflow series over short-time and
long-time periods (Valipour et al., 2013; Daniell, 1991) along with Adaptive Neural-
based Fuzzy Inference System (ANFIS), yielding good hydro inflow forecasting
performances (see Karunanithi et al., (1994), Golob et al., (1998), Campolo et al.,
(2003), Cheng et al. (2005), Jain & Kumar (2007), Trichakis et al., (2011), Nawaz
et al., (2015). Besides, genetic programming and support vector machine models are
also utilized for monthly river flow time series (Wang et al., 2009). Furthermore,
Random Forest (RF) is utilized in forecasting the hydro inflow, which produces
good predictions by handling non-linear and non-Gaussian data series, especially in
reservoir operations (see Yang et al., (2017); Papacharalampous & Tyralis (2018)).
In electricity markets, especially in wholesale and retail sides, pricing electricity
is essential. Generation mix is formed by regarding the marginal costs of the different
fuel-typed power plants for each hour to supply the electricity demand. For electricity
producers and traders, forecasting the value of electricity is essential to make a profit.
For this reason, many models have been studied to simulate the day-ahead market
(DAM) hourly prices since the deregulation of electricity markets. These models
include multilinear regression (Vu et al., 2014) and ARIMA family models (Haario
& Kauranne, 2010; Taylor & McSharry, 2007), ANNs (Mandal et al., 2006), fuzzy
logic systems (Mori & Kobayashi, 1995), Kalman Filter (Al-Hamadi & Soliman,
2004), and hybrid models (Xiao et al., 2015; Zheng et al., 2000). Among these,
ANN is known to exhibit higher performance for non-linear multivariate problems
with large data and requires fewer assumptions to find out the relation between input
and output. However, autoregressive models in some cases outperform ANN-based
models due to the seasonality effect. Burger et al. (2004) perform the Spot Market
Price Simulation (SMaPS) Model for Germany market price.
The virtual power plant (VPP) concept in the hydropower business as an addi-
tion to the physical asset of a hydroelectricity power plant can be represented as a
nonstandard product or an option. The physical constraints and execution decisions
belong to the physical part, whereas bidding decisions and settlement are included
in the financial part (Zurborg, 2010). Although physical constraints can be included
in the VPP concept, it distinguishes the technical operation from economic dispatch.
Forecasting the Hydro Inflow and Optimization … 127

The optimal operation in the VPP in both day-ahead and balancing markets as a
two-stage stochastic mixed-integer linear programming to maximize the expected
profit of generation companies using CVaR (Tajeddini et al., 2014), or in weekly
virtual power plant profit subject to the long-term bilateral contracts and technical
constraints to define the optimal dispatch problem formulated as a mixed-integer
linear programming (Pandžic et al., 2013) can be shown as optimization approaches
done in the literature. Moreover, VPP can be used for balancing or hedging portfolios
by power traders. VPP option buyers get a right to use capacity within the scope of
defined contractual constraints, such as marginal cost, operating hours and amount
of capacity (Burger et al., 2014).
As an emerging country, Turkey has a growing demand for electricity which is
commonly supplied by natural gas and LNG, hydro, import and hard coal, lignite
and wind. Moreover, investments in renewable energy sources have also been rising
in recent years. As an exported commodities, natural gas and LNG, which are the
most significant energy source in Turkey and their supply have a strong dependence
on international markets, and hydraulic capacity is taken as the most significant terri-
torial source in the Turkish electricity portfolio. Additionally, VPP option is a brand
new concept in the Turkish electricity market, and VPPs are traded by electricity
producers and power traders through bilateral agreements or auctions. Due to the
necessity and vitality of the hydraulic capacity in the Turkish electricity market,
hydro inflow forecasting, its optimization, pricing and risk management brings in
enormous advantage in energy planning. A justifying reason for this is that the hydro-
electric power comprises run-of-river hydropower plants and dammed hydropower
plants in Turkey, and these cover 35% of the total installed capacity according to
the figures by the end of 2015. Therefore, forecasting inflow with high accuracy is
essential for effective hydro optimization, weather derivatives, virtual power plant
pricing and volume risk management.
Having this as the main motivation, we construct an optimization model formu-
lated as the mixed-integer programming that calculates the fair value of defined VPP
for a given time horizon using the generated inflow and price scenarios. In line with
the literature (Burger et al., 2004), inflow scenarios for stochastic behavior of elec-
tricity prices with the purification of seasonality and 24-h periodicity by using the
SARIMA model are constructed for each month separately. In doing so, we aim to
achieve to (i) forecast hydro inflow and generate inflow scenarios for a defined virtual
power plant (VPP), (ii) simulate electricity prices by using historical hourly shapes,
(iii) optimize the hydraulic capacity of VPP by regarding inflow and price scenarios,
(iv) calculate the value of VPP as a financial contract, and (v) propose a hydro inflow
modeling, VPP pricing, and volume risk management by considering power traders
and producers. The outcome of this study is expected to guide policy makers and
investors on optimal contribution of hydropower capacity in energy markets in an
emerging market such as Turkey.
128 S. Çabuk et al.

2 Turkish Electricity Market

Turkish energy sector consists of electricity, natural gas, petroleum and LPG markets
which are regulated by the Energy Market Regulatory Authority (EMRA). Although
state-owned authority dominates these markets, the privatized portion in financial
markets is increasing year by year.
Even though the electricity generation with hydropower goes back to earlier
in the 1900s, a remarkable milestone started in 1970. The state controlled “Turkish
Electricity Administration Commission” (TEK), a monopolistic electricity market,
conducts the generation, transmission, and distribution of electricity for 23 years.
Afterward, between 1994 and 2003, the generation and transmission of electricity
is carried out by the “Turkish Electricity Joint Stock Company” (TEAS), and the
distribution is separated and administrated by the “Turkish Electricity Distribution
Joint Stock Company” (TEDAS). Between these years, two important hydropower
sources, Karakaya Dam (1800 MW, in 1987) and Ataturk Dam (2400 MW, in 1994),
are constructed, which still hold the largest installed capacity in Turkey and lead
to the hydroelectricity portfolio to develop substantially. After 1993, because of the
increase in the number of natural gas power plants, demand for thermal capacity
has also risen. On the other hand, to incorporate the private sector for building and
operating the electricity generation, the “Built Operate and Transfer (BOT) Model”
is released in 1984, allowing the private investors to produce the electricity and
sell it to the national grid, the state-owned electricity authority, or even to private
end-users (Ozturk et al., 2007). In 2004, TEAS was divided into three subgroups:
“Turkish Electricity Generation Joint Stock Corporation” (EUAS), “Turkish Elec-
tricity Transmission Joint Stock Corporation” (TEIAS) and “Turkish Electricity
Trading Joint Stock Corporation” (TEDAS). Moreover, TEDAS is separated into
21 distribution regions, and as of 2013, share transfer agreements for their privati-
zation are completed (TEDAS, 2015). On the energy generation side, privatization
is improved, especially for the thermal portfolio of EUAS. In 2015 “Turkish Energy
Stock Market” (EPIAS) is established having its shareholder structure consisting
of 3 groups: (i) Type A (30%) belongs to TEIAS, (ii) Type B (40%) belongs to the
national stock market, Borsa Istanbul (BIST), and Type C (30%) is allocated to market
participants and can be transferred among the companies. The historical outline of
the Turkish electricity market is also summarized in Fig. 1.
The Turkish Electricity market is divided into four sub-market: (i) the derivatives
market, (ii) the day-ahead market (DAM), (iii) the intraday market, and (iv) the
balancing power market (BPM) (Balancing and Settlement Regulation-EMRA, Law
no. 6446) (EMRA, 2015). The Derivative market is mid- and long-term market
(weeks/months/years) in which optimization of production and meeting supply needs
are the essentials. This new market, called Futures and Options Market (VIOP), is
operated by BIST and regulated jointly by Energy Market Regulatory Authority
(EMRA), Capital Markets Board (CMB) and Competition Authority (CA). Whereas
the Day-ahead market, which has a short-term, (D + 1), time horizon and intact with
the market since 2011, requires the equilibrium between generation and consumption.
Forecasting the Hydro Inflow and Optimization … 129

Fig. 1 Timeline on Turkish electricity market’s development

It is operated by EPIAS and regulated by EMRA and CA. Another one whose market
operators and regulators are the same as DAM is the Intraday market (very short-
term/hours) introduced in 2015. The most recent one is the balancing power market
operated by TEIAS for the system’s security, which is the real-time market and
regulated by EMRA and CA (EMRA, 2015).
Turkish electricity generation portfolio by companies comprises of Independent
Power Producers (IPPs) (64.73%), State-owned Power Plants (EUAS and power
plants under EUAS control) (21.22%), Build Operate and Transfer (BOT) Power
Plants, Build Operate (BO) Power Plants, The Autoproducers, The Transferring of
Operating Rights (TOR) Power Plants and Unlicensed Power Plants (14.05%). The
most dominating authority is IPPs (64.73%), whose share is expected in the next
years for the liberalization of the market (TEIAS, 2015).
Since electricity is a non-storable commodity, securing physical delivery is very
crucial for balancing the system. In Turkey, the system operator is the National Load
Dispatch Center, a subsidiary of TEIAS, and manages ancillary services (EMRA,
2015). To understand the physical market operations better, we consider the electricity
market in five parts: (i) bilateral agreements and contracts on physical delivery, (ii)
day-ahead market (DAM), (iii) balancing power market (BPM), (iv) intraday market,
(v) settlement and reporting.
All financial and settlement issues run under EPIAS’s responsibility. For day-
ahead (D + 1) operations, bilateral agreements, bid and offers, options, forwards,
and contracts for physical electricity delivery are logged into its online system the
day before within a certain time frame. After offer verification, market clearing and
evaluation of objections, the final notification of Market Exchange Price (MEP) (in
130 S. Çabuk et al.

(a) MEP (b) SMP

Fig. 2 Pricing mechanisms in the market

TL/MWh), and transaction volume are issued by EPIAS on the same day after the
closing for the offers (Yorukoglu, 2015). MEP is determined by the intersection of
sales and purchase offers for each hour of the day ahead and is a stepwise function
created for each hour via ranking the available sources of energy to fulfill the demand
with the object of maximizing social welfare (Fig. 2a). Additionally, the balancing
power market (BPM) is a real-time market in that physical delivery is realized,
and according to imbalance direction in the system (loading or de-loading), System
Marginal Price (SMP) is determined (Fig. 2b). By comparing the supply and demand
amounts, loading and de-loading instructions are delivered to the market partici-
pants for the system’s balance. According to its direction, SMP is determined in this
real-time market. The intraday market has real-time operation as well as balancing
power market and also provides trading opportunities for the market participants.
The last issue for the physical market is the settlement and reporting of DAMactivi-
ties, BPMactivities and imbalances. Settlement of day-ahead market is a daily basis
while balancing power market, energy imbalance, and ancillary services are settled
and reported monthly.
In Turkey, hydraulic capacity is the second biggest source for electricity produc-
tion 123.32%, following natural gas 128.44%. Fossil energy sources have third
place in electricity production 122.77%. Hydro-Run of River source 18.64% is
followed by the other renewable energy alternatives, mostly wind power 18.05%.
In the hydropower generation aspect, the three largest Turkish dams, 2. 405 MW
Ataturk Dam with an estimated mean annual generation of 8.17 TWh/year, 1,800 MW
Karakaya Dam with approximately 7.6 TWh/year, and 1,330 MW Keban Dam with
7.01 TWh/year contribute to the electricity generation. These dams are state-run
power plants located at the Fırat River basin, operated by EUAS. Hydraulic installed
capacity (both Run-of-River and Reservoir) is 31.96% of the electricity produc-
tion portfolio generating approximately 91.3 TWh/year in 2019 (TEIAS, 2015). The
annual hydro generation calculated as the ratio of annual generation by the annual
hydro installed capacity with respect to the capacity factors is shown in Fig. 3a. Due
to drought and some other environmental reasons, the fall in the hydraulic genera-
tion fluctuates similar to the capacity factor but shows an increasing trend in the last
decade. The feed-in tariffs (US¢/kWh) and incentives for electricity generation via
utilization of renewable sources for hydro, wind, geothermal, bio-fuel (incl. solid
Forecasting the Hydro Inflow and Optimization … 131

(a) Annual Hydro Generation

(b) Monthly Inflow and Prices

Fig. 3 Hydro inflow, capacity and electricity prices (TEIAS, 2015)

waste), and solar are set by the State to 7.3. 7.3. 10.5. 13.3 and 13.3, respectively.
The power plants commissioned during the period 2005–2015 are expected to benefit
from guaranteed prices for ten years.
Electricity prices depend on fundamental parameters such as electricity demand,
capacity development, climate changes and GDP growth. Additionally, hydraulic
generation capacity plays an important role in electricity production, and there is an
uncontrovertible relation between hydro generation and electricity prices. Signifi-
cantly, must-run hydropower plants affect prices crucially because of taking place
132 S. Çabuk et al.

on the left side of the Merit Order Curve. This implies that there is a negative corre-
lation between hydro inflow and electricity prices. When we observe the monthly
inflow to dams (Mio m3 ) (TEIAS, 2015) and monthly electricity prices (TL/MWh)
(Fig. 3b), these two behave in the opposite direction to each other. Concerning the
annual averages, for example, in 2010, inflow to dams with 76,246 (Mio m3 ) with
a market exchange price of 121.1 (TL/MWh) decreases to 58,947 inflow amount
with an increasing price of 137.9. As 2014 is a dry year, the electricity generation
from hydraulic power is 20% lower than in 2013, leading to an increase in electricity
prices in 2014, which also concludes that forecasting hydro inflow is one of the
critical factors for electricity pricing.

3 The Proposed Methodology

In Turkey, the importance of energy derivatives gets higher with the contributions
such as more conscious market participants, better compatibility in regulation, insti-
tutionalization, recognition of derivatives, and comprehension ability in risk manage-
ment (Basoglu, 2005). While hydropower is significantly dominant in the genera-
tion energy generating portfolio of Turkey as a physical commodity, its importance
as a financial product is increasing as well. Optimizing the capacity against the
spot market is essential for electricity producers. Moreover, optimizing hydro assets
against the forward/futures market and forming hedging strategies are other neces-
sities for the producers. Along with those, the liquidity of trade of hydro capacity as
an option gains importance in the energy market. Having an influence on financial
markets, the Turkish electricity sector is considered in the frame of this motivation
and the choices for capacity are aimed to be considered as a real option and VPP
concept.
In more detail, in hydro flow scenarios, our first target is to determine the inflow
forecast. To do so, we employ the most appropriate method, SARIMAX, as hydro
inputs strongly dependent on weather, illustrating strong seasonal dependence. The
output of this stage, the forecasts, will be the input for the next one at which hydro
optimization algorithm whose results are found using MIP is implemented to forecast
the water capacity for the future terms. On the other hand, we determine the price
behavior, whose values are estimated and resembled using Monte Carlo simulations.
Afterward, having the possible monthly electricity price values along with the hydro
capacity, we employ VPP to estimate the virtual hydropower plant values.
VPP as a new concept in the Turkish electricity market is analyzed in two
subgroups: (i) Financial part, which consists of the valuation of VPP, bilateral
agreements/auctions, and settlement, (ii) Physical part, which considers the physical
delivery and real-world constraints. When the VPP buyer has a right to buy elec-
tricity regarding the constraints of agreement, the VPP seller is obliged to provide
electricity to the counterparty by generating (if production is possible) or buying
from the market (if the VPP seller is not a producer). Therefore, the ability to predict
the value of VPP is very important. Under this setup, an optimization model giving
Forecasting the Hydro Inflow and Optimization … 133

the fair value of defined virtual hydropower plants (VHPP) with respect to inflow
and price scenarios is introduced.
The utilization of VPP in energy studies become very useful as it aims to monitor,
forecast, optimize, and dispatch the generation or consumption of network distributed
energy resources, such as renewable energy resources like solar, wind, hydropower,
and biomass units and flexible power consumers and storage systems (Dielmann &
van der Velden, 2003; Lukovic et al., 2010; Kasaei et al., 2017). The assets aggre-
gated in a VPP can be forecasted, optimized, and traded, which are considered as
a single power plant (Zdrilić et al., 2011). This helps to balance the fluctuations
in the generation of renewables by inclining and declining power generation and
power consumption of controllable units One of the objectives of a VPP is to inte-
grate renewable energy sources into existing markets. Since individual small plants
may not satisfy the minimum bid size of the markets, in general, they do not provide
balancing services or offer flexibility on the power exchanges. Furthermore, there are
exacting requirements about the availability and reliability of the flexibility offered in
the market (Höschle et al., 2017; Lannoye et al., 2012). A VPP can provide the same
service and redundancy by aggregating the power of multiple units and subsequently
trade in the same markets as large central power plants or industrial customers.
The steps of VHPP optimization and pricing are given in Fig. 4, which starts
with the estimation of hydro inflow in terms of its contributing factors. A time series

Fig. 4 The flowchart in determination of the hydro inflow optimization and pricing
134 S. Çabuk et al.

approach that considers the influence of the seasons (SARIMAX) is constructed

(Sect. 3) based on the hourly Day-Ahead-Market prices to generate electricity price
scenarios. Based on the predictive model obtained from this step, hydro optimization
with the association of price scenarios is performed. Using the generated inflow and
price scenarios, we employ MILP optimization to construct various optimization
cases allowing flexibility (Sect. 4). By means of the change in reservoir levels (initial,
end, minimum, maximum) and operating hours, we obtain four different optimization
cases (Case 1–4). In Case 1, VHPP values are obtained according to different initial
and end reservoir levels. In cases 2 and 3, VPP values are evaluated according to
different initial, end, and maximum reservoir levels, respectively. Lastly, in Case 4,
operating hours are constrained, and the virtual hydropower plant is required to be
dispatched only at the peak hours (08:00–19:00/all days in a week). By this way, we
create different states and analyze different VHPP values for the same inflow and
price levels with respect to change in mentioned constraints. This allows us to define
a flexible and reliable tool that values defined VHPP.

4 Hydro Inflow Forecasting

Hydro inflow behavior during a year plays an essential role in estimating the value
of hydroelectricity power plant. The monthly average inflows (m3 s) of Arkun Dam
and Hydroelectricity Power Plant located on Coruh River in the provincial border
of Erzurum and Artvin between October 1979 and September 2011 are taken as
input to build a plausible model. Arkun Dam, whose construction was started in
2010, belongs to EnerjiSA Generation Company, has been in operation since the
second quarter of 2014. It has three 78 MW capacity main turbines and two 5.4 MW
capacity environmental turbines, summing up to 244.8 MW. It has a fourteen km
energy tunnel, the longest tunnel in EnerjiSA projects, and the annual generation
of Arkun HEPP is approximately 780.1 GWh. The data on inflow obtained as an
outcome of a feasibility study and on precipitation (Fig. 5) show strong time and
seasonal dependence. Speedwell precipitation data that consists of various stations
with different weights with a lag of 5 months is taken as an exogenous variable.
Among 52 precipitation stations over Turkey, we select 15 significant stations to
achieve the highest correlation with the inflow series. The weights given for selected
stations are based on their spatial distances to the selected dams, as illustrated in
Table 1.
The basic descriptive statistics of the monthly average inflow of 384 observations
and precipitation data obtained from 15 stations (Table 2) show that both series have
right-skewed distributions and do not follow Normal distribution (Jargue-Bera and
Shapiro Wilk tests). Therefore, as supported by the goodness of fit tests justifying
log-normal characteristics of the original data, log-transformed series are employed
in modeling.
Due to time influence in both series, we check the stationarity using Augmented
Dickey Fuller (ADF) (Dickey & Fuller, 1979) and Kwiatkowski, Phillips, Schmidt
Forecasting the Hydro Inflow and Optimization … 135

Fig. 5 Hydro inflow and precipitation series

Table 1 Station weights for

City name Station ID Weight
indexed precipitation series
ALANYA TURK_17310 0.0249
ANAMUR TURK_17320 0.0656
BALIKESIR SYNOP_WMO_17150 0.0571
BINGOL TURK_17203 0.0039
BODRUM TURK_17290 0.1293
CANAKKALE TURK_17112 0.0433
DALAMAN SYNOP_WMO_17295 0.1939
EDIRNE TURK_17050 0.0334
FINIKE TURK_17375 0.1032
GIRESUN TURK_17034 0.0311
HOPA TURK_17042 0.0915
INEBOLU TURK_17024 0.0075
MUGLA TURK_17292 0.0232
SILIFKE TURK_17330 0.0459
SINOP TURK_17026 0.1463

Table 2 Descriptive statistics

Property Arkun inflow series Precipitation series
Mean 62.230 77.910
Standard deviation 68.071 60.833
Skewness 1.7468 1.3202
Kurtosis 2.3573 1.7131
Shapiro–Wilk 0.7170 0.8737
Jarque_Bera 287.85 160.73
136 S. Çabuk et al.

(a) ACF

(b) PACF

Fig. 6 Inflow (upper left–right) and precipitation (lower left–right) serial dependence

and Schin (KPSS) tests whose p-value is found to be 0.1 for both inflow and
precipitation series, supporting the stationarity (Shumway & Stoffer, 2017).
The autocorrelation function (ACF) and partial ACF (PACF) for a lag of
120 months of original inflow and precipitation data (Fig. 6a, b) also show the
seasonality.
When there exists a seasonal component that repeats every s observations, as in
monthly hydro inflow data showing a cycle, from the beginning of October to the end
of September, SARIMA is expected to capture the seasonality. On the other hand,
as an exogenous variable, precipitation has an important contribution to the inflow
variation. For this reason, SARIMAX for a series Xt (inflow) is expressed in terms
of its own realizations as well as an exogenous variable Yt (precipitation) can be
defined as follows (Shumway & Stoffer, 2017);
   
ARIMA(p, q, d ) × (P, Q, D)s ; P Bs φ(B)∇sD ∇ d Zt = Q Bs θ (B)ωt , (1)

Zt = Xt − β1 Y(1,t) − β2 Y(2,t) − · · · − βb Y(b,t) . (2)

Here, ωt denotes the white noise process, φ(B) and θ (B) of orders p and q, show
AR and MA polynomials, respectively; P (Bs ) and Q (Bs ) of orders P and Q, denote
seasonal AR and MA components with ∇ d = (1 − B)d and ∇sD = (1 − B)D differ-
ence operators. Zt is the auto-correlated regression residuals with paramaters βb ,
b = 1, . . . , k, in case of k exogenous variables.
Based on the ACF-PACF properties, plausible models are fitted on in-sample data.
Among trials of 39 SARIMAX models, the best fitting model is found to be ARIMA
(3, 0, 0)(4, 0, 0)12 with drift performance measures, log-likelihood (130.82), AIC
(131.41) and BIC (170.32), coming up the smallest among others. The estimated coef-
ficients of the SARIMAX model are given in Table 3, whose significance is justified
with p-values < 0.001. The inflow has three lags dependence on its historical obser-
vations, four lags dependence on influential seasonal history with σ 2 = 0.07219.
An increase in the precipitation causes 1.37% incline in the inflow. The observed
Forecasting the Hydro Inflow and Optimization … 137

Table 3 The parameter

Coefficient s.e P-values
estimations of SARIMAX
AR1 0.6811 0.0529 0.000000e+00
AR2 −0.0934 0.0628 1.371190e−01
AR3 0.1632 0.0525 1.902464e−03
SAR1 0.1233 0.0528 1.954585e−02
SAR2 0.3243 0.0491 3.841327e−11
SAR3 0.3387 0.0506 2.172418e−11
SAR4 0.1750 0.0546 1.347480e−03
Intercept 3.6600 0.6410 1.132342e−08
Exogenous 0.0137 0.0229 5.490811e−01

and in-sample estimated values are illustrated in Fig. 7a, which show the estimated
model to imitate the pattern of the original data. For the diagnostic checks on the
residuals on randomness and normality, we perform Bartels test (Bartels, 1982) and
Shapiro–Wilk, respectively. As can be seen in Fig. 7c, residuals are random but not
justifying the Normal distribution (p > 0.05) (Fig. 7b).
Additionally, Mean Error (ME), Root Mean Square Error (RMSE), Mean Abso-
lute Error (MAE), and Mean Absolute Percentage Error (MAPE) are calculated for

(a) In-sample model fit of Inflow (b) Q-Q plot of residuals

(c) Residuals and its ACF Plots (d) ACF and PACF of Squared Residuals

Fig. 7 Estimated SARIMAX model and diagnostic checks

138 S. Çabuk et al.

Table 4 Statistics of
Residuals t-Residuals St-Dev. (t-Residuals)
residuals
Mean 0.00030 −0.01353 0.01346
St-Dev 0.269036 0.22650 0.01550
DF – 6.61958 2.49588
Loglik – −34.5959

checking the efficiency of the training set yielding the values 0.000304, 0.268686,
0.2028392. 0.508461 and 5.471788, respectively, justifying the goodness in fit as
well. On the other hand, the heteroscedasticity check is done. ACF and PACF of
squared residuals (Fig. 7d) show that the seasonality is still persistent but remains
within the 95% confidence band. Lagrange Multiplier (LM) test for autoregressive
conditional heteroscedasticity (ARCH) assures the constant variance (p-value <0.05).
As the residuals are non-normally distributed, the goodness of fit test on t-distribution
is justified (p − value < 0.05) whose certain statistics are shown in Table 4. We esti-
mate the parameters (mean and standard deviation) of t-distribution (Cowan, 2011)
to employ this for simulation and forecasting purposes.
The SARIMAX model is utilized to forecast point estimates of monthly inflows,
as shown in Fig. 8a. The parameters of t-distributed residuals are used to generate
100 scenarios for a 12-months inflow (Fig. 8b). We depict that the monthly inflow
scenarios remain within the 80% confidence band around the point estimates.

5 Virtual Power Plant Pricing

A Virtual Power Plant also requires the determination of price scenarios, which
constitute the input for hydro optimization. Based on the historical spot electricity
market prices, hourly Day-Ahead Market (DAM) prices realized between 2011
and 2015, we aim to fit a model which enables us to obtain price simulations.
Among 43,800 observations, some special dates, national and religious holidays
are extracted. Additionally, the historical electricity prices pose some characteristics
such as seasonal patterns, periodicities and spikes (Burger et al., 2004). Because of
the Natural Gas curtailment for electricity generation in the history, that is insuffi-
ciency in the NG supply for the NG power plants, some spikes (e.g., 2000 TL/MWh)
are eliminated by truncating the data based on mean and standard deviation of the
historical data and obtain truncated hourly prices as seen in Fig. 9. Two-sided three
sigma deviations from the average value of price series are used as truncation limits
and truncated price series. The hourly data pose different characteristics than monthly
ones, which may change for each month. For this reason, we rearrange the prices for
each month and consider the five-year average prices for each month, as shown in
Fig. 10.
Based on each month’s prices, we find the best fitting model supported by the
required diagnostic tests. For each month, 25 different models are fitted whose best
Forecasting the Hydro Inflow and Optimization … 139

(a) Forecast

(b) Scenarious

Fig. 8 Inflow forecasts and simulations based on SARIMAX model

ones are chosen to simulate hourly prices based on monthly behavior. We summarize
the best selected SARIMA models for each month in Table 5. It should be noted
that most of the months come up with similar orders and the same period of 24 h.
For illustration, the time series analyses are presented for January data in Fig. 11.
According to the seasonal behavior, each series corresponding to a month within
a year, we desasonalize the series to eliminate the periodicity. Using the residuals
from deseasonalized models, we combine monthly simulations by adding their values
successively so that the price series for 8760 h can be generated. In Fig. 12, the dark
blue color represents the base case.

5.1 Hydro Power Optimization

After determining the hourly price behavior and hydro inflow pattern, we construct
an optimized system that adjusts itself in position to the contributing variables.
The objective function and its constraints are defined and expressed for virtual
140 S. Çabuk et al.

Fig. 9 Hourly prices (left panel) versus truncated hourly prices (right panel)

hydropower pricing (VHPP) that will be accounted for hourly. The hydro optimiza-
tion problem is constructed by using volume levels of the reservoir, Vk, forecasted
inflow series to the reservoir, Ik, by means of an integrated inflow forecasting model,
market price, pk, electricity generation, Gk, marginal cost for VPP, Ck, and spillage
of water, Sk for the time increments of k = 1, . . . , T . Spillage of water is an unde-
sirable condition for the hydro optimization. Since it is favorable that the spillage
should converge to zero, such a penalty is associated with the constraints (Olivares,
2008).
The volume levels for each time k is a function of a period ahead volume, inflow,
electricity generation and spillage of water expressed as follows:

Vk = Vk−1 + Ik − Gk − Sk
Vk−1 = Vk−2 + Ik−1 − Gk−1 − Sk−1
.. (3)
.
VT = VT −1 + IT − GT − ST

Hence, we set an optimization problem

Forecasting the Hydro Inflow and Optimization … 141

Fig. 10 Five-year average hourly prices for each month (2011–2015)

Table 5 Orders in SARIMA

Month (p, d , q, P, D, Q, s) AIC
for each month-average prices
January 3, 0, 3, 0, 0, 2, 24 with drift 5205.624
February 2, 0, 2, 0, 0, 2, 24 with drift 4660.324
March 3, 0, 2, 0, 0, 2, 24 with drift 5282.268
April 2, 0, 3, 0, 0, 2, 24 with drift 5138.051
May 2, 0, 1, 0, 0, 2, 24 with drift 5529.803
June 1, 0, 4, 0, 0, 2, 24 with drift 5349.095
July 4, 0, 2, 0, 0, 2, 24 with drift 5395.587
August 6, 0, 4, 0, 0, 2, 24 with drift 5468.479
September 3, 0, 2, 0, 0, 2, 24 with drift 5232.774
October 3, 0, 1, 0, 0, 2, 24 with drift 5249.348
November 2, 0, 2, 0, 0, 2, 24 with drift 5116.503
December 2, 0, 4, 0, 0, 2, 24 with drift 5392.938
142 S. Çabuk et al.

Fig. 11 Some important plots of January price and its modeling

Fig. 12 Price scenarios (TL/MWh)


T
 
max E (pk − Ck )Gk − MSk
k=1 (4)
s.t.0 ≤ Gk < Gmax < Gcap ,
0 ≤ Vmin ≤ Vk < Vmax , k = 1, 2, . . . , T ,
Forecasting the Hydro Inflow and Optimization … 143

where M is sufficiently large, with respect to the Initial and End Reservoir Levels
V1 and VT , respectively. The maximum level Vmax , minimum level Vmin , maximum
generationGmax , Installed capacity Gcap and time horizon of the option 1, . . . , T are
given in any special contract as specific constraints.
To solve the proposed nonlinear optimization problem, we employ MILP because
dispatch problems have a mixed linear-integer structure (Eydeland & Wolyniec,
2003) and it restricts the decision variables to be integer during the selection
of optimal solution as a process of minimizing or maximizing a linear function
(Goodarzi et al., 2014). Hourly hydro optimization for defined VPP is constructed
for a hydro year (T = 8760 h) by regarding the generated inflow and price scenarios
under the given constraints by using MILP and Xpress Solver (FICO, 2009). Because
of the hourly bidding in the market, the model is constructed discretely, and it gives
the optimized generation curve by regarding the objective function.
We generate 15 inflow scenarios for each Ik and 100 price scenarios for each pk,
k = 1, . . . , T , which are employed in the optimization model. Price scenarios are
ordered according to the increasing annual average of series where inflow scenarios
are arranged by considering the increasing annual sum. This means that we expect
the highest VPP value at the intersection of the last price scenario and the last inflow
scenario among all scenarios.
Under the PC environment, for 100 prices and one inflow scenario, the run time is
approximately 40 min. However, parallel runs are possible in FICO Xpress Solver,
enabling us to obtain 1500 VPP values approximately in 2 h. For the solution of
joint behavior (a price and an inflow scenario) at which we aim to solve the one-year
hourly (8760 h) problem, the optimization algorithm deals with numerous iterations
in the dimensions of 17,517 constraints (in rows), 35,040 variables (in columns),
52,556 non-zero elements and 8760 global entities.

5.2 VPP Pricing

MILP is also suitable for calculating option delta and implementing the dynamic delta
hedging for VPP (Burger et al., 2014). The optimization results based on certain
constraints with respect to given inflow and price scenarios for defined VPP are
performed with respect to the assumptions as given in Table 6 and called as Base
Case. Here, the maximum stored energy of defined VPP, Vmax is taken as 702 GWh,
whereas V0 and VT are assumed as the half of Vmax based on the expert opinions.

Table 6 Base case

Initial uptime 1
assumptions
Minimum uptime 1
Initial level (GWh) 351
End level (GWh) 351
Maximum level (GWh) 702
144 S. Çabuk et al.

Each VPP obtained from inflow scenarios, the fair value of VPP is found as

1 
n m
V PP = Zi,p , (5)
mn p=1 i=1

It should be noted that VPP is obtained for each inflow and price scenario as the
equally weighted average value of simulated VPPs, Zi,p ’s. It should be noted that joint
information on monthly electricity price values based on n-scenarios along with the
optimized hydro capacity values, i.e., m-inflow scenarios, we calculate the average
price required.
For VPP pricing as a real option, we need to consider additional concepts, such
as intrinsic and extrinsic values. In the frame of the hydropower plant, VPP value
for the expected market price curve gives the intrinsic value of VPP, whereas by
subtracting intrinsic value from its fair value, an extrinsic value for VPP is attained.
According to each price and inflow scenario, we obtain VPP values (TL) as shown
in Fig. 13.
Moreover, to demonstrate the sensitivity of VPP values, for the first price scenario
(named as Case Inflow), change in VPP values according to inflow scenarios and
for the first inflow scenario, change in VPP values according to price scenarios
(called as Case Price) whose patterns are seen in Fig. 14a and b, respectively. The
minimum VPP value is obtained at the intersection of the first price and the first inflow
scenarios. As mentioned before, price scenarios are sorted ascendingly according to
the annual price average. Whereas inflow scenarios are sorted descendingly according

Fig. 13 VPP values for base case

Forecasting the Hydro Inflow and Optimization … 145

(a) Inflow Scenarios

(b) Price Scenarios

Fig. 14 VPP values according case inflow and case price

to the annual total inflow amount. Therefore, we expect the minimum value at the
intersection of the first scenarios and the maximum value at the intersection of the
last scenarios. In Fig. 13, the inflow scenario axis shows that a higher inflow level
leads to higher VPP values as expected. On the other side, when we analyze VPP
values according to the price scenario axis, we conclude that monthly price scenarios
are more significant than shape scenarios for the change in value.
146 S. Çabuk et al.

5.3 Sensitivity to Parameters in Hydro Optimization

Along with the Base case analyses, we depict the influence of some variations on
the parameters in four groups to illustrate the sensitivity of the optimization model
on VPP values. We define four cases to demonstrate the influence of the choices
in the values of variables. The VPP values are obtained based on different initial
and end reservoir levels. Taking into account different initial, end and maximum
reservoir levels, VPPs are found in Case 2 and Case 3. When within the operating
hours the peak hours (08:00–19:00/all days in a week) are considered as the main
constraint, VPP is evaluated as Case 4. Table 7 shows that cases 1–3 account for all
operating hours as in the Base case, whereas Case 4 considers only the peak hours.
The maximum level only in cases 2 and 3 is half of the Base Case. In terms of initial
and end levels, Case 1 and Case 3 have these values to be zero, Case 2 is half, Case
for the same as the Base case.
In Case 1, Vmax is the same as Base Case, V0 = VT = 0, implying that VPP is
defined as empty at the initial time. Therefore, the generation is possible by employing
the inflow for the initial time regardless of the price levels. Optimization results with
these assumptions (Fig. 15a) in comparison to Base Case yield VPP values much
lower than Base Case. Moreover, another difference in Case 1 from Base Case is that
the increase in inflow amount causes sharper increases in VPP values. On the other
hand, in lower price levels, VPP values are smoother and have similar sight with Base
Case. However, at the higher price levels, because of the generation decision (holding
water for the higher-priced hours or generating electricity), transitions between price

Table 7 Assumptions for cases, optimized VPP values and VoaR

Base case Case 1 Case 2 Case 3 Case 4
Initial uptime 1 1 1 1 1
Minimum uptime 1 1 1 1 1
Year end 8760 8760 8760 8760 8760
Days 365 365 365 365 365
Initial level (GWh) 351 0 176 0 351
End level (GWh) 351 0 176 0 351
Maximum level (GWh) 702 702 351 351 702
Operating hours All All All All Peak
Optimal VPP prices
Mio. TL Base case Case 1 Case 2 Case 3 Case 4
VPP Value 138.9 125.5 130.8 125.1 120.6
VoaR for each case
Mio. TL Base case Case 1 Case 2 Case 3 Case 4
VoaR −29.8 −21.2 −22.6 −20.7 −15.6
Forecasting the Hydro Inflow and Optimization … 147

(a) Case 1

(b) Case 2

(c) Case 3

(d) Case 4

Fig. 15 The optimization of VPP under different constraints

148 S. Çabuk et al.

scenarios are also sharper than Base Case. Again, in this case, shape scenarios create
small differences in VPP values.
Case 2 aims to compare the value of the defined hydro virtual power plant in
case of the maximum reservoir level is half of the maximum reservoir level in Base
Case, and initial and end reservoir levels to be as half as the ones in Base Case. The
optimization results (Fig. 15b) illustrate that the holding water for higher-priced hours
is constrained by the half of maximum reservoir level since the maximum reservoir
level is lower as expected. Compared to Case 1, because of the non-zero initial
and end levels, smoother surfaces for VPP values are obtained. Additionally, at the
intersection of higher prices and higher inflow scenarios, VPP values are increasing
progressively. Change in constraints leads to change in intrinsic and extrinsic values
of VPP, despite the price scenarios are the same as in Base Case.
In Case 3, VPP is defined as empty at the initial time, so generation is possible
through inflow for the initial time regardless of the price levels. Optimization results
(Fig. 15c) show that since we need to inflow and price scenarios stay the same for the
initial generation, VPP values appear to be lower than Base Case and the increase
in inflow amount cause sharper increases in VPP values as in Case 1 because of the
zero initial and end reservoir levels. On the other hand, at lower price levels, VPP
values are smoother and have similar sight with Base Case. However, at the higher
price levels, because of the generation decision (holding water for the higher-priced
hours or generating electricity), transitions between price scenarios are also sharper
than Base Case as in Case 1.
In the last case, Case 4, we assume that Vmax , V0 , and VT are the same as Base
Case, but VPP could produce electricity at only peak hours (08:00–19:00/all days
in a week). Optimization under these assumptions (Fig. 15d) leads to a decrease in
VPP value since a penalty is imposed for spillage in the optimization model.
By considering four different cases, our aim is to demonstrate the effects of initial,
end, and maximum reservoir levels and operating hours. Case 1 presents the effects
of initial and end reservoir levels; Case 2 shows the influences of both initial and end
levels, and maximum levels; Case 3 offers a situation concentrating on the change in
maximum reservoir level; Case 4 explains a situation of taking the operating hours
at peak. Figure 15 indicates that zero initial and end reservoir levels result in a sharp
increase in the surface VPP values, the change in maximum reservoir levels causes
a smoother and sharper type of VPP surfaces with respect to lower and higher price
levels, respectively, and letting the operating hours only for peak makes VPP values
to decrease.
In VPP prices for each case are shown in Table 7. By using an hourly price forward
curve or market expectation, intrinsic value can be calculated for VPP. We need to
note that if a stock has a significantly lower intrinsic value than market value, its
value is overestimated. Therefore, according to the trade side, i.e., option buyer or
seller, the intrinsic value must also be taken into account.
Defining constraints of VPP allows constructing the deal terms for a targeted
trading position better. Moreover, the ability to calculate the fair value of VPP with
the various constraints helps to understand the sensitivity of VPP to fundamentals,
that is, inflow in our study and electricity price. Therefore, according to the risk
Forecasting the Hydro Inflow and Optimization … 149

appetite of the trader, VPP constraints are negotiated as contract terms between VPP
buyer and seller. In addition to the risk-focused approach, a study on different cases
is essential to model the technical flexibility of power plants because, under some
circumstances, operational hours can be limited for the technology of defined VPP.

5.4 Volume at Risk

As we determine both price and volume risk for the hydroelectricity power plant, the
volatility in prices should also be encountered as the price movements affect the value
of generated electricity (Fiorenzani, 2006). For this reason, we need to determine the
maximum risk to be taken at a certain level of confidence. Which levels of electricity
prices can oscillate, and which level the value of VPP can decline can be answered
by considering the volume at risk in the similar frame of Value at Risk (VaR). On
the other hand, the Volume at Risk concept is expressed in defining a risky volume
calculated using the distribution of VPP values obtained from the optimization model.
The Volume at Risk (VoaR) for each optimization case shown in Table 7 gives the
highest risk for the base case, the lowest in Case 4.

6 Concluding Comments

VPP deals are traded actively on the electricity exchange markets, mostly as a trading
instrument to hedge natural positions, especially for peak seasons, days, or hours.
In the Turkish Electricity Market, most liquid products are base-load shaped ones.
This means that with the purpose of capturing a profile of electricity, VPP deals are
quite effective. By regarding the system operator, in the long run, the profile-based
imbalance can be reduced with respect to the increase in VPP trading volume.
This study aims to establish a broad framework forVirtual Power Plant (VPP)
pricing. For this purpose, we work on forecasting hydro inflow and generate inflow
scenarios for a defined VPP, simulating electricity prices by using historical hourly
shapes, optimizing the hydraulic capacity of VPP by regarding inflow and price
scenarios. Moreover, we calculate the value of VPP as a financial contract, generate
various optimization cases by use of the constraints of the optimization model and
aim to contribute to the literature of hydro inflow modeling, VPP pricing, and volume
risk management by considering power traders and producers. In consideration of
real options in financial markets, we define a Virtual Hydro Power Plant with the
possible real-world constraints such as reservoir levels, generation constraints, costs.
This study presents a complete step by stepwork for the purposes of valuation for a
defined VPP. The ability to define and estimate the value of a VPP is very beneficial
for the developing derivative market, especially in emerging markets as Turkey.
It establishes all details of the VPP concept and contributes to risk management
150 S. Çabuk et al.

strategies in the electricity market. Using this constructed flexible pricing tool that
values defined VHPP, options market making can be easier for power traders.

References

Al-Hamadi, H. M., & Soliman, S. A. (2004). Short-term electric load forecasting based on Kalman
filtering algorithm with moving window weather and load model. Electric Power Systems
Research, 68(1), 47–59.
Bartels, R. (1982). The rank version of von Neumann’s ratio test for randomness. Journal of the
American Statistical Association, 77(377), 40–46.
Basoglu, T. (2005). Energy derivatives and its applicability in turkey. Unpublished M.Sc. Thesis,
Uludag University.
Box, G. E. P., & Jenkins, G. M. (1976).Time series analysis forecasting and control. Holden-Day,
Inc., 1976.
Bucher, M. (2011). Hydro power planning optimization. Unpublished M.Sc. Thesis. Swiss Federal
Institute of Technology (ETH), Zurich.
Burger, M., et al. (2004). A spot market model for pricing derivatives in electricity markets.
Quantitative Finance, 4(1), 109–122.
Burger, M., Graeber, B., & Schindlmayr, G. (2014). Managing energy risk: a practical guide for
risk management in power, gas and other energy markets. Wiley.
Burlando, P., et al. (1993). Forecasting of short-term rainfall using ARMA models. Journal of
Hydrology, 144(1–4), 193–211.
Campolo, M., Soldati, A., & Andreussi, P. (2003). Artificial neural network approach to flood
forecasting in the River Arno. Hydrological Sciences Journal, 48(3), 381–398.
Carlson, R. F., MacCormick, A. J. A., & Watts, D. G. (1970). Application of linear random models
to four annual streamflow series. Water Resources Research, 6(4), 1070–1078.
Cheng, C.-T., et al. (2005). Long-term prediction of discharges in Manwan Hydropower using
adaptive-network-based fuzzy inference systems models. In International Conference on Natural
Computation. Berlin, Heidelberg: Springer.
Chov, V. T. (1964). Handbook of applied hydrology: A compendium of water-resources technology.
McGraw-Hill Book Company.
Cowan, G. (2011). Monte Carlo techniques. Royal Holloway University (RHUL).
Daniell, T. M. (1991). Neural networks. Applications in hydrology and water resources engineering.
National Conference Publication-Institute of Engineers.
Dickey, D. A., & Fuller, W. A. (1979). Distribution of the estimators for autoregressive time series
with a unit root. Journal of the American Statistical Association, 74(366a), 427–431.
Dielmann, K., & van der Velden, A. (2003). Virtual power plants (VPP)-a new perspective for
energy generation?. In Proceedings of the 9th International Scientific and Practical Conference
of Students, Post-graduates Modern Techniques and Technologies, 2003. MTT 2003. IEEE.
EMRA. (2015). www.epdk.org.tr/index.php/elektrik-piyasasi/mevzuat?id=36.
Eydeland, A., & Wolyniec, K. (2003). Energy and power risk management. Wiley.
FICO. (2009). Xpress Mosel User Guide, Fair Isaac Corporation.
Fiorenzani, S. (2006). Quantitative methods for electricity trading and risk management. Palgrave
Macmillan.
Golob, R., Štokelj, T., & Grgic, D. (1998). Neural-network-based water inflow forecasting. Control
Engineering Practice, 6(5), 593–600.
Goodarzi, E., Ziaei, M., & Hosseinipour, E. (2014). Introduction to optimization analysis in
hydrosystem engineering. Springer International Publishing.
Forecasting the Hydro Inflow and Optimization … 151

Haario, H., & Kauranne, T. (2010). A comparison of electricity spot prices simulation using ARMA
GARCH and mean reverting models. Unpublished M.Sc. Thesis, Lappeenranta University of
Technology.
Hipel, K. W., & McLeod, A. I. (1994). Time series modelling of water resources and environmental
systems. Elsevier.
Höschle, H., et al. (2017). Electricity markets for energy, flexibility and availability—Impact of
capacity mechanisms on the remuneration of generation technologies. Energy Economics, 66,
372–383.
Jain, A., & Kumar, A. M. (2007). Hybrid neural network models for hydrologic time series
forecasting. Applied Soft Computing, 7(2), 585–592.
Karunanithi, N., Grenney, W. J., Whitley, D., & Bovee, K. (1994). Neural networks for river flow
prediction. Journal of Computing in Civil Engineering, 8(2), 201–220.
Kasaei, M. J., Gandomkar, M., & Nikoukar, J. (2017). Optimal management of renewable energy
sources by virtual power plant. Renewable Energy, 114, 1180–1188.
Lannoye, E., Flynn, D., & O’Malley, M. (2012). Evaluation of power system flexibility. IEEE
Transactions on Power Systems, 27(2), 922–931.
Lukovic, S., et al. (2010). Virtual power plant as a bridge between distributed energy resources and
smart grid. In 2010 43rd Hawaii International Conference on System Sciences. IEEE.
Mandal, P., Senjyu, T., & Funabashi, T. (2006). Neural networks approach to forecast several hour
ahead electricity prices and loads in deregulated market. Energy Conversion and Management,
47(15–16), 2128–2142.
Marco, J., Harboe, R., & Salas, J. (2014). Stochastic hydrology and its use in water resources
systems simulation and optimization. Springer Science and Business Media.
Mohan, S., & Vedula, S. (1995). Multiplicative seasonal ARIMA model for longterm forecasting
of inflows. Water Resources Management, 9(2), 115–126.
Mori, H., & Kobayashi, H. (1995). Optimal fuzzy inference for short-term load forecasting. In
Proceedings of Power Industry Computer Applications Conference. IEEE.
Nawaz, N., Harun, S., & Talei, A. (2015) Application of adaptive network-based fuzzy infer-
ence system (ANFIS) for river stage prediction in a tropical catchment. Applied Mechanics and
Materials, 735. Trans Tech Publications Ltd.
Olivares, M. (2008). Optimal hydropower reservoir operation with environmental requirements.
Unpublished M.Sc. Thesis, University of California.
Ozturk, K., Yilanci, H., & Atalay, O. (2007). Past, present and future status of electricity in Turkey
and share of energy sources. Renewable and Sustainable Energy Reviews, 11, 183–209.
Pandžic, H., et al. (2013). Offering model for a virtual power plant based on stochastic programming.
Applied Energy, 105, 282–292.
Papacharalampous, G. A., & Tyralis, H. (2018). Evaluation of random forests and Prophet for daily
streamflow forecasting. Advances in Geosciences, 45.
Salas, J. D., Tabios III, G. Q., & Bartolini, P. (1985). Approaches to multivariate modeling of water
resources time series 1. JAWRA Journal of the American Water Resources Association, 21(4),
683–708.
Seim, T. O., & Thorsnes, O. R. (2007). Analyzing the price- and inflow relationships in hydroelectric
scheduling. Unpublished M.Sc. Thesis.
Shumway, R. H., & Stoffer, D. S. (2017). Time series analysis and its applications: with R examples.
Springer.
Spolia, S. K., & Chander, S. (1974). Modelling of surface runoff systems by an ARMA model.
Journal of Hydrology, 22(3–4), 317–332.
Tajeddini, M. A., Rahimi-Kian, A., & Soroudi, A. (2014). Risk averse optimal operation of a virtual
power plant using two stage stochastic programming. Energy, 73, 958–967.
Tao, P.-C., & Delleur, J. W. (1976). Seasonal and nonseasonal ARMA models in hydrology. Journal
of the Hydraulics Division, 102(10), 1541–1559.
Taylor, J. W., & McSharry, P. E. (2007). Short-term load forecasting methods: An evaluation based
on European data. IEEE Transactions on Power Systems, 22(4), 2213–2219.
152 S. Çabuk et al.

TEDAS. (2015). www.tedas.gov.tr/en/pages/aboutus.aspx.

TEIAS. (2015). www.teias.gov.tr/eng/dispatchreports.aspx.
Thomas, H., & Fiering, M. (1962). Mathematical synthesis of stream ow sequences for the analysis
of river basin by simulation, in design of water resource systems. Harvard University Press.
Trichakis, I., Nikolos, I., & Karatzas, G. P. (2011). Comparison of bootstrap confidence intervals
for an ANN model of a karstic aquifer response. Hydrological Processes, 25(18), 2827–2836.
Valipour, M., Banihabib, M. E., & Behbahani, S. M. R. (2013). Comparison of the ARMA, ARIMA,
and the autoregressive artificial neural network models in forecasting the monthly inflow of Dez
dam reservoir. Journal of hydrology, 476, 433–441.
Vu, D. H., Muttaqi, K. M., Agalgaonkar, A. P. (2014). Short-term load forecasting using regression
based moving windows with adjustable window-sizes. In 2014 IEEE Industry Application Society
Annual Meeting (pp. 1–8). IEEE.
Wang, W.-C., Chau, K.-W., Cheng, C.-T., & Qiu, L. (2009). A comparison of performance of
several artificial intelligence methods for forecasting monthly discharge time series. Journal of
Hydrology, 374(3–4), 294–306.
Xiao, L., Wang, J., Yang, X., & Xiao, L. (2015). A hybrid model based on data preprocessing for
electrical power forecasting. International Journal of Electrical Power & Energy Systems, 64,
311–327.
Yang, T., Asanjan, A. A., Welles, E., Gao, X., Sorooshian, S., & Liu, X. (2017). Developing reservoir
monthly in ow forecasts using artificial intelligence and climate phenomenon information. Water
Resources Research, 53(4), 2786–2812.
Yorukoglu, S. (2015). Mathematical modeling and solution approaches for balancing Turkish
electricity day ahead market. Unpublished M.Sc. Thesis, Middle East Technical University.
Zdrilić, M., Pandžić, H., & Kuzle, I. (2011). The mixed-integer linear optimization model of virtual
power plant operation. In 2011 8th International Conference on the European Energy Market
(EEM). IEEE.
Zheng, T., Girgis, A. A., & Makram, E. B. (2000). A hybrid wavelet-Kalman filter method for load
forecasting. Electric Power Systems Research, 54(1), 11–17.
Zurborg, A. (2010). Unlocking customer value: the virtual power plant. World Power 1–5.
Comparing the Renewable Energy
Technologies via Forecasting Approaches

Fazıl Gökgöz and Fahrettin Filiz

1 Introduction

During the last decades, attention has turned to the use of renewable energy sources
from fossil energies. The share of renewable energy sources in the world continues
to increase significantly. There are different renewable energy types, such as wind
energy, hydro energy, solar energy. It has become obligatory among countries to
manage energy resources most appropriately. The fluctuation of renewable power
can influence the quality of power systems. An integrated energy forecasting system
is necessary for the sustainable development of power systems. The need for a stable
and sustainable renewable energy supply is constantly increasing among the world.
The primary energy consumption in the world is about 400 Exajoules/year, 80%
of primary energy is provided by fossil sources. Renewable energies supply 14% of
the primary energy. Nuclear energy provides 6%. When comparing the potential of
renewable energy sources hydro energy has a potential of 50 Exajoules/year, biomass
has a potential 276 Exajoules/year, solar energy has a potential 1575, wind energy
has a potential of 640 Exajoules/year, geothermal energy has a potential of 5000
Exajoules/year (Fridleifsson, 2003).
Currently, fossil sources are still the most used energy source in the world. Fossil
energy resources account for 80% of the world’s energy use. If the current consump-
tion continues worldwide, it is forecasted that energy consumption will increase
by over 50% before 2030 (Suganthi & Samuel, 2012). Renewable electricity can
be the largest power generation source in 2040. The fluctuation of renewable power
generation has important challenges, especially in the integration (Chen et al., 2020).

F. Gökgöz (B)
Quantitative Methods Division, Faculty of Political Sciences, Department of Management,
Ankara University, Ankara, Turkey
e-mail: [email protected]
F. Filiz
Social Sciences Institute, Doctoral Program in Management, Ankara University, Ankara, Turkey

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 153
A. B. Dorsman et al. (eds.), Applied Operations Research and Financial Modelling
in Energy, https://doi.org/10.1007/978-3-030-84981-8_8
154 F. Gökgöz and F. Filiz

The chaotic and volatile nature of renewable data increases these challenges. Many
models are studied in the literature to get more accurate renewable energy prediction
results.
The use of renewable energy resources continues to increase in the global energy
market. Renewable energy cost decline and technology improvements are the most
important factors in the occurrence of this situation. These factors have brought a
positive perspective to renewable energy technologies all over the world. It is reported
that the feasibility and economic effects of renewable energy systems are developing
for many countries (Gulagi et al., 2020).
Another driver of the development of renewable energy sources is climate change
issues (Kholodnyi, 2014). Fossil fuels have negative effects on climate such as global
warming. Renewable energy is clean and has low carbon emissions. Renewable
energies do not emit gas or produce waste. This proliferates greater attention by
researchers from all over the world to investigate more in this type of energy.
Renewable energy forecasting is a useful tool for generators and consumers
to design and implement their plans. Recently, several models are being used for
renewable energy generation forecasting in literature. In this study, the LSTM,
GRU, multiple linear regression, and polynomial regression are implemented for
the prediction of renewable energy generation in Turkey.
The major objectives of this study are as follows.
1. To compare the deep learning methods with linear regression and polynomial
regression methods.
2. To forecast wind energy and hydro energy using different methods and analyze
our findings.
3. To analyze wind energy and hydro energy behavior and compare their
forecasting performance.
The rest of this study is organized as follows. Section 2 gives literature on the
wind and hydro forecasting methods. Section 3 gives a brief introduction to applied
methods. Section 4 is dedicated to a dataset, data preprocessing, and empirical
study on wind and hydro forecasting with the help of linear regression, polynomial
regression GRU, and LSTM. Finally, conclusions were discussed.

2 Literature Review

2.1 Wind Energy Forecasting

Wind energy is the fastest and most widely used renewable energy to meet the
increasing energy needs. Likewise, the production capacity of wind energy is
increasing rapidly around the world. Along with this situation, the volatility of wind
energy adversely affects the quality of the power systems already existing. To reduce
Comparing the Renewable Energy Technologies … 155

the harmful effect of wind energy integration, it is necessary to plan energy produc-
tion with wind energy. Therefore, an accurate wind energy forecast system is essential
for power system operators.
Wind energy potential is affected by weather conditions mainly the air mass
movements (Durak & Şen, 2002). Wind energy has a volatile characteristic over
time. This situation makes it difficult to forecast wind energy. However, wind energy
forecasting optimization is more important for huge energy importing countries like
Turkey.
Many researchers have proposed methods to predict wind energy generation.
These approaches can be grouped into three headings: physical, statistical, and intel-
ligent forecasting. Physical forecasting models are based on weather conditions.
The statistical forecasting models attempt to find a relationship from the acquired
historical data. Regression methods are conventional statistical models, while deep
learning forecasting models are generally trying to solve complex relationships
among variables with different learning algorithms. LSTM and GRU models are
in these groups.
Wind speed and direction parameters are used in the polynomial regression
method to forecast wind energy (Stathopoulos et al., 2013). Wu et al. reviewed deep
learning models for wind energy forecasting. They argued recurrent neural networks
(RNN) and LSTM models to predict wind energy. They compare them according
to MAE, MSE, and mean average percentage error (MAPE) (Wu et al., 2016). Cao
and Gui proposed a model of wind energy forecasting with the help of LSTM. They
compare proposed models with artificial neural networks (ANN) and Support Vector
Machine (SVM) models and found that the MAPE and the MAE of their proposed
model are better than those using neural network and SVM models (Cao & Gui,
2019). It is observed that LSTM methods are actively used in energy prediction
models. LSTM method is used as a prediction method in this study because of its
success in predicting.
Niu et al. proposed a wind energy forecasting method with the GRU method.
They used input matrix as wind energy, wind speed, wind direction, temperature,
air pressure, air density, and hour data. They compare methods with other neural
networks and support vector models (Niu et al., 2020). Ke et al. proposed a load
forecasting method with gated recurrent units (GRU) method. They used power
load, weather data, and holiday data as input to the model. With auto-encoding input
data is compressed and GRU uses it for forecasting. After creating models, they
also compare their GRU results with SVM and LSTM models (Ke et al., 2019).
Considering the relevant literature; temperature, wind speed, cloud cover, humidity,
dew point are used in this study.
Zu and Song combine wavelet decomposition and gated recurrent units to forecast
the wind energy. They argued that the GRU model is more efficient than LSTM and
simpler than LSTM. They compared proposed models with GRU models (Zu & Song,
2018). GRU is another method that used in this study. GRU and LSTM models have
best forecasting results in this study. GRU’s MAE and MSE results are similar to
LSTM results.
156 F. Gökgöz and F. Filiz

2.2 Hydro Energy Forecasting

Hydro energy continues to develop, and many developing countries abandon fossil-
based energy resources and turn to renewable energy resources, such as hydro energy
and wind energy. Hydro energy depends on water management. Its energy refers to
the process of converting the kinetic energy of water into electrical energy. This leads
to the storing of energy and provides the opportunity to produce according to power
needs. The storage of energy and using it according to the need creates a significant
advantage.
However, the daily operation of hydro energy highly depends on weather and
climatic conditions. Hydro energy supports sustainable, safe, and competitive energy
systems that have been influenced by other renewable energy sources, such as wind
and photovoltaic (Ardizzon et al., 2014). The techniques used for wind energy fore-
casting are similar to hydro energy forecasting techniques. Multiple linear regression
methods and polynomial terms were used in load forecasting for power systems. It
is concluded that multiple linear regression is relatively easy to develop, and it is a
contemporary method. MAPE values change between 3.52 and 4.34% (King, 2007).
Multiple linear regression and principal component analysis are used for weekly
electrical load and calculated that the MAPE of the method is 3.68%. The prin-
cipal component is used as input for multiple linear regression methods. Also, it is
concluded that multiple linear regression methods have the simplicity of skipping the
need to train the model (Torkzadeh et al., 2014). Hydro energy dam behavior under
different environmental conditions is compared with the multiple linear regression
and neural network models (Mata, 2011).
Coulibaly et al. study on hydro energy forecasting systems with recurrent neural
networks and principal component analysis. They compare model results with root
mean squared error (RMSE) and argue that recurrent neural performance is better
than autoregressive and autoregressive mean average models (Coulibaly et al., 2000).
Niu et al. investigate the performances of regression, neural network, extreme
learning machine, and SVM methods to forecast hydro energy. They indicate that
ANN, extreme learning machines, and SVM provide better performances than linear
regression (Niu et al., 2019).
Akpinar et al. study on hydro energy generation with regression analysis
concluded that Turkey’s hydro energy generation is going to increase, and the hydro
energy share within electricity energy is going to decrease (Akpinar et al., 2011).
Grey prediction model with integrated polynomial regression was used in hydro
energy generation forecasting in Japan and evaluated by MAE and MSE (Li et al.,
2016). Ma et al. studied LSTM methods to make projections on hydro energy for
2009–2016. They compared the results of the proposed technique with the results of
the linear regression technique (Ma et al., 2018).
Li et al. proposed a hybrid model to forecast hydro energy. They compare proposed
models with the historical average, autoregressive moving average, gated recurrent
unit, seasonal autoregressive integrated moving average, LSTM models. They used
RMSE, MAE, and MAPE for evaluation criteria (Li et al., 2019). Li et al. proposed
Comparing the Renewable Energy Technologies … 157

the hybrid model based on the echo state network and Bayesian regularization to
forecast hydro energy. They compared hybrid models with the feed-forward neural
networks and generic echo state networks are also employed. They concluded that
the hybrid model outperforms both ANN and echo state networks (Li et al., 2015).
In this study, we compare regression and deep learning methods within the
Turkey case. This comparison gives an idea about traditional and current forecasting
methods. In further, both wind energy and hydro energy forecasting performance
have also been evaluated.

3 Brief Information of the Applied Methods

3.1 Principal Component Analysis (PCA)

Principal component analysis (PCA) aims to reduce the dimension of a dataset. PCA
transforms interrelated variables into a new component called the principal compo-
nents. Principal components have uncorrelated, orthogonal, and ordered properties
and still have most of the variation present in the actual dataset.
Standardization, covariance matrix computation, eigenvectors, and eigenvalues
computations are PCA computation steps. Standardization is achieved with the
following equation.

x −u
z= (1)
σ
After data standardization, the covariance matrix computations are given by the
following formula.

1 n
cov(X, Y ) = (X i − x)(Y i − y) (2)
n−1 i=1

With the help of the above formula, the covariance matrix of A is calculated. The
eigenvector of a matrix A is the vector for which has stratified the following formula:
 
Av − λv = 0 (3)

Lambda is a scalar value called the eigenvalue. The eigenvectors with the largest
eigenvalues are the principal components. The principal components are uncorrelated
of each other.
158 F. Gökgöz and F. Filiz

3.2 Linear Regression and Polynomial Regression

Statistical methods are based on the utilization of mathematical statistics and

stochastic processes. The statistical model’s ability to process non-linear data is
insufficient. Linear regression is a conventional statistical method. Multiple linear
regression aims to calculate the linear relationship between independent variables and
dependent variables with the help of a linear equation. Linear regression forecasting
capability is compared with other deep learning methods for energy forecasting (Liu
et al., 2017). The multiple linear regression is calculated using the following formula.

y = a + b0 ∗ x0 + b1 ∗ x1 + · · · + bn ∗ xn (4)

Polynomial regression is a kind of regression method. Polynomial regression tries

to find a relationship between the dependent and independent variable with nth degree
polynomial. The polynomial regression is calculated using the following formula.

y = a + b1 ∗ x1 + b2 ∗ x12 + · · · + bn ∗ x1n (5)

3.3 Long Short-Term Memory (LSTM)

The traditional ANN consist of the input layer, hidden layer, and output layer.
Deep neural networks have many layers and complex learning algorithms. Deep
learning network methods have powerful nonlinear mapping capability. Feedfor-
ward artificial neural networks do not entail memorizing characteristics, but recurrent
neural networks can memorize. LSTM and GRU are classified under deep learning
methodologies which are special kinds of recurrent neural networks.
LSTM block uses forget gate, input gate, and output gate to neurons. Figure 1
shows LSTM blocks.
Unlike recurrent neural networks, LSTM can prevent the long-term dependency
problem. LSTM has the ability to learn from long and short-term input data. This
feature prevents the negative effects of long and short-term data variables on the
forecast results and positively affects the prediction ability. The forget gate, input
gate, and output gate formulas in the LSTM block are given by (6)–(8). “i” stands for
input gate, f stands for forgetting gate and o stands for output gates. c is dedicated
to the memory cell and c̃ is dedicated to the new memory cell content.

f t = σ (W f x xt + W f h h t−1 + b f ) (6)

i t = σ (Wi x xt + Wi h h t−1 + bi ) (7)

Comparing the Renewable Energy Technologies … 159

Fig. 1 LSTM block

ot = σ (Wox xt + Woh h t−1 + bo ) (8)

W f x W f h , Wi x ,Wi h , Wox , and Woh are weight matrices for the inputs of the network
activation functions; σ represents the sigmoid activation function. h t−1 is the hidden
state of the previous time step. b f , bi , and bo are biases.

3.4 Gated Recurrent Unit (GRU)

GRU is also another extended model of RNN. GRU is different from the LSTM
according to utilizing gating information to prevent vanishing gradient problems.
LSTM can overcome the problem of recurrent neural networks that cannot handle
long-distance dependence. GRU has a simpler and more efficient structure compared
to LSTM (Ke et al., 2019).
Jozefowicz found that GRU is easier to train and has comparable results with
LSTM models (Jozefowicz et al., 2015). Zhou et al. argued that GRU had stable
performance and with fewer parameters (Zhou et al., 2016). Figure 2 shows reset
and update gates, activation, and candidate activation. r and z represent the reset and
update gates, and h and h̃ denote the activation and the candidate activation.
A reset gate controls how much of the previous state to remember and helps
capture short-term dependencies. An update gate control how much of a copy of the
old state to the new state and helps capture long-term dependencies in sequences.
r shows the reset and z shows the update gates. The reset gate and update gate
formulas in the GRU block are given by (9)–(10).
160 F. Gökgöz and F. Filiz

Fig. 2 GRU block

rt = σ (Wr x xt + Wr h h t−1 + br ) (9)

z t = σ (Wzx xt + Wzh h t−1 + bz ) (10)

Wr x Wr h ,Wzx ,Wzh are weight matrices for the inputs of the network activation func-
tions; σ shows the sigmoid activation function, h t−1 is the hidden state of the previous
time step, br and bz are biased. h and h̃ are dedicated to the activation and the candidate
activation functions.

4 Empirical Study

4.1 Dataset and Data Preprocessing

In this study, a dataset of wind and hydro energy in Turkey for the year 2019 and
2020 are used. Corresponding meteorology data is also available. The wind and hydro
generation dataset, each contains 24,096 samples recorded every 1 h-resolution.
Figure 3 shows electricity generation from 2016 to 2019. There is a slight increase
in wind energy between 2016 and 2019. Comparing to wind energy, there is a
significant increase in hydro energy between these years.
Figure 4 shows the installed capacity of Turkey from 2016 to 2019 according to
Turkey Energy Market Regulatory Authority’s report. Likewise, installed capacity
and electricity generation are constantly rising between 2016 and 2019. Hydro
Comparing the Renewable Energy Technologies … 161

Fig. 3 Electricity generation of Turkey

Fig. 4 Installed capacity of Turkey

energy has the largest share between 2016 and 2019 (Electricity Market Development
Report, 2019).
Temperature, wind speed, cloud cover, humidity, dew point values were used. PCA
applies to compute the principal components and use them to perform forecasting.
During the preprocessing phase, input variables are normalized to improve the model
performance of deep learning and regression models.
An effective model should be built on high-quality input data. Renewable energy is
directly affected by atmospheric conditions due to its natural characteristic. During
the process of collecting and measuring atmospheric data, there may be drops or
inconsistencies in the measurement quality of atmospheric data. Model inputs have
a direct effect on the predictive results. With data cleaning of abnormal data, deter-
mining the relationships between input data should be done in a preprocessing state.
Renewable energy generation is affected by many factors, such as wind speed, global
warming, temperature, drought, precipitation, evaporation, humidity, radiation, and
climatic parameters (Mun & Sailor, 1998). Data is sometimes missing or seriously
high or low because of measurement errors. After detecting anomalous samples and
preprocessing them positively affects the learning performance of the models.
162 F. Gökgöz and F. Filiz

Fig. 5 Wind energy (KWh) and PCA components

After applying PCA to meteorology data, four principal components are created.
These principal components are used as input to the deep learning and regression
models. Principal component analysis and multiple linear regression can be used
together (Wuttichaikitcharoen and Babel, 2014).
Figures 5 and 6 shows that wind energy generation and hydro energy generation.
Figure 5 shows wind energy generation data with one-hour resolution and four
principal components input data. Wind energy generation is more volatile than hydro
generation.
Figure 6 shows hydro energy forecasting input data with 1-h resolution and four
principal components input data. Hydro energy generation has a periodic pattern.
Temperature, wind speed, cloud cover, humidity, dew point values produce PCA_1,
PCA_2, PCA_3, and PCA_4.
In this study, the data is divided into two parts: a training part and a test part. The
training part is used to train the forecasting model. The test part is applied to evaluate
forecasting performance. A total of 90% of the samples are used to train the power
forecasting model, while the remaining of the sample data are used to test the model.
Comparing the Renewable Energy Technologies … 163

Fig. 6 Hydro energy (KWh) and PCA components

4.2 Empirical Methods

4.2.1 Multiple Linear Regression

MAE and MSE are used to measure the forecasting performance of the models. The
MAE and MSE are defined as follows.
1 n 

M AE = |X i − X i | (11)
n i=1

1 n 
2
MSE = (X i − X i ) (12)
n i=1

MAE and MSE values are 0.86 and 0.95, respectively, in wind energy forecasting.
Figure 7 shows the actual data and prediction data of the wind energy forecasting
model. Wind energy generation volatility affects models forecasting performance
negatively.
Figure 8 shows actual data and prediction data of the hydro energy forecasting
model. Hydro energy test data tends to decrease. It can be seen that the forecast data
cannot catch this trend. MAE value for hydro energy forecasting is 0.61 and MSE
value is 0.49.
164 F. Gökgöz and F. Filiz

Fig. 7 Wind energy forecasting with multiple linear regression

Fig. 8 Hydro energy forecasting with multiple linear regression

4.2.2 Polynomial Regression

Figure 9 shows actual and prediction data for the wind energy polynomial regression
model. Wind energy actual data trend is achieved but all polynomial models have
prediction spikes. MAE value for wind energy forecasting is 0.78 and the MSE value
is 0.91.
Comparing the Renewable Energy Technologies … 165

Fig. 9 Wind energy forecasting with polynomial regression

Multiple linear regression and polynomial linear regression forecasting perfor-

mance relatively the same. The polynomial regression model for hydro energy also
has prediction spikes. Figure 10 shows actual and prediction values for the model.
MAE and MSE values are 0.59, 0.57, respectively.
In wind energy and hydro energy forecasting, the occurrence of sudden increases
in polynomial regression estimation values creates the weak point of the model.

Fig. 10 Hydro energy forecasting with polynomial regression

166 F. Gökgöz and F. Filiz

Fig. 11 Wind energy forecasting with LSTM

4.2.3 LSTM

In wind energy forecasting with LSTM, the MAE value and the MSE value are 0.60
and 0.57, respectively. Figure 11 shows the actual and prediction data of the LSTM
model. The LSTM model seems to catch the actual data trend much better than the
multiple linear regression and polynomial regression.
In hydro energy forecasting, the LSTM model has caught the trend of actual
data. Figure 12 illustrates actual and prediction values for hydro energy forecasting
models. MAE value is 0.42 and MSE value is 0.35.
Since the LSTM model has memory cells, we do not see sudden changes in
predicted values. Also, the LSTM model can alleviate vanishing and exploding
gradients. This makes LSTM suitable for prediction.

4.2.4 GRU

GRU performance and catching the actual data trend success are similar to LSTM.
Although volatility in wind energy made structure prediction difficult, LSTM versus
GRU were able to catch these trends. The production trend that slows down after
a rapid decrease in hydro energy has been also caught. MAE value in wind energy
forecasting is 0.63 and MSE value is 0.61. Figure 13 shows wind energy forecasting
and actual results.
In hydro energy forecasting, the MAE value is 0.58 and the MSE value is 0.57.
Figure 14 shows hydro energy forecasting and actual results.
The GRU model is as successful as the LSTM model in catching the trends of
real data. The GRU model advantage is that the training time of the model is short,
Comparing the Renewable Energy Technologies … 167

Fig. 12 Hydro energy forecasting with LSTM

Fig. 13 Wind energy forecasting with GRU

although the GRU model and the LSTM model achievements look approximately
the same. Because of the relatively simpler structure of the GRU block, it is easier
to train the GRU model.
168 F. Gökgöz and F. Filiz

Fig. 14 Hydro energy with GRU

4.2.5 Comparison of Empirical Discussion

MAE and MSE values are used as metrics in wind energy and hydro energy study.
Table 1 shows that multiple linear regression and polynomial regression result in
high MSE values.
According to the comparison results, LSTM and GRU offer better forecasting. The
improved forecasting error can be a helpful tool for power quality. In this study, we
have used Turkey’s hydro and wind energy data between 2019 and 2020. LSTM and
GRU models are the most successful models in wind and hydro energy forecasting
with respect to the data that used in this study.
Comparing to GRU models, LSTM has performed slightly better forecasting
performance. In hydro energy forecasting, all forecasting models have more
successful results. The more volatile structure of wind energy causes this situation. It

Table 1 MAE and MSE values

Forecasting Wind energy Wind energy Hydro energy Hydro energy
methods MAE value MSE value MAE value MSE value
Linear 0.86 0.95 0.61 0.49
regression
Polynomial 0.78 0.91 0.59 0.57
regression
LSTM 0.60 0.57 0.42 0.35
GRU 0.63 0.61 0.58 0.57
Comparing the Renewable Energy Technologies … 169

was observed that the variability in the data had a negative effect on the performance
of the estimation methods.

5 Conclusion

The comparative models in this study can help to decide which model has better
forecasting error. This study fills the existing gaps to compare deep learning and
regression models applied for renewable energy prediction. Compared to the deep
learning method, the linear regression model does not show the same success in
determining nonlinear relationships.
Wind energy is more sensitive to meteoritical changes, these affect the forecasting
performance of models. Hydro energy forecasting methods have better forecasting
performance in all applied methods.
This study compared the forecasting performances of methods in wind energy
and hydro energy forecasting with including multiple linear regression, polynomial
regression, LSTM, GRU. The historical wind energy, hydro energy, and weather
data of Turkey are used. MAE and MSE are selected to evaluate the forecasting
performance of different methods. The results indicate that LSTM provides better
forecasting performances than multiple linear regression, polynomial regression, and
GRU. Deep learning methods solve complex nonlinear problems successively. The
forecasting error results prove the effectiveness of the deep learning models. The
training computing cost of the deep learning models is acceptable.
Accurate prediction of renewable energy is an important part of realizing power
market quality. Turkey should use its renewable energy sources effectively with the
help of forecasting methods. To achieve effective usage of renewable energy, deep
learning models should be constantly developed and tested.

References

Akpinar, A., Kömürcu, M. I., Özölçer, I. H., & Şenol, A. (2011). Total electricity and hydroelectric
energy generation in Turkey: Projection and comparison. Energy Sources, Part B: Economics,
Planning and Policy, 6(3), 252–262.
Ardizzon, G., Cavazzini, G., & Pavesi, G. (2014). A new generation of small hydro and pumped-
hydro power plants: Advances and future challenges. Renewable and Sustainable Energy Reviews,
31, 746–761. https://doi.org/10.1016/j.rser.2013.12.043
Cao, Y., & Liai, G. (2019). Multi-step wind power forecasting model using LSTM networks, similar
time series and LightGBM. In 2018 5th International Conference on Systems and Informatics,
ICSAI 2018 (Icsai) (pp. 192–197).
Chen, S., Liu, P., & Li, Z. (2020). Low carbon transition pathway of power sector with high
penetration of renewable energy. Renewable and Sustainable Energy Reviews, 130, 109985.
https://doi.org/10.1016/j.rser.2020.109985
Coulibaly, P., Anctil, F., & Bobée, B. (2000). Neural network-based long-term hydropower
forecasting system. Computer-Aided Civil and Infrastructure Engineering, 15(5), 355–364.
170 F. Gökgöz and F. Filiz

Durak, M., & Şen, Z. (2002). Wind power potential in Turkey and Akhisar case study. Renewable
Energy, 25(3), 463–472.
Electricity Market Development Report. (2019). Turkey. https://www.epdk.gov.tr/Detay/Icerik/3-
0-24-3/elektrikyillik-sektor-raporu
Fridleifsson, I. B. (2003). Status of geothermal energy amongst the world’s energy sources.
Geothermics, 32(4), 379–388.
Gulagi, A., et al. (2020). Current energy policies and possible transition scenarios adopting renew-
able energy: A case study for Bangladesh. Renewable Energy, 155, 899–920. https://doi.org/10.
1016/j.renene.2020.03.119
Jozefowicz, R., Zaremba, W., & Sutskever, I. (2015). An empirical exploration of recurrent network
architectures. In Lille, France: MLR.org Proceedings of the 32nd International Conference on
International Conference on Machine Learning (Vol. 37).
Ke, K., Hongbin, S., Chengkang, Z., & Brown, C. (2019). Short-Term electrical load forecasting
method based on stacked auto-encoding and GRU neural network. Evolutionary Intelligence,
12(3), 385–394. https://doi.org/10.1007/s12065-018-00196-0
Kholodnyi, V. A. (2014). Quantitative Energy Finance.
King, D. (2007). Short term load forecasting using multiple linear regression, pp. 1192–1198.
Li, G., Li, B. J., Yu, X. G., & Cheng, C. T. (2015). Echo state network with Bayesian regularization
for forecasting short-term power production of small hydropower plants. Energies, 8(10), 12228–
12241.
Li, G. D., Masuda, S., & Nagai, M. (2016). Prediction of hydroelectric power generation in Japan.
Energy Sources, Part B: Economics, Planning and Policy, 11(3), 288–294. https://doi.org/10.
1080/15567249.2012.708097
Li, L, Yao, F., Huang, Y., & Zhou, F. (2019). Hydropower generation forecasting via deep neural
network. In Proceedings—2019 6th International Conference on Information Science and Control
Engineering, ICISCE 2019 (pp. 324–328).
Liu, B., Fu, C., Bielefield, A., & Liu, Y. Q. (2017). Forecasting of Chinese primary energy
consumption in 2021 with GRU artificial neural network. Energies, 10(10).
Ma, J., Oppong, A., Acheampong, K. N., & Abruquah, L. A. (2018). Forecasting renewable energy
consumption under zero assumptions. Sustainability (Switzerland), 10(3), 1–17.
Mata, J. (2011). Interpretation of concrete dam behaviour with artificial neural network and multiple
linear regression models. Engineering Structures, 33(3), 903–910. https://doi.org/10.1016/j.eng
struct.2010.12.011
Mun, J. R., & Sailor, D. J. (1998). A modelling methodology for assessing the impact of climate
variability and climatic change on hydroelectric generation. Energy Conversion and Management,
39(14), 1459–1469.
Niu, W. J., et al. (2019). Comparison of multiple linear regression, artificial neural network, extreme
learning machine, and support vector machine in deriving operation rule of hydropower reservoir.
Water (Switzerland), 11(1).
Niu, Z., et al. (2020). Wind power forecasting using attention-based gated recurrent unit network.
Energy, 196, 117081. https://doi.org/10.1016/j.energy.2020.117081
Stathopoulos, C., Kaperoni, A. Galanis, G., & Kallos, G. (2013). Journal of Wind Engineering Wind
Power Prediction Based on Numerical and Statistical Models. Journal of Wind Engineering and
Industrial Aerodynamics, 112, 25–38. https://doi.org/10.1016/j.jweia.2012.09.004
Suganthi, L., & Samuel, A. A. (2012). Energy models for demand forecasting—A review. Renewable
and Sustainable Energy Reviews, 16(2), 1223–1240. https://doi.org/10.1016/j.rser.2011.08.014
Torkzadeh, R., Mirzaei, A., Mirjalili, M. M., Anaraki, A. S., Sehhati, M. R., & Behdad, F. (2014,
May). Medium term load forecasting in distribution systems based on multi linear regression &
principal component analysis: A novel approach. In: 2014 19th Conference on Electrical Power
Distribution Networks (EPDC) (pp. 66-70). IEEE.
Wu, W., Chen, K., Qiao, Y., & Lu, Z. (2016). Probabilistic short-term wind power forecasting based
on deep neural networks. In 2016 International Conference on Probabilistic Methods Applied to
Power Systems, PMAPS 2016—Proceedings (November 2018).
Comparing the Renewable Energy Technologies … 171

Wuttichaikitcharoen, P., & Babel, M. S. (2014). Principal component and multiple regression anal-
yses for the estimation of suspended sediment yield in ungauged basins of Northern Thailand.
2412–2435.
Zhou, G.-B., Wu, J., Zhang, C.-L., & Zhou, Z.-H. (2016). Minimal gated unit for recurrent neural
networks. 13, 226–234.
Zu, X. R., & Song, R. X. (2018). Short-term wind power prediction method based on wavelet packet
decomposition and improved GRU. Journal of Physics: Conference Series, 1087(2).
Business Cycles and Energy Real
Options Valuation

Turalay Kenc and Mehmet Fatih Ekinci

1 Introduction

Energy projects are mostly large, irreversible and highly risky investments. The real
options valuation approach is widely used to value such investment projects. Indeed,
papers covered in the survey article by Fernandes et al. (2011) underscore the rele-
vance of the real options approach to value energy sector investments. However,
the related literature overlooks the distributional nature of project cash flows. The
work on energy real options overwhelmingly considers log-normally-distributed
cash flows, despite the well-documented evidence that cash flows are normally-
distributed (Burg, 2018; Kanniainen, 2009; Veronesi, 1999). The log-normality cash
flows assumption solves the tractability of the problem at expense of ruling out nega-
tive cash flows. In other words, the popularity of this assumption stems from its
tractability rather than its realism. Given the large and irreversible nature of energy
projects, distributional assumptions characterising cash flows are crucially impor-
tant in these investments. Furthermore, the interaction of normally distributed cash
flows with macroeconomic risk associated with business cycles can be different than
that of log-normal cash flows. This paper therefore uses a real options approach to
value energy projects whose cash flows follow a normal distribution and subject to
macroeconomic risk.
Volatile energy prices, supply and demand conditions in energy markets create
an environment in which negative cash flows can not ruled out. After the Covid-19
outbreak, most economies suffered from the resulting great lock-down which led to
the collapse of energy prices. Oil producers across the world faced an episode of
negative cash flows. According to the IMF World Economic Outlook, October 2020:

T. Kenc (B)
TOBB ETU, Ankara, Turkey
M. F. Ekinci
Atilim University, Ankara, Turkey
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 173
A. B. Dorsman et al. (eds.), Applied Operations Research and Financial Modelling
in Energy, https://doi.org/10.1007/978-3-030-84981-8_9
174 T. Kenc and M. F. Ekinci

“protracted low oil prices led to shut-ins, sharply reduced drilling activity, and a surge
in shale gas producer bankruptcy filings.” Similar outcomes did take place before the
Covid-19 pandemic. The recent ones include periods in the aftermath of the Great
Financial Crisis of 2008–9 and the plummeted oil prices in 2015. These episodes
clearly show that incorporating negative cash flows as well as macroeconomic risks
is essential to value energy investment projects.
In modelling energy investments, researchers mostly use continuous-time frame-
works and work with prices or values instead of cash flows. Hence, it is assumed
that prices do not take negative values. In mathematical terms, this amounts to
the geometric Brownian motion (GBM) representation of prices, i.e., prices follow
lognormal distribution. In this framework, replacing prices with cash flows does not
make any changes. However, letting cash flows take negative values requires an arith-
metic Brownian motion (ABM) specification. In addition to the randomness modelled
as Brownian motion processes, we incorporate another uncertainty which is related
to changes in macroeconomic conditions to characterize cash flows. The macroeco-
nomic risk often surfaces as economies follow business cycles and other fluctuations.
This uncertainty or macroeconomic risk is modelled as a regime-switching process.1
This paper is related to several strands of the literature. The first strand is the
real options theory and applications in energy economics and finance. Since the mid-
1980s,2 academics have widely applied the real options approach to value investments
with a greater degree of irreversibility, uncertainty and flexibility.3 Energy projects
are good examples of such investments as discussed above. From the beginning,
the approach has found a great deal of popularity in energy investments than in
other areas (see Fernandes et al., 2011). It is because these investments carry a
higher degree of uncertainty; specificity of investment projects makes them more
irreversible; and projects contain several flexibilities over their timing of investing,
scale and abandonment.
Secondly, it is the regime-switching modelling literature. In several fields of
economics and finance, many papers have considered real options with regime shifts
in key variables. In the area of irreversible investment, works by Hassett and Metcalf
(1999), Guo 2001, Guo and Zhang 2004, Guo et al. 2005, Elliott Miao, and Yu (2009,
Driffill et al. (2013) have applied regime-switching frameworks to value investment
projects. They allowed project values or cash flows to incorporate the two sources of
uncertainty. The first source stems from the usual continuous randomness, and the
second is the jump due to shifting regimes. Papers with regime-switching features

1 The normal distribution assumption is further useful in other energy-related valuations. For
example, the valuation of spread options on energy instruments and mergers of energy firms can
be mathematically relatively easy under this assumption as it is likely to obtain their closed-form
solutions. It is a well-known problem in statistics that only the convolution of normal distributions
gives the same normal distribution (see Leland, 2007).
2 In fact, almost the very first real options paper by Brennan and Schwartz (1985) was on the

valuation of natural resource investments.

3 The book by Schwartz and Trigeorgis (2001) includes the key articles till 2001 and the remaining

ones can be found in Lambrecht (2017), Kenc and Driver (2020).

Business Cycles and Energy Real Options Valuation 175

have appeared in corporate finance4 and Chen et al. (2018). Recent work also decom-
posed the risk premia into its factor and macro-economic (regime-switching) compo-
nents (see Driffill et al., 2013). Researchers and practitioners model stock prices,
dividends, interest rates, exchange rates and commodity prices increasingly using
regime-switching approaches.
The third introduces normal distribution to characterize cash flows of the under-
lying assets. In technical terms, this amounts to using ABM to model dynamics of
the underlying assets. It is a departure from the work-horse model of finance, GBM,
whose popularity usually stems from its tractability rather than its realism. This last
one has borders with spread options, basket options and mergers. Recently several
papers have used ABM to model dynamics of the underlying assets. They include
works by DéCamps et al. (2011), Genser (2006), Leland (2007), Veronesi (1999),
Weber et al. (2011), Burg (2018).
In this chapter, we follow the cash-flow tradition of Veronesi (1999, Goldstein,
Nengjiu Ju, and Hayne Leland (2001), Guo et al. (2005) and accordingly model
the cash flows generated by energy projects. The model will assume that cash-flows
follow a regime-switching diffusion process. More precisely, it is a regime-switching
ABM process which leads to both positive and negative cash flows. We adopt the risk-
neutral pricing approach to value energy investment projects. The risk-neutralisation
applies to both continuously occurring small random risks and infrequent large jumps
in cash flows.
In the next section, we will start with the traditional GBM case and obtain the
mathematical derivation of a valuation model with regime shifts. Section 3 will extend
the model in Sect. 2 to the ABM case. In Sect. 4, we will provide some numerical
analysis, including comparisons between the single regime and multiple regimes as
well as analysing the GBM and ABM cases. Section 5 provides concluding remarks.

2 The Basic Real Option Valuation with Regime Shifts

This section explains a simple real option investment model with regime shifts by
following the work of Driffill et al. (2013). They consider a perpetual real option
framework to analyse investment decisions on new projects under macroeconomic
risks as well as business risk. So, the randomness in the model is not simply governed
by a diffusion process with constant drift and volatility but instead characterized by an
additional source of uncertainty that the dynamics of the state variable shifts between
different regimes at random times. In other words, there are two types of shocks in
the economy: small but continuous shocks that are generated by a Brownian motion
Z(t) and large but infrequent shocks that are generated by a marked point process,
more precisely multi-state Markov chain process st . To simplify the explanation of

4 see Bhamra, Kuehn, and Strebulaev (2010), Chen (2010); Arnold, Wagner, and Westermann
(2013), Chen and Manso (2017).
176 T. Kenc and M. F. Ekinci

the model, at the beginning we will skip the risk-neutralisation aspects of their model.
So, the parameters of the model are risk-neutral ones.
A simple case involves only two regimes which can be characterised by a boom
and recession states: s ∈ {B, R}. The probability of the economy switches from
state B to state R within a small period t approximately equal to hB t while the
probability of switching from regime R to B is approximately hR t. The long-run
probability of the economy being in state B is hR /(hB + hR ).
The Continuous-Time Markov Chain Specification. The Markov chain st has
n states with a finite state space S := {e1 , e2 , . . . , en } which is identified by a finite
set of unit vectors ei = (0, · · · , 1, . . . , 0)T and follows the process:

dst = Hst dt + dNt (1)

where H is the constant intensity matrix and dN a Poisson process.

The transition probability over any interval of time τ can be constructed from the
intensity matrix H via the exponential formula:

P(τ ) = e−τ H

where element hi of the matrix H denotes the rate of leaving state i. We can obtain
the probability that a transition occurs from state i to state j in a small time interval
(t, t + t) is equal to hi t. Similarly, 1 − hi t is the probability that the process
remains in state i. We assume that there are two regimes in the economy: boom and
recession and hence i = B, R.
The transition matrix between t and t + t is then
 
1 − hB t hB t
Pt =
hR t 1 − hR t

Furthermore, we can write that dPt = HPt dt with

 
−hB hB
H= (2)
hR −hR

Solving the stochastic differential equation (SDE) in Eq. (1) yields:

  t 
st = eHt s0 + e−Hu dNu , (3)
0

So E[st ] = eHt s0 (4)

Elliott et al. (2009) proves that

Business Cycles and Energy Real Options Valuation 177

 
1 hR hR e−ht
e Ht
= + H
h hB hB h

where h = hB + hR .
Technology. Consider an investor evaluates a project whose cash flows follow
a regime switching diffusion process with the properties described above. Changes
in macro economic conditions govern large changes in cash flows while continuous
randomness in cash flows is generated by a Weiner process. The dynamics of cash
flows is given by:

dX (t) = X (t)[μst dt + σst dZ(t) (5)

where Z(t) is standard Brownian motion, st a Markov process independent of Z(t),

and μst and σst are regime-specific drift and volatility parameters: The pair takes
different values when the process st is indifferent states. In addition, the risk-free
f f
rate r f also takes regime specific values: rB and rR . The regime specific values of μ,
σ and r are known. With this feature of the model we loose the Markovian property
f

of X (t). Guo et al. (2005) points out that st , X (t) is jointly Markovian if at any
time t the state of st is known.
We first deal with the present value of this project, V (X , st ). The fundamental
pricing equation with regime shifts in μ and σ takes the form of

1
Xdt + Vx dX + Vxx (dX )2 + V, dst  = rst V (X , st )dt
2
Note that here we assume that the processes for both dX and ds are risk-
neutral. Plugging the equations for dX and ds into the above formula and taking
the expectation of the resulting expression yields

1
X + μst Vx + σs2t Vxx + V, Hst  = rst V (X , st )
2
This gives the following two equations when st = B or st = R:

X + μB VxB + σB2 21 VxxB + hB (V R − V B ) = rB V B

X + μR VxR + σR2 21 VxxR + hR (V B − V R ) = rR V R

As stated above, the parameters μB , μR , hB and hR are all risk-neutral.5 Following

Bhamra et al. (2010), Chen (2010), Driffill et al. (2013), we provide the project value
in the following proposition.
Proposition 1 (Project Valuation) The fundamental value of the project in each
regime, which rules out speculative bubbles, can be obtained as

5 We will show how we converted them from real to risk-neutral below.

178 T. Kenc and M. F. Ekinci

Vi (t) = κi X (t), i = B or i = R, (6)

where κi being the price-earnings ratio is given by

(rj − μj ) − (ri − μi )
κi−1 = ri − μi + pfj (7)
rj − μj + p

The first term on the right-hand side ri − μi represents the single regime case
while the second one is the adjustment term for regime switching. The former refers
to the case that the economy stayed in regime i forever, and the last term accounts
for future time spent in regime j. Here, p = hB + hR is defined as the risk-neutral
rate of news arrival, and (fB , fR ) = (hR /p, hB /p) represents the long-run risk-neutral
distribution. The assumption μB > μR and σB < σR implies that the price–earnings
ratio is higher in boom than in recession, i.e., κB > κR .

2.1 Valuing the Option to Invest

In this section we consider regime switching and optimal investment timing in a real
option framework in the spirit of McDonald and Siegel (1986). We find the value
of the option to invest, F, and the critical profit stream, X ∗ , at which it is optimal
to pay a sunk cost, I , for a project, V . The investment opportunity is assumed to
last forever and therefore it is equivalent to a perpetual call option—the right but
not the obligation to buy a share of stock at a prespecified price. Since the profit
process follows the two-regime-switching process defined in Eq. (5), there exists
a different trigger threshold, Xi∗ , for each regime i. That is, there are two optimal
critical values, XB∗ and XR∗ , with XB∗ < XR∗ , (as in Driffill et al. (2013), Hassett and
Metcalf (1999), Guo et al. (2005) and thus there are three relevant regions in the
profit space: In Region 1 (X ≤ XB∗ ), there is no investment regardless of the value of
X until the profits stream reaches XB∗ , where, provided that the economy is in state
B, investment takes place. In Region 2 (XB∗ ≤ X ≤ XR∗ ), there is investment if the
economy is in regime B, but there is no investment if the economy is in regime R. In
Region 3 (XB∗ < XR∗ ≤ X ) investment project is acceptable regardless of the state of
the economy.6
The investment problem in each region is defined below, which shows the equa-
tions that characterize the investment opportunities, F, across regions. In Region 1,
it is the case that (X ≤ XB∗ ), hence we have the following second order differential
equation in regimes B and R:

6 Because of the corresponding investment action, the sets (0, XB∗ ) and (XR∗ , ∞) are often referred
as to the inaction region and the action region, respectively. The set [XB∗ , XR∗ ] is called transient
region (Guo, 2001).
Business Cycles and Energy Real Options Valuation 179

1   f
μB XFxB + σB2 XB2 Fxx
B
+ hB F R − F B − rB F B = 0, (8)
2
1   f
μR XFxR + σR2 XR2 Fxx
R
+ hR F B − F R − rR F R = 0, (9)
2

where FXi and Fxx i

denote, respectively, the first and second partial derivatives of F
with respect to X in state i. F i = F(X (t), st = i) is the value of the function in state
i, and so on.
In Region 2, it is XB∗ ≤ X ≤ XR∗ and hence we have the second order differential

FB ≤ V B − I, (10)

1   f
μR XFxR + σR2 XR2 Fxx
R
+ hR F B − F R − rR F R = 0. (11)
2
Clearly, when XB∗ ≤ X the investor go ahead with project investment and hence
we have F B ≤ V B − I .
In Region 3, it is (XB∗ < XR∗ ≤ X ) and hence in both regimes, the investor exercises
her investment options

FB ≤ V B − I, (12)

FB ≤ V R − I. (13)

This system of equations can be solved subject to eight (boundary) conditions,

namely.
1. zero value conditions

Fi (0) = 0, i = B, R;

2. value matching conditions

F i (Xi∗ ) = V i (Xi∗ ) − I , i = B, R;

3. smooth-pasting conditions7

F i (Xi∗ ) = Vi (Xi∗ ), i = B, R;

7Technically these conditions are meant to ensure that the slopes of the functions Fi and Vi − I are
equal at the investment thresholds XB∗ and XR∗ . Intuitively, they imply that the rate of returns to the
option and the project net of the investment cost are equal at the investment thresholds.
180 T. Kenc and M. F. Ekinci

4. continuity conditions8

lim F R (X ) = lim∗+ F R (X ),
X ↓XB∗− X ↑XB

 
lim∗− F R (X ) = lim∗+ F R (X ).
X ↓XB X ↑XB

Proposition 2 (Solution for Investment Problem in Each Region) Option valuation

solutions for F B and F R across regions are given by the expressions below.
In region 1, it is the case that (X ≤ XB∗ ), the solutions accross regimes are given
by

F B = A1 X γ3 + A2 X γ4 , (14)

F R = B1 X γ3 + B2 X γ4 , (15)

where γ3 and γ4 are distinct real roots of the quadric equations defined in Eq. (28); ξ1
and ξ2 are similarly distinct real roots of the quadratic equation defined in Eq. (35);
A1 , A2 , XB∗ , XR∗ , C1 , C2 are constants to be determined by the model‘s two value
matching conditions, two smooth-pasting conditions and two continuity conditions
as described below. B1 and B2 are determined as function of A1 and A2 .
In region 2, it is the case that (XB∗ < X ≤ XR∗ ), the solutions accross regimes are
given by

F B ≤ kB X − I , (16)

hR kB hR I
F R (X ) = C1 X ξ1 + C2 X ξ2 + f
X− f
. (17)
rR + hR − μR rR + hR

In region 3, it is the case that (XB∗ < XR∗ ≤ X ), the solutions accross regimes are
given by

F B ≤ kB X − I , (18)

F R ≤ kR X − I , (19)

8 They ensure the smoothness of the option value function at the boundary between the inaction
region and the transient region (see Driffill, Kenc, and Sola 2013), Guo, Miao, and Morellec (2005).
Business Cycles and Energy Real Options Valuation 181

2.1.1 Solution for Investment Problem in Region 1

Plugging the solutions in Eqs. (14) and (15) into Eqs. (8) and (9) gives us:
   
μB X γ3 A1 X γ3 −1 + γ4 A2 X γ4 −1 + 21 σB2 X γ3 (γ3 − 1)A1 X γ3 −2 + γ4 (γ4 − 1)A2 X γ4 −2
+hB (B1 X γ3 + B2 X γ4 − A1 X γ3 − A2 X γ4 ) − rB (A1 X γ3 + A2 X γ4 ) = 0
f

(20)
   
μR X γ3 B1 X γ3 −1 + γ4 B2 X γ4 −1 + 21 σR2 X 2 γ3 (γ3 − 1)B1 X γ3 −2 + γ4 (γ4 − 1)B2 X γ4 −2
(21)
+hR (A1 X γ3 + A2 X γ4 − B1 X γ3 − B2 X γ4 ) − rB (B1 X γ3 + B2 X γ4 ) = 0
f

Equating coefficient of X γ3 and X γ4 in Eqs. (20) and (21) we must have:

1 f
μB γ3 A1 + σB2 γ3 (γ3 − 1)A1 + hB (B1 − A1 ) − rB A1 = 0 (22)
2
1 f
μB γ4 A2 + σB2 γ4 (γ4 − 1)A2 + hB (B2 − A2 ) − rB A2 = 0 (23)
2
1 f
μR γ3 B1 + σR2 γ3 (γ3 − 1)B1 + hR (A1 − B1 ) − rB B1 = 0 (24)
2
1 f
μR γ4 B2 + σR2 γ4 (γ4 − 1)B2 + hR (A2 − B2 ) − rR B2 = 0 (25)
2
From Eq. (22)

B1 1 f
hB = μB γ3 + σB2 γ3 (γ3 − 1) − hB − rB (26)
A1 2

From Eq. (24)

hR AB11 = μR γ3 + 21 σR2 γ3 (γ3 − 1) − hR − r f (27)

Therefore, γ3 is a solution of the fourth order equation:

  
1 2 1 2
hB hR = μB γ + σB γ (γ − 1) − hB − r f
μR γ + σR γ (γ − 1) − hR − r f
2 2
(28)

Note that from Eqs. (23) and (25) we obtain the the same result as Eq. (28). The
real option must have a finite value at X = 0 so only the positive roots (0 < γ3 , γ4 )
of Eq. (28) are of interest. To relate B’s to A’s we use Eqs. (22) and (23)
 
B1 1 1 2
l1 = = hB + r − μB γ3 − σB γ3 (γ3 − 1)
f
(29)
A1 hB 2
182 T. Kenc and M. F. Ekinci

 
B2 1 1 2
l2 = = hB + r − μB γ4 − σB γ4 (γ4 − 1)
f
(30)
A2 hB 2

Given that l1 and l2 are known, then B1 = l1 A1 and B2 = l2 A2 .

2.1.2 Investment Problem in Region 2

In Region 2, since X > XB∗ the firm invests in regime B but keeps the real option
alive in regime R. The former case implies that

F B ≤ VB − I , (31)

the latter is given by

1 f
μR XFXR + σR2 X 2 Fxx
R
+ hR (F B − F R ) − rR F R = 0. (32)
2
Substituting Eq. (31) into Eq. (32) and simplifying yields

1 f
μR XFXR + σR2 X 2 Fxx
R
+ hR (VB − I − F R ) − rR F R = 0 (33)
2
Given that VB (X ) = kB X , we rewrite Eq. (33) as

1 f
μR XFXR + σR2 X 2 Fxx
R
+ hR (kB X − I − F R ) − rR F R = 0 (34)
2
The homogeneous part of Eq. (34),

1 f
μR XFXR + σR2 X 2 Fxx
R
− hR F R − rR F R = 0,
2

has a trial solution of the form

F h (X ) = C1 X ξ1 + C2 X ξ2 ,

where ξ1 < 0 and ξ2 > 0 satisfy the following quadratic equation:

σR2 f
μR ξ + ξ (ξ − 1) − hR − rR = 0. (35)
2
with
Business Cycles and Energy Real Options Valuation 183

 2 f
1 μR 1 μR 2(rR + hR )
ξ1 = − 2 + − 2 +
2 σR 2 σR σR2

and
 2 f
1 μR 1 μR 2(rR + hR )
ξ2 = − 2 − − 2 +
2 σR 2 σR σR2

Using the method of undetermined coefficients, we postulate that the forcing term
hR (kB X − I ) of the non-homogeneous part of Eq. (34) takes the following particular
solution

F nh (X ) = C3 X + C4 . (36)

To determine C3 and C4 in the particular solution, we differentiate the candidate

solution Eq. (36) and plug it into Eq. (34) to obtain

μR C3 X − (hR + r f )(C3 X + C4 ) − hR I + hR kB X = 0.

Applying the constant coefficients technique gives us

μR C3 − (hR + r f )C3 X = −hR kB ,

−(hR + r f )C4 = hR I .

The solutions of these equations are

C3 = (hR +r f )−μ ,
hR kB
R
C4 = − (hR +r f ) .
hR I

Therefore we finally obtain the solution for F R :

hR kB hR I
F R (X ) = C1 X ξ1 + C2 X ξ2 + X− .
(hR + r f ) − μR (hR + r f )

2.1.3 Investment Problem in Region 3

Finally, in Region 3, investment is made. The value functions are given by

F B ≤ VB − I ,
(37)
F R ≤ VR − I .
184 T. Kenc and M. F. Ekinci

2.2 Solution of the Constants and Critical Triggers

In order to solve the above equations for the constants Ai , Bi , Ci (i = 1, 2) and critical
triggers Xi∗ (i = B, R) we utilize the fact that the value of investment opportunity
F i (X ), (i = B, R) must also satisfy the following boundary conditions:
1. the Value-Matching conditions

F i (Xj∗ ) = V i (Xj∗ ) − I , i = B, R; j = B, R

2. the Smooth-Pasting conditions

F i (Xi∗ ) = V i (X ∗ ), i = B, R

3. the Continuity Value-Matching condition

lim F R (X ) = lim∗ F R (X ),
X ↓XB∗ X ↑XB

4. the Continuity Smooth-Pasting Condition

lim F R (X ) = lim∗ F R (X ).
X ↓XB∗ X ↑XB

More precisely, the value-matching conditions imply that

∗γ3 ∗γ4
A1 XB + A2 XB = kB XB∗ − I , (38)

∗ξ ∗ξ hR kB hR I
C1 XR 1 + C2 XR 2 +   X∗ −   = kR XR∗ − I , (39)
hR + r f − μR R hR + r f

using the smooth-pasting conditions, we obtain

∗(γ3 −1) ∗(γ4 −1)
γ3 A1 XB + γ4 A2 XB = kB , (40)

∗(ξ1 −1) ∗(ξ2 −1) hR kB

ξ1 C1 XR + ξ2 C2 XR +  = kB , (41)
hR + r f − μR

continuity value-matching and smooth-pasting conditions yield,

∗γ3 ∗γ4 ∗ξ ∗ξ hR kB hR I
B1 XB + B2 XB = C1 XB 1 + C2 XB 2 +   X∗ − , (42)
hR + r f − μR B f
rR + hR
Business Cycles and Energy Real Options Valuation 185

∗(γ3 −1) ∗(γ4 −1) hR kB

γ3 B1 XB + γ4 B2 XB = ξ1 C1 XB∗(ξ1 −1) + ξ2 C2 XB∗(ξ2 −1) + .
f
rR + hR − μR
(43)

This is a system of 6 equations—in 8 unknowns (XB∗ , XR∗ , A1 , A2 , B1 , B2 , C1 , C2 ).

We have already obtained expressions for (B1 , B2 ) in terms of A1 and A2 as B1 = l1 A1
and B2 = l2 A2 .
Using Eqs. (38) and (40) we first solve for A1 and A2 as follows:
   ∗γ ∗γ −1  
A1 XB 3 XB 4 kB XB∗ − I
= ∗(γ3 −1) ∗(γ −1) (44)
A2 γ3 XB γ4 XB 4 kB

Then from Eqs. (39) and (41) we solve for C1 and C2 :

   ∗ξ ∗ξ −1  
C1 XR 1 XR 2 kR XR∗ − I − C3 XR∗ − C4
= (45)
C2 ξ1 XR∗ξ1 ξ2 XR∗ξ2 kR − C3

After substituting A1 , A2 , C1 and C2 from Eqs. (44) and (45) into Eqs. (42) and (43)
we finally solve the resulting equations for XB∗ and XR∗ using a nonlinear equations
system solver.

2.3 The Single Regime Case

Letting hB = hR = 0, μB = μR and σB = σR leads to the single regime real option

valuation model. In the single regime world, a single critical threshold value of X ∗
creates two regions. When X (t) < X ∗ , the project operates in Region 1 and whose
investor has an option to invest. When X ∗ ≤ X (t), we are in Region 2 and the option
is already exercised. We obtain the relevant equations for the single regime case
as follows. In Region 1, i.e., X (t) < X ∗ , the investor values her option using the
following equation:

1
μXFx + σ 2 X 2 Fxx − r f F = 0, (46)
2
This equation is a special case of the regime switching real options valuation
Eqs. (8) and (9) and gives us the expression for γ :

 2
1 μ 1 μ2 2r f
γ = − 2+ − + .
2 σ 2 σ2 σ2
186 T. Kenc and M. F. Ekinci

Similarly, when X ∗ ≤ X (t), it is an inaction region and the value of the project
after the investment is a special case of the regime switching Eqs. (13) and (13):

F ≤ V − I, (47)

Solving the system of value-matching and smooth-pasting conditions yields

expressions for the critical cash flow value X ∗ and the option value F(0):
γ
X ∗ = [r f − μ] I
γ −1
 γ  
X0 X∗
F(0) = −I
X∗ rf − μ
∗
Note that the value matching condition is F(X ∗ ) = V (X ∗ ) − I = r fX−μ and the
associated smooth-pasting condition is γ AX ∗γ −1 = r f −μ
1
given that F(X ) = AX γ .

3 Valuing Regime Switching Real Options Under

Arithmetic Brownian Motion

In this section, we will depart from the GBM assumption in describing the dynamics
of cash flows and adopt an arithmetic Brownian motion (ABM). Equation (5) for
cash flows in GBM modifies to cash flows in ABM as follows:

dX (t) = μst dt + σst dZ(t). (48)

This section also describes the risk-neutralisation process what we skipped at the
beginning. To this end, we assume the following stochastic discount factor:

dM (st )
= −rsf t dt − λst dZt − [exp(κst ) − 1][dst − Hst dt],
M (st )

where λ is the risk price for Brownian shock affecting cash flows and κi is the relative
jump size of the discount factor when the Markov chain leaves state i implying that
exp(κsi ) − 1 is the price of switching risk from regime i to j. The model implies
thatκi = 1/κj 9 The risk neutralisation requires that μB = μPB − λB ρσB σM and
μR = μPR − λR ρσR σM for the parameters of Brownian motion processes. Here, σM
is the standard deviation for the market portfolio and ρ is the correlation coefficient
between market portfolio and energy project in hand. Given λi , ρ and σM one can
convert μPi under the physical probability measure to its risk-neutral counterpart μi

9 For a careful derivation of this result see Driffill, Kenc, and Sola (2013).
Business Cycles and Energy Real Options Valuation 187

in each regime i = B, R. We apply a similar conversion to the transition probability

terms: hi = eκi hPi for i = B, R.

Proposition 3 (Project Valuation under Regime Switching ABM) The fundamental

value of the project with regime switching cash flows in ABM in each regime, which
rules out speculative bubbles, can be obtained as
   −1    
VH r + hB −hB μB BB BB
= B + X
VR −hR rR + hR μR BR BR

In a more compact form, Bi can be written as

1
Bi = p
ri

p
where ri is the perpetual risk-free rate given by

p rj − ri ˜
ri = ri + p̃fj
p̃ + rj

in which p̃ = hB + hR .

1  
X + μB VxB + σB2 VxxB + hB V R − V B = rB V B (49)
2
1  
X + μR VxR + σR2 VxxR + hR V B − V R = rR V R (50)
2
To solve the system of equations above we postulate

V B = AB + B B X
(51)
V R = AR + BR X

Substituting the above equations into Eqs. (49) and (50), we get the following
equations

X + μB BB + hB [AR + BR X − AB − BB X ] = rB (AB + BB X )
X + μR BR + hB [AB + BB X − AR − BR X ] = rR (AR + BR X )

By separating the terms with X from the rest, we obtain:

X + hB (BR − BB )X = rB BB X μB BB + hB (AR − AB ) = rB AB
and ,
x + hR (BB − BR )X = rR BR X μR BR + hR (AB − AR ) = rR AR
188 T. Kenc and M. F. Ekinci

where
   −1      −1  
BB r + hB −hB 1 AB rB + hB −hB μB BB
= B = ,
BR −hR rR + hR 1 AR −hR rR + hR μR BR
  rR +hB +hR
BB
= rB rRrB+h B rR +rB hR
+hB +hR
BR rB rR +hB rR +rB hR

If there exists only a single interest rate irrespective of regimes, then the formulas
simplify to Bj = 1/r and to
  μB hB (μB −μR )
AB − r 2 (r+hB +hR )
= r2
μR hR (μR −μB ) .
AR r2
− r 2 (r+hB +hR )

3.1 Valuing the Option to Invest in Projects with Regime

Switching ABM Cash Flows

Since the problem is already described in the previous section, we only cover the
differences in solutions to the problems with GBM and ABM in this section. First,
we modify the GBM case solutions provided in Table ~ for the ABM case as follows.
Proposition 4 (Option Valuation with Regime Switching ABM) Solutions for F B and
F R across regions are as follows.
In Region 1, it is (X ≤ XB∗ ) and we have the following solutions:

F B = A1 eγ3 X + A2 eγ4 X ,
(52)
F R = B1 eγ3 X + B2 eγ4 X .

In Region 2, it is (XB∗ < X ≤ XR∗ ) and we have the following solutions:

F B ≤AB + BB X − I ,
hR kB hR I
F R =F R (X ) = C1 X ξ1 + C2 eξ2 X + f
X− f
. (53)
rR + hR − μR rR + hR

In Region 3, it is (XR∗ ≤ X ) and we have the following solutions:

F B ≤ AB + BB X − I ,
(54)
F R ≤ AR + BR X − I .

In the ABM case, the fourth-order equation solution for γ3 takes the following
solution,
Business Cycles and Energy Real Options Valuation 189
  
hB hR = μB γ + 21 σB2 γ 2 − hB − r f μR γ + 21 σR2 γ 2 − hR − r f , (55)

and expressions for l1 and l2 are given by:

 
l1 = B1
= 1
h
hB  B
+ r f − μB γ3 − 21 σB2 γ32
A1  (56)
l2 = B2
A2
= 1
hB
hB + r f − μB γ4 − 21 σB2 γ42

Given that l1 and l2 are known, then B1 = l1 A1 and B2 = l2 A2 .

In Region 2, the solutions for homogeneous and nonhomogeneous parts are of
the forms:

F h (X ) = C1 eξ1 X + C2 eξ2 X ,
(57)
F nh (X ) = C3 X + C4 .

where ξ1 < 0 and ξ2 > 0 satisfy the following quadratic equation

σR2 2 f
μR ξ + ξ − hR − rR = 0. (58)
2

with
 2 f  2 f
μR μR 2(rR + hR ) μR μR 2(rR + hR )
ξ1 = − + + and ξ2 = − 2 − + .
σR2 σR2 σR2 σR σR2 σR2

To determine C3 and C4 in the particular solution, we differentiate the candidate

solution Eq. (36) and plug it into Eq. (34) to obtain

f
μR C3 − (rR + hR )(C3 X + C4 ) − hR I + hR AB + hR BB X = 0.

Applying the constant coefficients technique gives us

f
−(rR + hR )C3 X = −hR BB X ,
f
μR C3 − (rR + hR )C4 = hR I − hR AB

The solutions of these equations are

C3 = hR BB
f ,
(rR +hR )
[μR C3 +hR (AB −I )]
C4 = f
(rR +hR )
μR hR BB
= f + hR Af B −I .
(rR +hR )2 (rR +hR )

Therefore we finally obtain the solution for F R :

190 T. Kenc and M. F. Ekinci

hR BB μR hR BB AB + I
F R (X ) = C1 X ξ1 + C2 X ξ2 + f
X+ f
− hR f
.
(rR + hR ) (rR + hR ) 2
(rR + hR )

In Region 3 when investment is made. The value functions are the same in the
GBM case. are given by

F B ≤ VB − I ,
(59)
F R ≤ VR − I .

To determine the constants Ai , Bi , Ci (for i = 1,2) we work with the value-matching

conditions
∗ ∗
A1 eγ3 XB + A2 eγ4 XB = AB + BB XB∗ − I , (60)

∗ξ ∗ξ hR BB μR hR BB AB + I
C1 XR 1 + C2 XR 2 + f
XR∗ + f
− hR f
= AR + BR XR∗ − I ,
(rR + hR ) (rR + hR ) 2
(rR + hR )
(61)

with the smooth-pasting conditions

∗ ∗
γ3 A1 eγ3 XB + γ4 A2 eγ4 XB = BB , (62)

∗ ∗
ξ1 C1 eξ1 XR + ξ2 C2 eξ2 XR + C3 = BR . (63)

Utilizing the continuity value-matching and smooth-pasting conditions, we write:

∗ ∗ ∗ ∗
B1 eγ3 XB + B2 eγ4 XB = C1 eξ1 XB + C2 eξ2 XB + C3 XB∗ + C4 , (64)

∗ ∗ ∗ ∗
γ3 B1 eγ3 XB + γ4 B2 eγ4 XB = ξ1 C1 eξ1 XB + ξ2 C2 eξ2 XB + C3 . (65)

From the above equations, we obtain a system of 6 Eqs. (13)–(18) in 8 unknowns

(XB∗ , XR∗ , A1 , A2 , B1 , B2 , C1 , C2 ). We have already obtained expressions for (B1 , B2 )
in terms of A1 and A2 as B1 = l1 A1 and B2 = l2 A2 .
Using Eqs. (60) and (62) we first solve for A1 and A2 as follows:
   ∗ ∗ −1  
A1 eγ3 XB eγ4 XB AB + BB XB∗ − I
= ∗ ∗ (66)
A2 γ3 eγ3 XB γ4 eγ4 XB BB

Then from Eqs. (61) and (63) we solve for C1 and C2 :

   ∗ ∗ −1  
C1 eξ1 XR eξ2 XR BR XR∗ − I − C3 XR∗ − C4
= ∗ ∗ (67)
C2 ξ1 eξ1 XR ξ2 eξ2 XR BR − C3
Business Cycles and Energy Real Options Valuation 191

After substituting A1 , A2 , C1 and C2 from Eqs. (66) and (67) into Eqs. (64) and (65)
we finally solve the resulting equations for XB∗ and XR∗ using a nonlinear equations
system solver.

3.2 The Single Regime Case

As in the GBM case, the single regime under the ABM assumption implies that
hB = hR = 0, μB = μR and σB = σR . The critical threshold value of X ∗ in the single
regime model divides the time into two regions.In Region 1 we are with the case
X (t) < X ∗ and has an option to invest while in Region 2, it is X ∗ ≤ X (t) and the
option is already exercised. The value of the real option is give by:

1
μFx + σ 2 Fxx − r f F = 0, (68)
2

which is the special case of the regime switching case’s Eqs. (49) and (50).
In Region 2, the value of the project after investment takes the form of

F ≤ V − I, (47)

The value-matching and smooth-pasting conditions gives us the critical cash flow
value X ∗ and the option value F = Aeγ X :

B+γ (I −A)
X∗ =
∗
γ (X −X )
 γB ∗

F =e A + BX − I
∗
Note that the value matching condition is eγ X = A + BX ∗ − I and the smooth-
∗
pasting condition is γ Aeγ X = B. The regime switching ABM model reveals that
μ
A = (r f )2 and B = 1/r (See Genser, 2006).
f

4 Numerical Analysis

In this section, we carry out a numerical analysis of the real options valuation models
based on both lognormally and normally distributed cash flows. We focus on the
regime specific cash-flow thresholds that trigger investment, the optimal time to
investing and the probability of investing within a certain period. To obtain their
values we run our models using the plausible values for the parameters based on
the literature. The GBM-based regime-switching papers by Arnold et al. (2013),
Driffill et al. (2013), Guo et al. (2005) guide us identifying these plausible values.
192 T. Kenc and M. F. Ekinci

Table 1 Parameters of the

Parameter Regime GBM ABM
GBM and ABM
f
regime-switching real options rB Boom 0.06 0.06
models f
rR Recession 0.06 0.06
μB Boom 0.04 0.43
μR Recession 0.015 0.34
σB Boom 0.06 0.69
σB Recession 0.15 1.51
hB Boom 0.16 0.16
hR Recession 0.52 0.52
I – 100 100
X0 – 5 5

The parameters needed in such frameworks include transition probability parameters

f f
hB and hB , the risk-free rates, rB and rR , diffusion process parameters μB , μR , σB ,
and σR . Besides, we need the cost of project investment I and initial cash flow X0
f f
of the energy project. The parameters rB rR , hB , hR , I , X0 are common for our both
models whereas μB , μR , σB , and σR are model-specific.
The intensity parameter hB denotes the rate of leaving state B and similarly hR is
the rate of leaving from regime R. Our chosen values hB = 0.16 and hR = 0.525
imply that the switch from the boom regime to the recession takes approximately
1/hB = 6.25 years on average while it is 1/hR = 1.9 years for a move from the
recession state to the boom state. In other words, the project operates for 6.25 years
in the boom regime while it spends 1.9 years for the recession state. Since the expected
fraction of time spent in regime i is hj /(hi +hj ), the chain stays in regime B 77% of the
time while it spends 23% in regime R.10 These values are in line with the observed
business cycle patterns. The interest rate is 6% at both regimes. Cash flow starts
with an initial value of $5 and I = $100. The GBM model’s diffusion parameters
(μi , σi , i = B, R) within the ranges of values used in Driffill et al. (2013), Guo et al.
(2005). Table 1 reports their values. We identify the values of the ABM model’s
(μi , σi , i = B, R) as follows.
Values for the parameters (μABMi , σiABM , i = B, R) are chosen such a way that
the GBM- and ABM-based real option models are comparable in terms of the
moments of observable variables. More specifically, the chosen values deliver us
comparable moments of project values after five years, the investment horizon in
energy sector. This is in the spirit of Alexander et al. (2012). Alternatively, we could
have matched the moments of cash flows rather than project values. As opposed to
the latter approach, the former yields similar outcomes in price-earning ratios under
both models.

10 The expression e−hi t measures the probability that the chain stays in state i longer than some
time t ≥ 0. For example, given hB = 0.16, the probability that the chain stays in regime B longer
than 5 years equals 45%.
Business Cycles and Energy Real Options Valuation 193

We describe the exact procedure of matching moments of the GBM and ABM
processes first using a one regime model. Given μ and σ of the GBM model, we
find the ABM parameters by matching the first two moments of energy project value
distribution:
√ √
T ε(T )] = κ GBM X0 e(μ − 2 (σ ) )T +σ T ε(T )
GBM 1 GBM 2 GBM
A + B[X0 + μABM T + σ ABM ,
(70)

μABM
where A = (r f )2
, B = 1/r f and the project value under the ABM process is V0ABM =
A + BX0 .
Taking expectation of both sides gives us an expression from which we solve for
μABM

μABM 1
+ f [X0 + μABM T ] = κX0 eμ T
GBM

(r )
f 2 r
 
μGBM T 1
⇒ μ ABM
= κX0 e − f /(1/(r f )2 + T /r f ).
r

From the variance of both sides of Eq. (70) we also determine the standard
deviation of the ABM process σ ABM as follows:
√ 2 √ 2
T ε(T ) = κX0 e(− 2 (σ ) )T +σ T ε(T )
1 GBM 2 GBM
ccc Bσ ABM ,
√ √
Bσ ABM 1.645 T = Quantile κX0 e(− 2 (σ ) )T +σ T ε(T )
1 GBM 2 GBM
, 0.9 ,
Quantile(·, 0.9)
σ ABM = √ , (71)
1.645 T r1f

where ε∗ is the critical value for a 90% confidence interval, implying that ε∗ = 1.645.
f f
Given that rB = 6%, rR = 6% μGBM B = 0.04, μGBMR = 0.015 σBGBM = 6% and
σBGBM = 15% we determine the parameters of one-regime as:

hR hB
μGBM = μGBM + μGBM
hB + hR B
hB + hR R
0.034 = 0.77 × 4% + 0.23 × (1.5%)

Given that the time horizon of investors is five years, i.e., T = 5, and the initial
value of cash flow is $5, i.e., X0 = $5, the ABM model’s μABM is given by

κX0 eμ T − r1f /[1/(r f )2 + T /r f ]

GBM
μABM =
 
0.572 = 38.7 × $5 × e0.027∗5 − 0.06
1
/[1/0.062 + 5/0.06].
194 T. Kenc and M. F. Ekinci

We calculate the standard deviation of cash flows under the single regime model
σ GMB as:

σ GBM = hBh+h
R
σ GBM + hBh+h
R B
B
σ GBM ,
R R
8.1% = 0.77 × 6% + 0.23 × 15%.

The ABM model’s σ ABM is then given by

Quantile(·, 0.9)
σ ABM = √ = 1.0143.
1.645 T /r f

Here we calculate the expression in Quantile(·, 0.9) by running a monte Carlo

simulation of 500,000 standard normal variates.
One-regime model’s earnings-price ratio can be derived as follows:

κ GBM = hBh+h
R
κ GBM + hBh+h
R B
B
κ GBM
R R
38.7% = 0.77 × 39.2 + 0.23 × 37.9

Figure 1 plots normally and lognormally distributed cash flows over a five years
horizon with daily increments. As expected, ABM-based cash flows take negative
values while cash flows governed by the GBM processes using one-regime parame-
ters are always positive values. The figure also demonstrates that the range of ABM
cash flows is wider than that of the GBM ones.
Table 2 reveals that the regime switching real options model based on ABM
processes produces comparable results to the model with the GBM dynamics.
However, there are significant differences across the models. Energy investors with
projects whose cash flows are normally distributed may invest later than under the
GBM-based model. Precisely, investment triggering critical value of cash flow is
$6.73 under the ABM model while it is $6.54 for investments with lognormally
distributed future cash flows. A similar conclusion can be drawn from the optimal
investment times as they are 3.2 and 3.53 years for the GBM and ABM models,
respectively. By design, price-earnings ratios and project values are rather similar.
Having dealt with the numerical implementation of the one-regime case under
both the GBM and ABM processes, we now explain how one can implement the
regime-switching real options under the two different diffusion processes. The same
basic idea developed in the single-regime case applies to the models with regime
shifts:

= κB X0 eμB
GBM
μABM
B
T
− BB /[AB (−μABM ) + BB T ] ⇒ μABM
B = 0.43
μGBM
μABM
R = κB X0 e R T
− BR /[AR (−μABM ) + BR T ] ⇒ μABM
R = 0.34
σBABM =
Quantile ( σ GBM
√
B ··· ,0.9) ⇒ σBABM = 0.69
1.645 T BB
Quantile(σR GBM
··· ,0.9)
σRABM = √
1.645 T B
⇒ σRABM = 1.51,
R
Business Cycles and Energy Real Options Valuation 195

Fig. 1 Simulation of normally and lognormally distributed cash flows

196 T. Kenc and M. F. Ekinci

Table 2 Results of the GBM-

Parameter GBM ABM
and ABM-based one-regime
real options models X 6.5 6.7
PE ratio 38.7 38.0
V 253.2 255.7
Option value 98.2 129.5
Optimal investment time 3.2 3.5

Table 3 Results of the GBM

Parameter Regime GBM ABM
and ABM regime-switching
real options models X0 Boom 5.0 5.0
X0 Recession 5.0 5.0
X∗ Boom 6.1 6.6
X∗ Recession 6.3 6.7
Price-earnings ratio Boom 39.2 39.6
Price-earnings ratio Recession 37.9 39.1
Project value Boom 196.2 197.8
Project value Recession 189.5 195.6
Option value Boom 99.9 100.8
Option value Recession 94.8 100.4

where expression Ai (−μABM ) represents the term Ai exclusive of μ’s.

Table 1 reveals that there are differences between the drift and volatility terms
under the two models. These differences reflect the fact that the GBM assumption has
multiplicative drift and volatility terms whereas those of the model with the ABM
cash flows are additive. We chose their values such that the model under the ABM
assumption generates similar regime specific values for the project values obtained
under the GBM case. The rest of the variables take identical values under both the
GBM and ABM cases. We omitted regime shifts in the interest rate and investment
costs11 and allowed regime specific values for other parameters.
Results from our numerical analysis, reported in Table 3, reveal that critical values
at which investors invest in the project vary across both regimes and models. Under
the GBM model, the threshold value is 6.1 when we are in a boom regime while it is
6.3 in a recession regime. They are 6.6 and 6.7 under the ABM model, respectively.
Both models produce sensible results. When the investment triggering critical cash
flow value between boom and recession regimes are compared, we observe smaller
threshold values under the boom regime than that under the recession regime as
expected from the fact that a boom regime is a better regime in the sense of high cash
flow values and lower cash flow volatility.

11One can also extend the models in this chapter to stochastic investment costs with regime shifts
(Hassett & Metcalf, 1999; Elliott, Miao, & Yu, 2009).
Business Cycles and Energy Real Options Valuation 197

Table 4 Probability of
Parameter Regime GBM ABM
Investing under the GBM and
ABM regime-switching Probability of investing Boom 0.093 0.16
models Probability of investing Recession 0.074 0.14
T∗ Boom 3.371 3.11
T∗ Recession 3.470 3.56

A comparison of critical values across models indicates considerable variation.

The threshold values under the ABM model are higher than those of the GBM’s.
The difference is 8.2% under the boom regime while it is 6.4% under a recession.
Furthermore, the models differ in terms of sensitivity to macroeconomic risk. It is
more pronounced under the GBM model while it is less relevant for the ABM model.
The insensitivity of the ABM model to changes in macroeconomic conditions may
stem from the additive nature of its diffusion process. Finally, given the values of the
initial cash flows and investment costs, it is trivial to compute the option values under
both boom and recession regimes for the two models using the formulas provided in
Propositions 2 and 4, respectively.
The numerical analysis also includes the calculation of probability of investing.
This experiment involves a simulation of cash flows over a fixed time period, (for
which we choose five years, T = 5), and counting the number of times X (t) ≥ Xi∗
within this period. It starts with simulating regimes st based on an initial state,
We assume that at the beginning the economy is in the boom regime. Recall that
1/hi determines the average switching times. Random numbers generated using
the exponential distribution,12 provide samples of inter-arrival times τt where τt
represents a regime change. Whenever rand(Exponential(1.0/hi )) is greater than
1/hi , a regime switch occurs. More precisely, given that the economy starts in boom
state, if rand (exp(1.0/hB )) > 1/hB , the economy switches to recession regime. We
then generate a new state of random numbers to determine a new regime switch:
rand (exp(1.0/hR )) > 1/hR . At the end of this process, we will have st for t =
0, t, 2t, · · · , 5. Having established at which regime the economy is going to
operate over time we are now in position to simulate cash flows according to both
models as follows:
   
σB2 √ σR2 √
GBM : X (t) = X0 exp μB t − t + σB ε t st + exp μR t − t + σR ε t (1 − st )]
2 2
√ √
ABM : X (t) = X0 + μB t + σB ε t st + μR t + σR ε t (1 − st ).

Table 4 underscores an important result. Probability of investing is higher for ABM

investors despite higher critical cash flows values. Investors in the ABM model are
likely to take actions with a probability of 16.3% (14.5%) under a boom (recession)
regime whereas they are 9.3 and 7.37% for GBM investors. This suggests that the

12 E.g. rand(Exponential(1.0/hi )) in Julia programming language.

198 T. Kenc and M. F. Ekinci

additivity of cash flows leads to shorter times to reach the threshold values. This
feature is evident in Fig. 1 as the range of cash flows under the ABM model is wider
than those of the GBM case. Optimal investment times across models and regimes
explain the variation in probability of investing between the models to a certain
extent. Optimal investment time under the ABM model takes a value of 3.1 years
in the boom state while it is 3.4 years for the GBM case. The numerical results on
probability of investing and optimal investment time suggest that macroeconomic
risks are pronounced for ABM investors as well.

5 Conclusions

We developed a real options approach to value energy projects whose cash flows
follow a normal distribution and subject to macroeconomic risk. In the lights of
the analytic and numerical results we obtained, we conclude that energy investors
must determine the probability distribution of cash flows from their investments.
Lognormal distribution—the most common assumption used in the literature—may
not readily characterize cash flows as they exhibit negative values from time to time.
The empirical literature suggests that the observed distribution for cash flows is
indeed a normal one.
In the continuous-time dynamic modelling, this amounts to using arithmetic Brow-
nian motion (ABM) processes to govern the stochastic dynamics in cash flows. By
adopting such stochastic differential equations in the real options valuation frame-
work, we have valued energy projects. We run a set of numerical calculations using
our ABM model as well as the standard GBM-based regime switching model. The
numerical calculations show that there are significant differences between models in
terms of investment triggering critical cash flow values, probability of investing and
optimal time to invest. ABM investors face higher critical values than GBM investors
while probability of investing values suggest that ABM firms are more likely to invest
over a fixed time period of 5 years. In other words, under the normally distributed
cash flows assumption, they may invest later than under the GBM model. Similarly,
our probability of investing numerical analysis also points to that it is more likely to
observe that investors take decisions in favour of investing more than under a model
with normally distributed cash flows. The latter is confirmed by optimal investment
time.
This paper can be extended in several dimensions. A natural extension would
involve using a richer arithmetic Brownian motion process for normally distributed
cash flows along the lines of a mean reverting process such as Ornstein–Uhlenbeck
or Vasicek processes. One can also incorporate capital structure decisions. Of course,
there is room to add stochastic dynamics into the parameters and variables of the
real options frameworks. They include interest rate and cost of investment. Finally,
this analysis can be combined with an estimation work to determine the statistical
distributional moments of cash flows.
Business Cycles and Energy Real Options Valuation 199

References

Alexander, D. R., Mo, M., & Stent, A. F. (2012). Arithmetic Brownian motion and real options.
European Journal of Operational Research, 219(1), 114–122. https://doi.org/10.1016/j.ejor.2011.
12.023
Arnold, M., Wagner, A. F., & Westermann, R. (2013). Growth options, macroeconomic conditions,
and the cross section of credit risk. Journal of Financial Economics, 107(2), 350–385. https://
doi.org/10.1016/j.jfineco.2012.08.017
Bhamra, H. S., Kuehn, L.-A., & Strebulaev, I. A. (2010). The levered equity risk premium and credit
spreads: A unified framework. Review of Financial Studies, 23(2), 645–703. https://doi.org/10.
1093/rfs/hhp082
Brennan, M. J., & Schwartz, E. S. (1985). Evaluating natural resource investments. The Journal of
Business, 58(2), 135–157.
van der Burg, J. (2018). Stochastic continuous-time cash flows: A coupled linear-quadratic model.
Ph.D. thesis, New Zealand: Victoria University Wellington.
Chen, H. (2010). Macroeconomic conditions and the puzzles of credit spreads and capital structure.
The Journal of Finance, 65(6), 2171–2212. https://doi.org/10.1111/j.1540-6261.2010.01613.x
Chen, H., Cui, R., He, Z., & Milbradt, K. (2018). Quantifying liquidity and default risks of corporate
bonds over the business cycle. The Review of Financial Studies, 31(3), 852–897. https://doi.org/
10.1093/rfs/hhx107
Chen, H., & Manso, G. (2017). Macroeconomic risk and debt overhang. The Review of Corporate
Finance Studies, 6(1), 1–38. https://doi.org/10.1093/rcfs/cfw007
DéCamps, J.-P., Mariotti, T., Rochet, J.-C., & Villeneuve, StéPhane. (2011). Free cash flow, Issuance
costs, and stock prices. The Journal of Finance, 66(5), 1501–1544. https://doi.org/10.1111/j.1540-
6261.2011.01680.x
Driffill, J., Kenc, T., & Sola, M. (2013). Real options with priced regime switching risk. International
Journal of Theoretical and Applied Finance, 16(05), 1–31. https://doi.org/10.1142/S02190249
13500283
Elliott, R. J., Miao, H., & Jin, Yu. (2009). Investment timing under regime switching. International
Journal of Theoretical and Applied Finance, 12(4), 443–463.
Fernandes, B., Cunha, J., & Ferreira, P. (2011). The use of real options approach in energy sector
investments. Renewable and Sustainable Energy Reviews, 15(9), 4491–4497. https://doi.org/10.
1016/j.rser.2011.07.102
Genser, M. (2006). A structural framework for the pricing of corporate securities: Economic and
empirical issues. Lecture Notes in Economics and Mathematical Systems 566. Berlin: Springer.
Goldstein, R., Nengjiu, J., & Leland, H. (2001). An EBIT-based model of dynamic capital structure.
The Journal of Business, 74(4), 483–512.
Guo, X. (2001). An explicit solution to an optimal stopping problem with regime switching. Journal
of Applied Probability, 38, 464–481.
Guo, X., Miao, J., & Morellec, E. (2005). Irreversible investment with regime shifts. Journal of
Economic Theory, 122(1), 37–59. https://doi.org/10.1016/j.jet.2004.04.005
Guo, X., & Zhang, Q. (2004). Closed-form solutions for perpetual american put options with regime
switching. SIAM Journal on Applied Mathematics, 64(6), 2034–2049.
Hassett, K., & Metcalf, G. E. (1999). Investment with uncertain tax policy: Does random tax policy
discourage investment. Economics Journal, 109, 372–393.
Kanniainen, J. (2009). Can properly discounted projects follow geometric brownian motion? Math-
ematical Methods of Operations Research, 70(3), 435–450. https://doi.org/10.1007/s00186-008-
0275-0
Kenc, T., & Driver, C. (2020). Leverage, uncertainty and investment decisions. Economics Letters,
190(May). https://doi.org/10.1016/j.econlet.2020.109052
Lambrecht, B. M. (2017). Real options in finance. Journal of Banking & Finance, 81(August),
166–171. https://doi.org/10.1016/j.jbankfin.2017.03.006
200 T. Kenc and M. F. Ekinci

Leland, H. E. (2007). Financial synergies and the optimal scope of the firm: Implications for mergers,
spinoffs, and structured finance. The Journal of Finance, 62(2), 765–807.
McDonald, R., & Siegel, D. (1986). The value of waiting to invest. Quarterly Journal of Economics,
101(4), 707–727.
Schwartz, E. S., & Lenos, T. (2001). Real options and investment under uncertainty. MIT Press.
Veronesi, P. (1999). Stock market overreaction to bad news in good times: A rational expectations
equilibrium model. The Review of Financial Studies, 12(5), 975–1007.
Weber, F., Schmid, T., Pietz, M., & Kaserer, C. (2011). Simulation-based valuation of project
finance—Does model complexity really matter? SSRN Electronic Journal. https://doi.org/10.
2139/ssrn.1594116
Understanding the Electricity Switching
Behavior of Industrial Consumers:
An Empirical Study on an Emerging
Market
Murside Erdogan, Selin Metin Camgoz, Mehmet Baha Karan,
and M. Hakan Berument

1 Introduction

Starting from the late 1990s, the paradigm change in the energy markets has been
inevitably affecting all energy segments, including electricity markets and players.
During the period, electricity production and services have privatized in line with
the implemented neo-liberalism policies, and then electricity services are unbundled
and opened to market competition. This process, initially led by the UK and Scan-
dinavian Countries in Europe, has become the basic policy of the European Union
(EU) and has quickly turned into a global trend. Such that the electricity market
has changed significantly from production to consumption, and the electricity used
by consumers has rapidly evolved from a social good provided by the states to a
commercial product. Furthermore, the new factors currently shaping this era, such
as global warming, energy security, and increasing global competitive conditions,
have put intense pressure on the markets to use electricity more efficiently (Karan &
Kazdağli, 2011).
The new electricity paradigm aims to open up competition by removing the
barriers in the retail electricity market and reducing the electricity prices by creating
pressure on energy companies. Retail competition is expected to decrease the
sourcing costs of electricity, as well as the operating costs of retail, such as billing,
metering, and credit assessment. When consumers’ switching operations become

M. Erdogan · M. H. Berument
Bilkent University, 06800 Cankaya, Ankara, Turkey
e-mail: [email protected]
S. Metin Camgoz (B) · M. B. Karan
Department of Business Administration, Hacettepe University, 06800 Cankaya, Ankara, Turkey
e-mail: [email protected]
M. B. Karan
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 201
A. B. Dorsman et al. (eds.), Applied Operations Research and Financial Modelling
in Energy, https://doi.org/10.1007/978-3-030-84981-8_10
202 M. Erdogan et al.

more accessible with low entry and exit costs, retail competition should lead to end-
user prices that reflect the value that moves along with wholesale prices (Poudineh,
2019). Therefore, the switching rates of final customers is a commonly used as an
indicator for the level of buyer commitment in a market, since it provides informa-
tion on the participation of consumers in the market activities and on the effects of
liberalization (CEER, 2016; Concettini & Créti, 2013). In a market where provider
switching is difficult, market regulators or policymakers can hold consumers despite
relatively high prices (Shin & Managi, 2017).
Notably, the EU has focused on its practices in this context and has rapidly
increased the rate of electricity switching among member countries in the past ten
years. The switching rate for electricity household customers in 2018 was the highest
in Norway (21.4%), and more than 10% in Finland, Germany, Great Britain, Ireland,
Norway, Portugal, Spain, and Sweden. In many countries, the switching frequency
in 2018 was on average higher than in past years. Annual switching rates of non-
household customers in the EU were also increased steadily due to economic incen-
tives. Czechia, Italy, Lithuania, Poland, Portugal, and Spain have more than a 25%
rate (CEER, 2018).
Turkey is one of the important emerging economies with a rapidly growing elec-
tricity market and consumption. According to the projections of the Ministry of
Energy and Natural Resources of Turkey (https://www.enerji.gov.tr/en-US/Pages/
Electricity), the electricity demand of Turkey is expected to reach 357.4 TWh in
2023. Following the implementation of the EU, Turkey has started reforms in the
energy market in 2001 by privatizing and unbundling manufacturing and distribution
sectors. Aiming to open the electricity market to full competition, serious steps are
taken towards liberalizing the retail market with the distribution of privatizations and
the elimination of the eligible consumer limit year by year. In line with this target, the
eligible consumer limit—which was 9 GWh/year in 2003—was decreased steadily
till 1,400 kWh/year in 2020. As a result, lowering the eligible consumer limit enabled
the theoretical market opening rate, which was 23% in 2003, to reach 99% in 2020.
Despite the aforementioned actions, the switching rates for electricity consumers
in Turkey are still meager, and the retail electricity market is not as competitive as
expected. Given that, little is known about the reasons why consumers remain loyal
to a particular electricity energy supplier or why they switch suppliers. As exhibiting
passive responses in switching might inevitably deteriorate the open market system
of the electricity sector, the comprehension of those factors is worth examining
particularly in Turkey where the electricity sector is being restructured.
Although the effects of economic factors are quite straightforward in switching
decisions, the switching behaviors of consumers do not solely depend upon legal
and administrative decisions (He & Reiner, 2015; Samuelson & Zeckhauser, 1988;
Shin & Managi, 2017). It is also noted that understanding the consumers’ behavioral
expectations and responses is also recommended to create competitive market condi-
tions. Therefore, the current study aims to explore several psychological/behavioral
factors (possible drivers and buffers) that might influence the electricity supplier
switching behaviors in Turkey. Particularly, it explores the influence of behavioral
variables such as the risk of switching, cost of switching, attractiveness of switching,
Understanding the Electricity Switching Behavior of Industrial Consumers … 203

perceptions of the service quality, and market competition in predicting the switching
behaviors of consumers. The theoretical framework of the proposed hypotheses is
built on the status quo (mooring) effect, push, and pull effects.
The primary foci and the significant contribution of this current research refer
to its sample of large-scale consumers. A vast proportion of electricity consump-
tion is employed by industrial consumers other than residential consumers. Despite
their potential importance and higher electricity consumption rates, to the authors’
knowledge, no research investigates which factors drive the industrial consumers to
respond to new switching opportunities. One might presume that organizations might
be more rational actors as they usually work with consultants and experts. Nonethe-
less, it might be possible that their routines, relationships with suppliers, interper-
sonal relationships, and/or other behavioral factors could affect their switching or
not switching decisions. The presence of inertia in industrial consumers compared to
domestic consumers might reveal even more significant problems which deteriorate
the liberalization of the electricity markets (Wieringa & Verhouf, 2007). Therefore,
providing insight into understanding the switching behavior of industrial consumers
in the Turkish electricity market as an emerging market is valuable.
The current chapter is organized as follows: After the introduction part, we provide
a section on the current state of the electricity market in Turkey along with the dereg-
ulation issues. Then a literature review with the theoretical background regarding the
proposed hypotheses is provided. In the methodology section, we define the data,
design of the survey questionnaire, the preliminary statistics, and the binary logistic
model. Subsequently, we present the study findings, draw conclusions, and highlight
the avenues for future research with respect to the limitations of the study.

2 An Overview of Turkish Electricity Retail Market

Developments

Following the changes in the energy paradigm in developed countries, Turkey has
begun to reform the energy market in the early 1990s. However, the essential step took
place in 2001 with the establishment of the Energy Market Regulatory Authority.
Organized Electricity Market has started to operate with the Balancing Mechanism in
2006. A few years later, Day-Ahead Planning, Balancing Power Market, and Hourly
Settlement is initiated. In the meantime, the Renewables Energy Support Mechanism
(YEKDEM) has been passed into a law to promote renewable investment in Turkey.
With the introduction of the Day-Ahead Market and Collaterals Mechanism, the
second stage of an organized electricity market establishment has started in 2009. In
2015, the Intra-day market was set which led to market sophistication and additional
means for balancing, especially for renewables (PWC, 2012).
After 2015, the privatization of the distribution companies and gradually
decreasing of the eligible consumer limit are the crucial steps that have been taken
204 M. Erdogan et al.

Eligible Consumer Limit Theoretical Competitive Market(%)

Actual Competitive Market(%)
35000 100
90
30000
80
25000 70
Kwh/Year

20000 60

%
50
15000 40
10000 30
20
5000
10
0 0
2011 2012 2013 2014 2015 2016 2017 2018
Years

Fig. 1 Eligible consumers and market competitiveness rates in Turkish Market. Source EMRA
(2018) Electricity Market Development Report

towards the liberalizing of the retail electricity market. Despite all these devel-
opments, the actual competitive electricity market could not grow as expected,
especially in recent years (see Fig. 1).
In 2018, the Energy Market Regulatory Authority (EMRA) has started to imple-
ment the Last Resort Supply Tariff (LRST) that is based on a cost-based tariff method-
ology to overcome this problem. The tariff has been introduced for the electricity that
is provided by the regionally-authorized supply companies to eligible customers, who
do not buy electrical energy through bilateral agreements and consume more than 50
GWh electricity. Then, the tariff limit has been decreased to 10 GWh electricity in
2019. In this way, the general practice of determining the active energy price within
the regulated tariff has been replaced by an upper price limit tariff that is formulated
as the sum of Market Clearing Price and YEKDEM Premium,1 multiplied by 1.128.
As a result of this transformation, the tariff has been indexed to the market price. The
new tariff enabled eligible consumers to lower their electricity prices by bargaining,
substantially; such as saving more than $ 50 million in 2016 (World Energy Council,
2018).

1 YEKDEM Premium is the amount that is reflected in consumer electricity tariffs for financing
the subsidy that the state pays to renewable energy producers. The addition of this premium to the
electricity tariff has not been welcomed by the electricity consumers. Nevertheless, all consumers
shared the high cost of producing renewable energy.
Understanding the Electricity Switching Behavior of Industrial Consumers … 205

2.1 Literature and Theoretical Framework

The rise of privatization in service companies such as telecom, banking, gas, and
electricity has initiated the supplier switching in the service industry of the devel-
oped countries as a result of the market competition (Keaveney, 1995). This attempt
has rapidly received attention from scholars and thus pioneering research stream
focused on identifying the factors involved in consumer supplier switching deci-
sions in service industries. Among those pioneering studies, the theoretical frame-
work of Keaveney (1995) suggested several factors of economic (e.g. price and
switching cost) (Klemperer, 1995; Waterson, 2003), inconvenience, service encoun-
ters, core service failure, response to failure, and market competition as the factors of
switching behaviors. However, the framework of Keaveney (1995) underestimated
the role of psychological, demographic, and behavioral patterns in switching deci-
sions of customers (He & Reiner, 2015). Building upon Keaveney’s (1995) work,
Bansal and Taylor (1999) argued that not only economic but also psychological
factors influence the switching behaviors of consumers. Accordingly, economic
factors refer to relatively more objective and straightforward factors that represent a
cost or benefit to consumers, whereas psychological factors refer to more subjective
and complex behavioral factors that might include consumer expectations, customer
satisfaction, perceived service qualities (Athanassopoulos, 2000), social influence,
perceived market conditions and consumer loyalty (He & Reiner, 2015). Some other
researchers also emphasized the importance of demographic variables in predicting
the probability of switching versus non-switching behaviors of consumers such as
income (Ek & Söderholm, 2008; Gamble et al., 2009), age (He & Reiner, 2015),
education (Ek & Söderholm, 2008) and gender (Gamble et al., 2009).
Within the last decades, scholars’ growing attention in exploring supplier
switching behavior has also shifted to the energy sector. Giulietti et al. (2005)’s study,
as one of the leading studies for the energy sector, argues that most of the house-
holds in the UK gas retail market seem to be reluctant to switch their gas providers
due to switching costs or searching costs. Likewise, Gamble et al. (2009) reported a
similar finding in the Swedish retail electricity market and reported that compared to
other markets, consumers are more reluctant to switching in the electricity market.
Their study findings indicate that the negative attitude toward switching suppliers
can increase with higher levels of loyalty to an existing supplier, higher levels of
information search costs, and can decrease with expected economic benefits. In their
follow-up recent study, Gamble et al. (2009) have underlined the barriers to switching
behaviors as loyalty to the incumbent supplier, high information search costs, and
too small economic benefits.
Ek and Söderholm (2008) have investigated the electricity switching behavior of
consumers in Sweden retail market and indicate that when consumers expect high
financial gains, they are more inclined to action-oriented towards supplier switching
behavior. When the level of expected economic benefit is low, consumers prefer to
maintain the current situation. Furthermore, when consumers perceive relatively high
information and search costs, they are less likely to switch to an alternative electricity
206 M. Erdogan et al.

supplier. The authors base their explanations on the theoretical framework of Push-
Pull-Mooring (status quo) effects. Accordingly, both the push (i.e. consumers are not
satisfied with their current suppliers) and the pull effects (i.e. the perception that the
alternative supplier is better than the current one) usually lead to active responses of
consumers in the forms of comparing and evaluating the other possible alternatives;
in contrast, the status quo refers to a state of inertia where consumers favor their
status quo, stay with the present supplier and neglect the potential cost savings.
Yang (2014) examined household switching behavior and the barriers among
residential energy consumption in the Danish market. Yang (2014)’s findings indi-
cated that perceived economic benefits and psychological factors had a limited power
to explain the switching behavior of Danish consumers. Nonetheless, Yang (2014)
also postulated the necessity of simplifying the switching behavior by decreasing
managerial barriers and improving communication with consumers. Giulietti et al.
(2005) argue that the attractiveness of switching behavior and tendency of switching
is reported as low due to the reason that consumers misperceive the search costs as
higher than their actual costs. With the data gathered from the Japanese household
survey, Shin and Managi (2017) emphasized the importance of switching behavior
in reflecting deregulation reform effectively. Their results showed that electricity
switching behavior is strongly associated with consumer satisfaction.
The only study exploring the switching and demand responses in the Turkish
electricity market is conducted by Şirin and Gönül (2016). With the data gathered
from 124 eligible household consumers, the study findings demonstrate that eligible
household consumers tend to search for cheaper alternatives consistent with the
traditional economic models. However, behavioral factors are also shown to influ-
ence consumers’ decisions in such a way that although consumers have access to
information for cheaper alternatives, they prefer to stick with their current electricity
supplier. In addition to economic factors, social and peer influences (e.g. sugges-
tions of neighbors and friends) seem to be the driving factors of electricity switching
behaviors of Turkish consumers.
Among the current study’s variables of interest, ‘Switching cost’ refers to the
perception of the additional costs that a customer is required to pay for terminating
the current relationship with its supplier or for switching from one service firm to an
alternative other (Porter, 1980). Patterson and Smith (2003, p. 108) define switching
costs as “the perception of the magnitude of the additional costs required to terminate
a relationship and secure an alternative one”. It is the cost paid for switching from
one service firm to an alternative other, or the termination cost for leaving the current
provider. Each consumer is expected to make a specific investment when entering
into a relationship with an alternative supplier which in turn results in switching costs.
Although the cost of switching seems to be related to economic issues, it also includes
the costs of psychological and emotional bonds that are developed with the service
firm (Sharma & Patterson, 2000). From this perspective, switching costs might be
regarded as the social bond, trust, and loyalty that have been developed between the
supplier and the customer over a period of time. It might also create a dependence
on the consumer on the service provider (Colgate & Lang, 2001). Accordingly, if the
cost of switching is perceived as high by the consumer, the consumer is more likely
Understanding the Electricity Switching Behavior of Industrial Consumers … 207

to stay passive and to continue with its current service provider. The theoretical
argumentation could be based on the status quo effect (or mooring effect), which
refers to a type of cognitive bias of individuals’ tendency to prefer to do things the
same by doing nothing, or sticking to their prior decision (Samuelson & Zeckhauser,
1988). Nebel (2015) points out that under the circumstances of high uncertainty,
choice overload, and high deliberation costs, the status quo effect is more likely. When
industrial consumers think that the cost for searching and evaluating information of
alternative electricity suppliers is high, and assume that they need to pay a high cost
for the benefits lost by changing their current electricity provider, they are more
prone to the status quo effect. Therefore, we propose that when switching costs are
perceived as high, the industrial customers tend to stick to their decisions that have
been made in the past and become less likely to switch their electricity supplier.
A similar argumentation can also be valid for the risk of switching. Dowling
and Staelin (1994) define the risk of switching as “the customer’s perception of the
uncertainty and adverse consequences of buying a product/service” (p. 119). The
perceived risk of represents the likelihood of negative consequence of loss and/or
danger in a possible switching situation. If the perceived risk of switching is high,
it means that consumers think that they will experience problems with continuity
of supply, payments, reply to complaints, when they switch to another electricity
supplier. Under those circumstances, the status quo might prevail and impede indus-
trial customers from switching one electricity service provider to another. Therefore,
we propose:
H1: When the cost of switching increases, the likelihood of electricity switching
behavior decreases.
H2: When the risk of switching increases, the likelihood of electricity switching
behavior decreases.
The attractiveness of switching is considered as a potential influencer that has a
role in influencing switching or non-switching behavior of the industrial consumers.
It refers to the assessment of the service providers the consumers intend to change
and switch to (Chang et al., 2017). The assessment regarding the attractiveness of the
other alternative suppliers can be reached by various information channels of media,
commercials, and/or expert recommendations. When the attractiveness of switching
is considered low, consumers may perceive that there are not many differences among
electricity suppliers in general and in terms of their prices and service qualities. This
situation can theoretically be framed with the pull effect of the Push–Pull-Mooring
theory. Accordingly, pull factors refer to the positive and appealing factors that attract
and motivate the target to move towards an attractive destination. In this case, the
quality and the number of the attractiveness of the alternative service providers
are seen as pull factors. Such that, once the consumers consider that there is no
difference among the alternative suppliers, then unattractiveness of the alternatives
might discourage industrial customers to stay with their current electricity supplier.
Thus:
H3: When the attractiveness of switching decreases, the likelihood of electricity
switching behavior decreases.
208 M. Erdogan et al.

The perception of Service quality, on the other hand, is defined by Parasuraman

et al. (1985) as the result of a comparison between what consumers consider the
service should be and their perceptions about the actual performance offered by the
service provider. Parasuraman et al. (1985)’s well-established SERVQUAL model
includes five dimensions of the service experience as reliability, responsiveness,
empathy, assurance, and tangibility. It is also defined as the overall assessment of
the service provider’s performance concerning the comparison of what they expect
and what they receive (Lien et al., 2017). Although service quality can be thought
of as a global dimension, the multidimensional view of service quality also captures
sub-dimensions of the qualities of the commercial, administrative, communicative,
and physical services. Research has shown that service quality evaluations of the
customers play an important role in influencing their staying or switching their service
providers (Chang et al., 2017). Building upon the push–pull-mooring theory, the
service quality perceptions of the consumers might act as a push effect. In the energy
sector, push factors include the negative perceptions of the currently used service
provider that leads the consumers to search for other alternatives (Chang et al., 2017).
That is to say when the industrial consumers perceive that their energy providers’
service qualities are low in several dimensions (e.g. commercial, administrative,
communicative, and physical), their tendency for switching their service supplier
increases. Alternatively, the social-exchange theory can also underlie the framework
for switching decision behavior. The main premise of the social exchange theory
is the norm of reciprocity that indicates the importance of an action by one party
leads to a response by another party (Eisenberger et al., 2004). Accordingly, positive
reciprocity includes the tendency to offer positive treatment in response to positive
treatment. If the consumers perceive that their current electricity supplier offers
them high-quality service, then they respond to their reciprocity by not switching
and maintaining their relationship with their current service supplier. Thus, the study
proposes that:
H4: When the service quality perceptions (a. commercial, b. administrative,
c. communicative, and d. physical quality) increases, the likelihood of electricity
switching behavior decreases.
Market competition refers to the competitiveness of the environment of the service
sector. If the environment is highly competitive, a large number of alternative prod-
ucts and services are available for consumers to select from. In those highly compet-
itive environments, customers are not expected to be loyal unless they are satisfied.
Nonetheless, in a low market competitive environment, there are not many alternative
products and services. Therefore, the satisfaction of the customers is not expected to
have an impact on their loyalty (Yee et al., 2010). We propose that:
H5: When the perceived market competition increases, the likelihood of electricity
switching behavior increases.
Understanding the Electricity Switching Behavior of Industrial Consumers … 209

3 Methodology

3.1 Data Description and Procedure

This study is a part of a larger study that has been carried out to estimate the electricity
switching behaviors of industrial consumers. Industrial consumers refer to the group
of large-scale electricity consumers that are allowed to either make a bilateral contract
with any preferred retailer or are allowed to purchase electricity from an incumbent
retailer with Last Resort Supply Tariff (LRST).2
The survey data used as the primary data source. The data is collected from
Turkish eligible industrial electricity consumers with the criteria of using LRST and
consuming more than 10 million kWh electricity/year during 2019. This criterion
corresponds to several diversified companies, with the majority of them are located
in Organized Industry Zones (OIZ) and the other non-OIZ companies compromise
the large industrial companies of iron, steel cement, ceramic, and automotive. The
survey package is distributed to energy consultants and/or supervisors of the compa-
nies by the authors. The data has been collected with the support from The Ministry of
Industry and Technology, The Turkish Steel Producers, The Turkish cement Manu-
facturers Association between the period of June–September 2019. After the elimina-
tion of incomplete survey questionnaires, a total of 82 cases remained in the sample.
Among those 82 companies nearly half of them (48%) are OIZ companies with a
high company diversity, while the remaining are iron, steel, cement, and other indus-
trial companies. Although the full descriptions of the companies cannot be provided
for confidentiality reasons, Appendix 1 offers brief information (e.g. sector, firm
size, annual profit, electricity consumption, etc.) regarding descriptive statistics of
the companies.

3.2 Measurement of the Study Variables

The survey questionnaire of the current research is constructed to capture the

measures of switching cost, risk of switching, the attractiveness of switching, percep-
tions of the service quality (commercial, administrative, communicative and phys-
ical), market competition, and some descriptive items. The measures were drawn
from well-established instruments based on the studies of Caceres and Paparoidamis
(2007), Jones et al. (2002), Wieringa and Verhoef (2007), Yee et al. (2010). The
Turkish translation of the measures is conducted by the authors using a back-
translation procedure. No anonymous items are detected as a result of the translation
procedure. The detailed information regarding the survey variables is given below:

2 The electricity consumption limit is determined by EMRA for the LRST, data for 2019 is used in
this chapter.
210 M. Erdogan et al.

The risk of switching was measured by 5 items borrowed from Wieringa and
Verhoef (2007). The sample item was ‘If I switch to another electricity supplier, I
expect problems with solving problems.’ The response format ranges from 1 (strongly
disagree) to 5 (strongly agree). Higher values denote higher risk of switching. The
internal reliabilities of the items are measured with Cronbach alpha coefficients, a cut
off value of 0.70 and/or higher are considered as acceptable (Nunnaly & Bernstein,
1994). In the current study, the alpha coefficient was found 0.87, suggesting that the
items have relatively high internal consistency.
Switching cost was measured with 4 items adapted from Jones et al. (2002) and
Yee et al. (2010). ‘Customers have to pay a high cost for the benefits lost by changing
the service provider’ denoted as a sample item for this factor. A five-point Likert-type
scale 1 (strongly disagree) to 5 (strongly agree) is used as a response format. The
Cronbach alpha reliability score was found as 0.90.
The attractiveness of switching was measured with a total of 3 items developed by
Wieringa and Verhoef (2007). The sample item was ‘There are not many differences
among electricity suppliers’. The response format ranged from 1 (strongly agree)
to 5 (strongly disagree). The higher scores were indicative of the attractiveness of
switching. The Cronbach alpha reliability score was 0.89.
Service quality perceptions are measured with a total of 13 items in four dimen-
sions adapted from Caceres and Paparoidamis (2007). Those dimensions are the
perceptions of commercial service quality (5 items, i.e. My supplier respects its
promises), administrative service quality (3 items, i.e. Invoices sent are always
precise), communicative service quality (2 items, i.e. My supplier informs me suffi-
ciently) and physical service quality (3 items, i.e. My electricity supplier helps me
in solving the problem with electricity distribution). The Cronbach alpha reliability
scores range from 0.81 to 0.92.
Market Competition is measured with 3 items borrowed from Yee et al. (2010).
‘Attractive promotions and tariffs are offered in the market’ is the sample item for
market competition. The Cronbach alfa was found as 0.82.
The scores of the measures are obtained by taking the averages of items within their
factor group. The respondents are also asked a binary variable (yes or no) whether
they have changed their electricity supplier over the past 5 years. This variable served
as the main dependent variable of the study. Additionally, the survey questionnaire
included a self-reported motivation questionnaire in which the respondents are asked
to provide the reasons for switching or non-switching their electricity service supplier.

4 Findings

4.1 The Confirmatory Factor Analyses

After data screening, the data of the study were checked for univariate and multi-
variate outliers. The closer investigation of the skewness and kurtosis scores displayed
Understanding the Electricity Switching Behavior of Industrial Consumers … 211

that the data was closed to normal distribution. To test the factorial structure of the
study variables and how well the measurement data fit the model, the analysis has
continued with confirmatory factor analysis (CFA) with the maximum likelihood esti-
mation method. Within the CFA model, measures in the data set were considered to be
the observed indicator of their underlying constructs. Chi-square/degrees of freedom
(χ2/df), Goodness of Fit Index (GFI), Comparative Fit Index (CFI), Adjusted Good-
ness of Fit Index (AGFI), Tucker-Lewis Index (TLI) and Root Mean Square Error
Approximation (RMSEA) indexes are used as the proxies of goodness-of-fit in eval-
uating the model fit. Accordingly, for the service quality items, 4 factor-structure
solution yielded a relatively good fit to data after deleting one item from adminis-
trative service quality and one item form commercial service quality due to their
low factor loadings (χ2 = 90.02, χ2/df = 1.87, GFI = 0.92, CFI = 0.92, AGFI =
0.89, TLI = 0.89, RMSEA = 0.07). Then, we conducted another CFA for the items
for measuring the risk of switching, switching cost, the attractiveness of switching,
market competition. After eliminating one item from switching cost, and one item
from risk of switching, the goodness of fit values of the model showed satisfactory
fit (χ2 = 155.7, χ2/df = 2.64, GFI = 0.98, CFI = 0.94, AGFI = 0.91, RMSEA =
0.08). After CFA, the Cronbach alpha reliability coefficients were calculated.

4.2 The Preliminary Analyses

The correlations among main study variables and descriptive statistics were displayed
in Table 1. The majority of the correlation scores among the study variables were

Table 1 Correlations between variables

1 2 3 4 5 6 7 8
1 Risk of 1.000
switching
2 Cost of 0.701 1.000
switching
3 Commercial −0.298 −0.274 1.000
service quality
4 Administrative −0.190 −0.161 0.815 1.000
service quality
5 Communicative −0.281 −0.262 0.694 0.547 1.000
service quality
6 Physical service −0.123 −0.092 0.463 0.481 0.389 1.000
quality
7 Unattractiveness 0.165 0.222 −0.318 −0.134 −0.227 −0.010 1.000
of switching
8 Market −0.041 −0.067 0.178 0.098 0.338 0.197 −0.068 1.000
competition
212 M. Erdogan et al.

not above 0.70 and VIF values were below 0.5, suggesting that multicollinearity was
not a serious issue.
The dependent variable of the current study model refers to the electricity
switching behavior of the company. To distinguish the companies that change their
electricity service supplier, and the companies that maintain with their existing
supplier, we create a dummy variable. If the company has switched in the last years,
the value of the switching variable is coded as 1, and otherwise 0. Approximately
80% of the companies (67 switched, 16 did not switch) have reported that they have
switched their electricity suppliers within the last 5 years. The switching ratios are
also investigated both for OIZ and non-OIZ companies. Out of 40 OIZ companies,
32 of them reported that they had switched their electricity supplier while eight of
those had not. Table 2 provides the summary statistics of the study variables for the
pooled sample group, the switcher group, and the non-switcher group separately.
To compare the mean differences in terms of key study variables between the
switcher and non-switcher groups, we conducted t-test statistics. The findings of t-
test statistics display that the mean values of the perceived service quality dimensions
(commercial, administrative, communicative, and physical) are statistically higher
in the switcher group than those of the non-switcher group. Likewise, t-test statistics
provide evidence that the mean values of risk of switching, switching costs, the
attractiveness of switching (higher scores indicate unattractiveness of switching) of
the switchers’ group are significantly lower compared to those scores in the non-
switcher group. No significant difference is reported between the switcher and non-
switcher group for the variable of market competition.
Before testing the binomial model, we also listed the frequencies of the self-
reported reasons for switching and non-switching decisions. As seen in Table 3,
‘lower prices of the new provider (71.6%)’ is reported as the main reason for
switching, followed by ‘Increase in the price of previous retailer (19.4%)’, ‘Dissolu-
tion of a contract (13.43%)’ and ‘Better quality of service of the new service provider
(11.9%)’. In terms of the reasons for not to switch; ‘familiarity with the current service
provider (37.50%)’, ‘satisfaction with the service of the current electricity provider
(25%)’, ‘satisfaction with price of current electricity retailer (18.75%)’, and ‘Do not
see benefits of switching providers (12.5%) were listed as the primary reasons as
anticipated.

4.3 Binomial Logistic Regression Model

In testing the hypotheses of the current study we utilize binary logistic regression
models. Logistic regression (LR) implements maximum probability estimation after
converting the dependent variable to a logit variable (the natural log of the proba-
bilities of the variables that occur or do not change) according to the independent
variables. The advantage of logistic analysis is that it does not require strict assump-
tions (e.g. large data set, normality assumption) as it does in OLS models (Tabachnick
& Fidell, 2001). In our model, the binary dependent variable (electricity switching)
Understanding the Electricity Switching Behavior of Industrial Consumers … 213

Table 2 Descriptive statistics of the pooled sample, the switchers, and non-switchers
Variables of Statistics Whole sample Switchers group Non-switchers t-stat
interest (1) group (0)
Risk of switching N 82 67 15
Mean 2.133 1.955 2.95 0.001***
St. Dev 0.932 0.849 0.902
Min 1.000 1.000 1.000
Max 4.250 4.000 4.2500
Cost of switching N 82 67 15
Mean 1.747 1.627 2.267 0.013**
St. Dev 0.717 0.642 0.842
Min 1.000 1.000 1.000
Max 4.500 3.000 4.500
Commercial N 79 64 15
service quality Mean 3.935 4.006 3.613 0.115
St. Dev 0.719 0.673 0.86
Min 2.000 2.200 2.000
Max 5.000 5.000 5.000
Administrative N 79 64 15
service quality Mean 4.146 4.219 3.844 0.139
St. Dev 0.701 0.646 0.881
Min 1.667 2.000 1.667
Max 5.000 5.000 5.000
Communicative N 79 64 15
service quality Mean 3.079 3.547 2.833 0.029**
St. Dev 0.854 1.129 1.047
Min 1.000 1.000 1.500
Max 5.000 5.000 5.000
Physical service N 79 64 15
quality Mean 3.419 3.094 2.956 0.617
St. Dev 1.138 0.828 0.975
Min 1.000 1.667 1.000
Max 5.000 5.000 5.000
Unattractiveness N 82 67 15
of switching Mean 2.755 2.667 3.2 0.019**
St. Dev 0.926 0.948 0.699
Min 1.000 1.000 2.000
Max 4.000 4.000 4.000
(continued)
214 M. Erdogan et al.

Table 2 (continued)
Variables of Statistics Whole sample Switchers group Non-switchers t-stat
interest (1) group (0)
Market N 82 67 15
competition Mean 3.175 3.164 3.2 0.853
St. Dev 0.809 0.855 0.621
Min 1.000 1.000 2.000
Max 5.000 5.000 4.500
Note t-statistics are reported under the corresponding estimated coefficient. *** is for the level of
significance at 1%, ** is for the 5% and * is for the 10%

Table 3 Reported reasons for switching

Reasons to switch (N = 67) Ratio (%) Reasons not to switch (N = 16) Ratio (%)
Lower prices of new retailer 71.64 Satisfied with the price of current 18.75
electricity retailer
Increase in the price of the 19.40 Familiar with current retailer 37.50
previous retailer
Better quality of service of new 11.94 Satisfied with the service of current 25.00
retailer electricity retailer
Worse quality of service of the 5.97 Not seeing the benefits of switching 12.50
previous retailer
Recommendation of energy 1.49 Switching takes too much time 12.50
consult
Breach of contract 7.46 Too much trouble to switch 6.25
Bankruptcy of retailer 2.99 Experienced difficulties while 6.25
switching
Dissolution of a contract 13.43 Difficult to compare of price 6.25
Note Respondents were asked to choose all causes that applied, so the sums exceed 100%

was assigned to a value of (1) when the companies switch their electricity supplier
or (0) when the companies have not switched.
We employ two binomial logistic models for the pooled dataset. In Model 1,
all of the predictor variables were entered into the model for predicting the likeli-
hood of switching the electricity supplier. Model 2 refers to the Stepwise Binomial
Logistic regression model in which the behavioral predictor variables are eliminated
concerning their significance levels.

Si = ai0 + ai1 RS + ai2 CS + ai3 COMSQ + ai4 ASQ + ai5 CSQ

+ ai6 PSQ + ai7 US + ai8 MC + ei (Model 1)

Si = ai0 + ai1 RS + ai7 US + ai2 MC + ei (Model 2)

Understanding the Electricity Switching Behavior of Industrial Consumers … 215

S Switching, If a switch, the value is 1, otherwise 0

CS Cost of Switching
COMSQ Commercial Service Quality
ASQ Administrative Service Quality
CSQ Communicative Service Quality
PSQ Physical Service Quality
US Unattractiveness of Switching
MC Market Competition

The examination of the binomial logistic regression results for both Model 1 and
Model 2 are displayed in Table 4 along with the goodness of fit tests, regression
coefficients, Wald statistics, odds ratios, and significance levels.
As shown in Table 4, Model 1 is found significant regarding the log-likelihood
test value as −26.886. The overall fit of the models was checked by a number of the
goodness of fit tests, including, Hosmer–Lemeshow (Hoslem), traditional chi-square
tests, and Nagelkerke tests. The Hoslem test was considered more robust than the
traditional chi-square test, particularly if continuous covariates were in the model or
the sample size is small (Garson, 2010). The Hoslem test was insignificant which
indicated that the model adequately fit the data. Likewise, the values of Nagelkerke,
Cox-Snell, McFadden Pseudo were reported as 0.40, 0.25, and 0.40 respectively.
The percentage of the overall accuracy of our estimates was more than 83% in
Model 1 in which the threshold of 50% accuracy was regarded as high. According
to Wald’s criterion, risk of switching, (un)attractiveness of switching, administrative
and communicative service quality factors were found to be significant in predicting
the likelihood of electricity switching behavior. Thus, an examination of the coeffi-
cient ratios of the significant variables revealed that the odds of switching behavior
were associated with an increase in service quality perceptions of administrative and
communicative dimensions. Such that one-unit increase in administrative service
quality perceptions score was associated with an approximate increase by a factor of
4.6, whereas a one-unit increase in communicative service quality perceptions was
associated with an approximate increase by a factor of 2.5 in the odds of switching
electricity supplier. Furthermore, the variables of risk of switching and attractive-
ness of switching were found to be decreasing the likelihood of switching behavior
for electricity consumers. The odds ratio indicated that a unit increase in switching
risk decreased the likelihood of switching behavior by 0.661 units. Likewise, a unit
increase in attractiveness of switching (higher scores represent unattractiveness of
switching), decreased the likelihood of switching behavior by 0.589 units.
Model 2 (Stepwise binomial logistic model), was reported as significant regarding
the loglikelihood test value as −30.40. The Hoslem test was insignificant which
indicated that the model adequately fit the data. Likewise, the values of Nagelkerke,
Cox-Snell, McFadden Pseudo were reported as 0.30, 0.19, and 0.22 respectively. The
percentage of the overall accuracy of the estimates was %67. According to Wald’s
criterion, only the variables of risk of switching and attractiveness of switching
were found to be significant in predicting the likelihood of electricity switching
behavior. Similar to Model 1, the odds ratio indicates that a unit increase in the risk
216 M. Erdogan et al.

Table 4 Results of logistic regression with the whole sample

Variables Model 1 Model 2
Coeff Wald Odds ratio Coeff Wald Odds ratio
Risk of switching −1.080** 4.457 0.339 −1.200*** 10.434 0.301
(−2.111) (−3.230)
Cost of switching −0.391 0.379 0.676
(−0.615)
Commercial service −1.617 2.004 0.198
quality (−1.416)
Administrative service 1.527* 2.723 4.604
quality (1.650)
Communicative service 0.920* 2.765 2.509
quality (1.663)
Physical service quality −0.555 0.993 0.574
(−0.997)
Unattractiveness of −0.890* 3.015 0.411 −0.711* 3.069 0.491
switching (−1.736) (−1.752)
Market competition −0.582 1.099 0.559
(−1.048)
Constant 8.213** 3687.92 6.544*** 695.29
(2.096) (3.777)
Observations 79 82
Log Likelihood −26.886 −30.4
Akaike Inf. Crit 71.772 66.801
McFadden pseudo 0.300 0.221
R-squared
Cox-Snell pseudo 0.253 0.19
R-squared
Nagelkerke pseudo 0.407 0.309
R-squared
Percent correctly 83.54% 80.49%
predicted
Hoslem test (p value) 0.2433 0.4674
Note t-statistics are reported under corresponding estimated coefficient in parenthesis. *** is for
the level of significance at 1%, ** is for the 5% and * is for the 10%

of switching decreased the likelihood of switching behavior by 0.699 units. Likewise,

a unit increase in attractiveness of switching (higher scores represent unattractiveness
of switching), decrease the likelihood of switching behavior by 0.509 units in Model
2.
The confirmation of the hypotheses were based on the results of Model 2-as a
more parsimonious model. The findings of Model 2 support that H2 and H3 are
confirmed, whereas the logistic regression results fail to support the H1, H4 and
H5. That is to say, when the risk of switching increases, the tendency of electricity
Understanding the Electricity Switching Behavior of Industrial Consumers … 217

switching behavior among industrial consumers decreases. A similar pattern is also

found for H3, suggesting that when the unattractiveness of switching increases, the
tendency of consumers’ electricity switching behavior decreases. To be more specific,
one might comment that as consumers consider that there are not many differences
among electricity suppliers, they perceive that switching is not attractive and thus
refrain from switching behavior.

4.4 Supplementary Analyses

For testing the robustness of the initial findings, supplementary analyses were
conducted. In this respect, the data set was divided into two subsamples of the orga-
nized industry zones (OIZ) and the other companies. This is done for two reasons.
First, a huge proportion of our data consists of OIZ companies. Second, OIZ consti-
tutes industrial consumers with high sector diversification compared to the other
industrial consumers in particular sectors. Thus, the binomial logistic regression
analysis was conducted separately for OIZ and non-OIZ companies to explore the
potential predictors of electricity switching behavior. Initially, Table 5 displays the
results of binomial logistic regression (both Model 1 and Model 2) for the sample of
OIZ companies.
As it can be seen in Table 5, the binomial logistic regression of Model 1 (where
all the variables are present within the model) is found significant with the log-
likelihood test value of −12.087. The Hoslem test was insignificant which indicates
that the model adequately fits the data. Likewise, the values of Nagelkerke, Cox-
Snell, McFadden Pseudo are reported as 0.51, 0.32, and 0.39 respectively. The overall
percentage of accuracy of our estimations is more than 87%. According to Wald’s
criterion, the attractiveness of switching is found to be significant in predicting the
likelihood of electricity switching behavior. Such that, the odds ratio indicates that
a unit increase in attractiveness of switching (higher scores represent unattractive-
ness of switching), decrease the likelihood of electricity switching behavior by 0.794
units within Model 1. Similarly, Model 2 (Stepwise binomial logistic regression) was
significant with a loglikelihood test value of −14.256. The overall percentage of accu-
racy of the estimations is reported as 82%. In line with Wald’s criterion, the factors
of risk of switching and attractiveness of switching variables are statistically found
to significant. That is to say, a unit increase in attractiveness of switching (higher
scores represent unattractiveness of switching) decreases the likelihood of switching
behavior by 0.693 units. Furthermore, a unit increase in the risk of switching, decrease
the likelihood of switching behavior by 0.643 units.
The findings reported in Model 2 for OIZ companies provide similar reports
regarding the findings of Model 2 for the pooled sample. That is to say, supplementary
analyses also confirmed the predictive utility of risk of switching and unattractiveness
of switching factors in decreasing the likelihood of electricity switching behavior.
Hence, for OIZ companies, risk of switching and unattractiveness of switching
218 M. Erdogan et al.

Table 5 Results of logistic regression with OIZ companies sample

Variables Model 1 Model 2
Coeff Wald Odds ratio Coeff Wald Odds ratio
Risk of switching −0.641 0.628 0.527 −1.030** 3.971 0.357
(−0.792) (−1.993)
Cost of switching −0.575 0.362 0.563
(−0.601)
Commercial service −1.779 0.801 0.169
quality (−0.895)
Administrative service 1.293 0.513 3.644
quality (0.716)
Communicative service −0.986 1.562 0.373
quality (−1.273)
Physical service quality 1.261 1.620 3.529
(1.250)
Unattractiveness of −1.579* 3.229 0.206 −1.181* 3.722 0.307
switching (−1.797) (−1.929)
Market competition −0.94 0.982 0.391
(−0.991)
Constant 12.359* 233,134.8 7.390*** 1618.97
(1.778) (3.044)
Observations 40 40
Log Likelihood −12.087 −14.256
Akaike Inf. Cri 42.175 34.512
McFadden pseudo 0.396 0.263
R−squared
Cox-Snell pseudo 0.327 0.219
R-squared
Nagelkerke pseudo 0.517 0.360
R-squared
Percent correctly 87.50% 82.05%
predicted
Hoslem test (p value) 0.7287 0.195
Note t-statistics are reported under corresponding estimated coefficient in parenthesis. ***is for the
level of significance at 1%, ** is for the 5% and * is for the 10%

were reported as significant factors in predicting electricity switching behavior of

consumers, thereby supporting the premises of H2 and H3.
The similar binomial logit estimations are also conducted for the Non-OIZ compa-
nies’ subsample. As can be seen in Table 6, the overall explanation of Model 1 was
satisfactory and the model is found significant with the log-likelihood test value of −
12.458. The Hoslem test was insignificant which indicates that the model adequately
fits the data. The overall percentage of accuracy of our estimations is more than
Understanding the Electricity Switching Behavior of Industrial Consumers … 219

Table 6 Results of logistic regression with Non-OIZ company subsample

Variables Model 1 Model 2
Coeff Wald Odds ratio Coeff Wald Odds ratio
Risk of switching −1.986* 3.165 0.137 −1.534** 6.9342 0.216
(−1.779) (−2.922)
Cost of switching −0.417 0.074 0.659
(−0.273)
Commercial service −0.733 0.127 0.481
quality (−0.356)
Administrative service 1.696 1.316 5.45
quality (1.147)
Communicative service −0.883 1.215 0.414
quality (−0.823)
Physical service quality 1.138 0.677 3.12 0.755* 2.7365 2.127
(1.102) (1.914)
Unattractiveness of −0.109 0.015 0.897
switching (−0.121)
Market competition −0.374 0.130 0.688
(−0.361)
Constant 4.623 101.831 3.256* 25.946
(0.727) (1.815)
Observations 39 39
Log Likelihood −12.458 −13.519
Akaike Inf. Cri 42.916 33.037
McFadden pseudo 0.321 0.263
R-squared
Cox-Snell pseudo 0.261 0.22
R-squared
Nagelkerke pseudo 0.428 0.36
R-squared
Percent correctly 84.62% 82.05%
predicted
Hoslem test (p value) 0.1624 0.195
Note t-statistics are reported under corresponding estimated coefficient in parenthesis. *** is for
the level of significance at 1%, ** is for the 5% and * is for the 10%

84.6%. According to Wald’s criterion, the only risk of switching is found as a signifi-
cant negative predictor in determining the likelihood of electricity switching behavior.
Similarly, Model 2 (Stepwise binomial logistic regression) was significant with a log-
likelihood test value of −13.519. The values of Nagelkerke, Cox-Snell, McFadden
Pseudo are reported as 0.36, 0.22, and 0.26 respectively. The percentage of the overall
accuracy of our estimates was 82.05%. According to Wald’s criterion, the risk of
switching was still significant in Model 2, suggesting that a one-unit increase in the
risk of switching factor, decrease the likelihood of switching behavior by 0.789 units.
220 M. Erdogan et al.

Likewise, physical service quality perceptions are also turned out to be a significant
positive predictor in determining the likelihood of electricity switching behavior in
Model 2 for Non-OIZ company sample. The odds ratio indicates that a unit increase
in physical service quality perceptions increases the likelihood of switching choice
by 2.12 units.

5 Discussion

The current study aimed to investigate the potential determinants of the electricity
supplier switching behavior of large scale consumers in an emerging market, Turkey.
Although some research explores the supplier switching behavior of consumers
particularly in the service industry (Wieringa & Verhoef, 2007) and energy sector
(Ek & Söderholm, 2008; Şirin & Gonul, 2016), this study contributes to the extant
literature by preliminary addressing the electricity switching behavior of large-scale
consumers rather than households. As industrial consumers represent an important
large segment of the total market, revealing the predictors of switching and not
switching actions are important in combatting with the inertia in the liberalized elec-
tricity market. Furthermore, moving beyond the rational factors, this study develops
and tests a model that incorporates switching cost, risk of switching, the attractive-
ness of switching, perceives service quality, and market competition in predicting
the switching behavior within the theoretical framework of the push–pull-mooring
effect.
The results of the current study highlight the roles of switching risk and attractive-
ness of switching as the psychological switching barriers in the Turkish electricity
market. Accordingly, when the risk of switching increases, the likelihood of switching
behavior decreases. Such that, there will be less switching behavior when eligible
industrial consumers anticipate experiencing various problems (i.e. payments, reply
on complaints, etc.) as they switch to a new electricity supplier. This finding also
supports the CEER report on European Retail Markets that includes ‘loyalty to
existing suppliers’ ‘insufficient gains available (or anticipated by consumers)’ and
‘lack of trust in the process or the players in the market’, as the barriers to switching
(CEER, 2016). Likewise, He and Reiner (2015) also pointed out the complexity
of the switching process as a major cause to remain with the current supplier and
argue that consumers perceive the switching process to be more complex than it
is. This finding also corroborates with the findings of Wieringa and Verhoef (2007)
conducted in the Dutch Market. Similarities with this previous research also exist
in terms of remarkable determinants that affect switching behavior. Switching cost
and attractiveness of switching are found to be the most important determinants of
switching behavior parallel to Wieringa and Verhoef (2007)’s findings. Nonetheless,
as in Ek and Söderholm (2008) that explore household electricity supplier behavior
in Swedish Market, psychological factors are found to have influential effects on
consumer’s switching behaviors, in addition to economic costs and benefits.
Understanding the Electricity Switching Behavior of Industrial Consumers … 221

Given that the notion of industrial consumers as rational consumers have strong
incentives to search for the optimum opportunities and switch as soon as they can
because the energy bills are the company’s significant part of operating cost, the
findings of the study highlight the opposite. Consistent with the status quo (mooring
effect), under high uncertainty and high-risk conditions, industrial consumers also
experience a cognitive bias in which they continue to stick to their prior decision and
prefer to do nothing and experience inertia (Samuelson & Zeckhauser, 1988).
Another insightful result of the study posits the attractiveness of switching as
an important predictor of switching behavior. The attractiveness of switching refers
to the perceptions of customers concerning the extent to which viable competing
alternatives are available in the market (Jones et al., 2000). According to findings
of the study, the negative association between attractiveness of switching (higher
scores indicate unattractiveness of switching) and switching behavior denotes that
when the industrial consumers perceive switching as unattractive, the tendency to
stick to their current provider increases. In other words, when consumers think that
there are not many differences between electricity suppliers in terms of their prices
and service qualities, the attractiveness of switching is perceived as low. Thus, when
the viable alternatives are lacking, or when the consumers anticipate insufficient
gains as indicated in CEER’s report (CEER, 2016), the probability of terminating the
relationship with the current electricity provider decreases. Alternatively, it could be
possible that consumers might be unaware of the potential choice of energy suppliers
and/or their differentiating products/services (Crampes & Waddams, 2017).
In the second stage of the study, we provide specific findings with respect to
the OIZ and Non-OIZ companies separately. For the companies located in OIZ, the
factors predicting switching behavior demonstrate a similar pattern for the pooled
sample. Such that, risk of switching and attractiveness of switching is reported as the
main determinants of electricity supplier change. However, for non-OIZ companies,
in addition to the risk of switching, the factor of physical service quality perceptions is
revealed as important predictors of switching behavior. As a consequence, we might
comment that OIZ and non-OIZ companies might be reactive to external signals in
different ways, thus alternative differing interventions could be designed with respect
to OIZ and Non-OIZ large scale customers.
Overall, the results of this study indicate that being emerging market experiencing
the electricity market liberalization later than other markets, the supplier switching
behavior of large-scale electricity consumers in the Turkish market is found to be
less likely relying on the fact that switching opportunities are new to the market that
may lead consumers to retain from take the risks of switching.
The current study is not without its limitations. We note that our study is cross-
sectional and exploratory but not causal. So, more research (causal or descriptive)
would need in inferring generalizations regarding how factors influence switching
behavior in Turkish markets after liberalization.
As a concluding remark, the findings of the current chapter might shed light
on the problems that deter electricity consumers from switching in the market. In
terms of practical considerations, policymakers might design several implications or
222 M. Erdogan et al.

interventions to encourage large-scale consumers to switch (e.g. offering incentives)

and then to promoting demand response.

Acknowledgements We gratefully acknowledge supports from the Republic of Turkey Ministry

of Industry and Technology, Energy Market Regulation Authority, and Association of Distribution
System Operators. We would like to express our gratitude to The Scientific and Technological
Research Council of Turkey for the grant 119K076.

Appendix 1: Descriptive Statistics

Name Number of observation Mean Standard dev.

How many years have you been 80 30.01 17.66
operating in this sector?
Have you ever switched your 82 0.807 0.39
electricity supplier
How many times have you switched 68 4
your electricity supplier?
Which sector does your business serve?
Organized industrial zone 40
Cement manufacturer 6
Plastic 7
Steel producer 10
Construction 5
Others 15
How many employees firm has?
0–100 38
100–250 12
250–1000 14
1000–2000 4
2000 and above 14
What is your annual profit?
0 TL–500,000 TL –
1,000,000 TL–3,000,000 TL –
3,000,000 TL–10,000,000 TL 6
10,000,000 TL–25,000,000 TL 5
25,000,000 TL–125,000,000 TL 11
125,000,000 TL and above 50
(continued)
Understanding the Electricity Switching Behavior of Industrial Consumers … 223

(continued)
Name Number of observation Mean Standard dev.
No answer 11
What is the share of electricity cost to your overall cost?
4% 14
5–9% 8
10–19% 9
20–29% 12
30–39% 19
No answer 9

References

Athanassopoulos, A. D. (2000). Customer satisfaction cues to support market segmentation and

explain switching behavior. Journal of Business Research, 47(3), 191–207.
Bansal, H. S., & Taylor, S. F. (1999). The service provider switching model (SPSM) a model of
consumer switching behavior in the services industry. Journal of Service Research, 2(2), 200–218.
Chang, H. H., Wong, K. H., & Li, S. Y. (2017). Applying push-pull-mooring to investigate channel
switching behaviors: M-shopping self-efficacy and switching costs as moderators. Electronic
Commerce Research and Applications, 24, 50–67.
Caceres, R. C., & Paparoidamis, N. G. (2007). Service quality, relationship satisfaction, trust,
commitment, and business-to-business loyalty. European Journal of Marketing, 41(7/8), 836–867.
CEER. (2016). Market monitoring report 2015, November. http://www.
acer.europa.eu/official_documents/acts_of_the_agency/publication/acer%20m
arket%20monitoring%20report%202015%20-%20electricity.pdf
CEER. (2018). Monitoring report on the performance of European retail markets in 2018. CEER
report monitoring retail markets WS of customers and retail markets WG.
Colgate, M., & Lang, B. (2001). Switching barriers in consumer markets: An investigation of the
financial services industry. Journal of Consumer Marketing, 18(4), 332–347.
Concettini, S., & Créti, A. (2013). Liberalization of electricity retailing in Europe: Coming back or
going forth? CSSD Working Paper.
Crampes, C., & Waddams, C. (2017). Empowering electricity consumers in retail and wholesale
markets. CERRE Report.
Dowling, G. R., & Staelin, R. (1994). A model of perceived risk and intended risk-handling activity.
Journal of Consumer Research, 21(1), 119–134.
Eisenberger, R., Lynch, P., Aselage, J., & Rohdieck, S. (2004). Who takes the most revenge? Indi-
vidual differences in negative reciprocity norm endorsement. Personality and Social Psychology
Bulletin, 30(6), 787–799.
Ek, K., & Söderholm, P. (2008). Norms and economic motivation in the Swedish green electricity
market. Ecological Economics, 68(1–2), 169–182.
Gamble, A., Juliusson, E. A., & Gärling, T. (2009). Consumers’ attitudes towards switching suppliers
in three deregulated markets. Journal of Socio-Economics, 39, 814–819.
Garson, D. (2010). Logistic regression, from Statnotes: Topics in multivariate analysis.
Giulietti, M., Price, C. W., & Waterson, M. (2005). Consumer choice and competition policy: A
study of UK energy markets. The Economic Journal, 115(506), 949–968.
224 M. Erdogan et al.

He, X., & Reiner, D. (2015). Why do more British consumers not switch energy suppliers? The
role of individual attitudes. EPRG Working Paper 1515 Cambridge Working Paper in Economics
1525.
Jones, M. A., Beatty, S. E., & Mothersbaugh, D. V. (2002). Why customers stay: Measuring
the underlying dimensions of services switching costs and managing their differential strategic
outcomes. Journal of Business Research, 55, 441–450.
Jones, M. A., Mothersbaugh, D. L., & Beatty, S. E. (2000). Switching barriers and repurchase
intentions in services. Journal of Retailing, 76(2), 259–274.
Karan, M. B., & Kazdağli, H. (2011). The development of energy markets in Europe. In Financial
aspects in energy (pp. 11–32). Springer.
Keaveney, S. M. (1995). Customer switching behavior in service industries: An exploratory study.
Journal of Marketing, 59(2), 71–82.
Klemperer, P. (1995). Competition when consumers have switching costs: An overview with applica-
tions to industrial organization, macroeconomics, and international trade. The Review of Economic
Studies, 62(4), 515–539.
Lien, C. H., Cao, Y., & Zhou, X. (2017). Service quality, satisfaction, stickiness, and usage intentions:
An exploratory evaluation in the context of WeChat services. Computers in Human Behavior, 68,
403–410.
Nebel, J. M. (2015). Status quo bias, rationality, and conservatism about value. Ethics, 125(2),
449–476.
Nunnally, J. C., & Bernstein, I. H. (1994). Psychometric theory (3rd ed.). McGraw-Hill.
Parasuraman, A., Zeithaml, V. A., & Berry, L. L. (1985). A conceptual model of service quality and
its implications for future research. Journal of Marketing, 49(4), 41–50.
Patterson, P. G., & Smith, T. (2003). A cross-cultural study of switching barriers and propensity to
stay with service providers. Journal of Retailing, 79(2), 107–120.
Porter, E. M. (1980). Competitive strategy. Academic Press.
Poudineh, R. (2019). Liberalized retail electricity markets: What we have learned after two decades
of experience? https://ora.ox.ac.uk/objects/uuid:7f3578a6-27b8-47d0-b2f2-c3aeff2b9053
Samuelson, W., & Zeckhauser, R. (1988). Status quo bias in decision making. Journal of Risk and
Uncertainty, 1(1), 7–59.
Sharma, N., & Patterson, P. G. (2000). Switching costs, alternative attractiveness, and experience as
moderators of relationship commitment in professional, consumer services. International Journal
of Service Industry Management, 11(5), 470–490.
Shin, K. J., & Managi, S. (2017). Liberalization of a retail electricity market: Consumer satisfaction
and household switching behavior in Japan. Energy Policy, 110, 675–685.
Sirin, S. M., & Gonul, M. S. (2016). Behavioral aspects of regulation: A discussion on switching
and demand response in the Turkish electricity market. Energy Policy, 97, 591–602.
Tabachnick, B. G., & Fidell, L. S. (2001). Using multivariate statistics. Allyn and Bacon.
Waterson, M. (2003). The role of consumers in competition and competition policy. International
Journal of Industrial Organization, 21(2), 129–150.
Wieringa, J. E., & Verhoef, P. C. (2007). Understanding customer switching behavior in a liberalizing
service market: An exploratory study. Journal of Service Research, 10(2), 174–186.
World Energy Council. (2018). Dünya Enerji Konseyi Türkiye Türkiye Enerji Piyasasının Geleceği
Çalışma Grubu Tepg1 –Türkiye Enerji Piyasalari Araştirma Raporu. https://www.dunyaenerji.
org.tr/wpcontent/uploads/2018/07/TEPG1pdf
Yang, Y. (2014). Understanding household switching behavior in the retail electricity market. Energy
Policy, 69, 406–414.
Yee, R. W., Yeung, A. C., & Cheng, T. E. (2010). An empirical study of employee loyalty,
service quality, and firm performance in the service industry. International Journal of Production
Economics, 124(1), 109–120.
https://www.enerji.gov.tr/en-US/Pages/Electricity. Ministry of Energy and Natural Sources. https://
www.pwc.com.tr/tr/publications/industrial/energy/assets/renewable-report-11-april-2012.pdf
Does the Market Value Clean
Innovation? Evidence from US Listed
Firms

Antoine Dechezleprêtre, Cal B. Muckley, and Parvati Neelakantan

1 Introduction

It is widely acknowledged that a rapid and radical change in the mix of technologies
used to produce and consume energy is needed to reduce harmful greenhouse gas
emissions (IPCC, 2014). Technological change is also undoubtedly one of the keys
to ensuring that climate change can be addressed without compromising economic
growth. This, in turn, requires that climate and innovation policies provide an eco-
nomic incentive for firms to redirect innovation away from dirty and towards clean
technologies. Whether such an incentive exists or is emerging for United States-based
firms is an important and challenging research question.
This study brings new insights to the corporate environmental–financial perfor-
mance debate by examining the value that capital markets accord to low-carbon
(‘clean’) and fossil fuel (‘dirty’) innovation in the United States, based on a large
data set of 2526 US listed firms and their patent applications to the US patent office
across 18 years, from 1995 to 2012. Informed by the seminal literature which accords

We are grateful to Daragh Ryan for excellent research assistance. Cal Muckley would like to
acknowledge the financial support of Science Foundation Ireland under Grant Number 16/SPP/3347
and 17/SP/5447.

A. Dechezleprêtre
OECD, Science, Technology and Innovation Directorate, Grantham Research Institute
on Climate Change and the Environment, London School of Economics, London, UK
e-mail: [email protected]
C. B. Muckley (B)
UCD School of Business and Geary Institute, University College Dublin, Dublin, Ireland
e-mail: [email protected]
P. Neelakantan
UCD School of Business, University College Dublin, Dublin, Ireland
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 225
A. B. Dorsman et al. (eds.), Applied Operations Research and Financial Modelling
in Energy, https://doi.org/10.1007/978-3-030-84981-8_11
226 A. Dechezleprêtre et al.

a market evaluation to firm-level eco-efficiency (e.g., Guenster et al. (2011)) and

knowledge stock (e.g., Hirshleifer et al. (2013)), we disaggregate innovation mea-
surements (e.g., Deng et al. (1999) and Gu (2005)) of US firms’ patent stock into
clean, dirty and other innovation. We, then, elicit their market evaluations over time.
The extant literature has focused on specific factors incentivizing clean innovation
(e.g., Aghion et al. (2016), Calel and Dechezleprêtre (2016), Popp (2002)). These
studies bring evidence that change in policies and market conditions can induce firms
to innovate in environment-friendly innovation, while discouraging them to innovate
in fossil-based innovation. For instance, Popp (2002) notes that a raise in energy
prices induce firms to substantially increase their production of energy-saving tech-
nologies. Similarly, Calel and Dechezleprêtre (2016) find that the introduction of the
European Union Emission Trading System caused regulated firms to increase their
patenting activity in low-carbon technology by 30%. However, a plethora of fac-
tors determine a company’s decision to produce clean or dirty technologies, includ-
ing relative production factors’ prices, public policies, and innovation capabilities
(Aghion et al., 2016; Popp, 2002). In addition, multiple public policies coexist in any
given jurisdiction, with some incentivizing clean innovation (e.g., carbon markets,
fuel taxes) and some discouraging it (e.g., fossil-fuel subsidies). Hence, the existing
literature does not shed light on the cumulative impact of the various factors that
incentivize or disincentivize clean innovation on the firms’ decision to produce clean
or dirty technologies. Additionally, a firm’s decision to invest in the production of
clean technology rests on the expected environmental policy stringency in the future,
which is unobserved to researchers.
In this study, we treat capital markets as a litmus test to assess whether there is a net
incentive or disincentive for corporations to innovate in an environmentally friendly
way. The capital market can give a summative signal to account for the complex
combination of factors which can yield an incentive or a disincentive for firm-level
innovation in radically new clean technologies or in improved dirty technologies.
To model the capital market’s response to firm-level innovation in clean and dirty
technologies, we construct objective measures of innovation productivity variables
which are inspired by prior literature (Chan et al., 2001; Deng et al., 1999; Gu, 2005).
Our chief contribution consists in eliciting capital market evaluations associated
with the disaggregated Deng et al. (1999) and Gu (2005) innovation productivity
measure to account for ‘clean’ and ‘dirty’ innovation production. In this analysis, we
refer to clean technologies as patented technologies in renewable energy generation,
electric vehicles, and energy efficiency technologies in the building sector. We refer
to dirty innovation as patented technologies in the fossil-fuel based energy generation
and ground transportation. Our study is the first to investigate, over time, the value
accorded by the capital markets to low-carbon (‘clean’) and fossil fuel (‘dirty’)
innovation in the United States.
Our baseline results are as follows. We find that the capital market accords a
1.20% higher valuation to the firms producing environment-friendly innovation and
decreases the market value of firms producing fossil-based technologies to the tune
of 0.45%, on average. In more demanding specifications including a range of firm-
Does the Market Value Clean Innovation? Evidence from US … 227

level controls, this negative association between fossil-based innovation and market
value becomes statistically insignificant but the clean innovation premium remains.
We further investigate whether clean innovation premium holds over time. We
divide our sample period into two sub-periods: 1995–2005 and from 2006–2012.
The first sub-period was characterized by low environmental policy stringency in the
U.S., while the later period saw a sharp increase in environmental policy stringency.
We note an increase in the association between clean innovation productivity and
firms’ market value across these periods and observe that only during the 2006–2012
period do we find a statistically and economically significant positive association. In
this period, the market value of firms producing environment-friendly technologies
increases by 1.3%. We hypothesize that the observed increase in environmental policy
stringency starting in 2005 may explain the statistically significant clean innovation
premium.

2 Hypothesis and Literature Review

Several IPCC reports claim that stabilizing global carbon emissions by 2050 requires
an 80% reduction in global greenhouse gas emissions. This motivates our investi-
gation of the existence of capital market incentives to decarbonize the US economy
through a radical change in the mix of technologies used to produce and consume
energy.
Hart (1995) distinguishes two ways to reduce pollution: a short-term, “end-of-
pipe” approach and a more long-term prevention approach which emphasizes source
reduction and process innovation. Prevention, in the early stages, can be achieved
by incorporating emission reducing methods in total quality management programs
(Makower, 1994). Such initiatives can improve utilization of “end-of-pipe” devices
in addition to helping firms in using the “end-of-pipe” control devices productively
and efficiently (Schmidheiny & Timberlake, 1992). However, to drastically reduce
emissions, firms are required to produce environment-friendly technologies (Hart &
Ahuja, 1996).
In light of the resource-based theory of firms, we hypothesize that firms inno-
vating in clean technologies and protecting these innovations through patents are
valued higher than firms having patented dirty technologies. The resource-based
theory of competitive advantage posits that a firm’s inimitable assets gives the firm
a sustainable competitive advantage over other firms (Hart, 1995). However, if an
asset/technology is available to other competing firms, then, such a technology cannot
generate premium profits (Russo & Fouts, 1997).
Prior studies have analyzed the association between environmental and financial
performance; however, these studies report two countervailing relationships between
investments in environmental innovation and financial performance.
228 A. Dechezleprêtre et al.

2.1 Environment-Friendly Innovation Can Improve a Firm’s

Financial Performance

Pro-active environment-friendly innovation by firms, being a measure of firms’ envi-

ronmental performance, prompts investors to associate these firms with lower risk
(Shane & Spicer, 1983). Such innovations can help firms to sidestep risk of losses
from crises or change in regulation (Reinhardt, 1999) and expenses due to lawsuits
and legal settlements (Karpoff et al., 2005). This leads investors to assign a lower
discount rate to firms innovating in clean technologies, resulting in higher market val-
uation. Additionally, firms with good credentials in environmental performance can
draw investments from ethical investors (Heinkel et al., 2001). Investments by ethical
investors indicate a firm’s resolve to produce environment-friendly technologies that
helps lower the cost of capital for these firms when they raise capital in the secondary
markets. Superior environmental performance can also improve firms’ operational
efficiency (Porter & Van der Linde, 1995), credibility in existing markets (Klassen &
McLaughlin, 1996), and opportunities to enter new markets (Porter & Van der Linde,
1995) as a result of better corporate and brand reputation with regulators, employees
and the public (Russo & Fouts, 1997). Hence, an increase in expected cash-flows and
reduction in risk for firms can lead capital markets to evaluate environment-friendly
firms positively.
Firms producing clean technologies can also be required to fundamentally trans-
form their culture and human resources (Hart, 1995). Firms having better environ-
mental performance are likely to have highly-skilled management (Bowman & Haire,
1975) and better labour conditions which could attract high-quality employees (Tur-
ban & Greening, 1997) and boost employee morale and productivity (Dowell et al.,
2000). Employees in firms committed to strong corporate social responsibility are
often more willing to accept lower wages and yet be highly productive as a result
of efforts towards promoting an environment-friendly corporate culture (Nyborg &
Zhang, 2013).

2.2 Environment-Friendly Innovation Can Deteriorate a

Firm’s Financial Performance

Some studies show that corporate investment in environmental innovation can lead
to a deterioration of a firm’s financial performance. These studies argue that pro-
duction of environment-friendly technologies and adoption of practices to mitigate
greenhouse gas emissions raise costs for firms, causing them to increase the prices
of their products, which cuts into their profits (Palmer et al., 1995). Fisher-Vanden
and Thorburn (2011) also find clean innovation to be associated with negative stock
market valuation. They note that voluntarily choosing to produce clean technologies
deteriorates the firms’ market value, suggesting that investors expect environmental
investments to raise costs for the firms.
Does the Market Value Clean Innovation? Evidence from US … 229

To sum up, the literature is inconclusive as to the expected direction of the rela-
tionship between clean innovation and firm market value. Therefore, this question
remains empirically open. In addition, we note several methodological shortcom-
ings in the prior literature which may help to explain these conflicting results. These
include small samples, subjective environmental performance measures, and the use
of single industry samples. In this study, we seek to overcome these methodological
shortcomings by examining a large sample of firms across industries, employing
objective measures of environmental performance, and accounting for differences in
industries.

3 Data and Variables

In this section, we present our sample of firms, the firm-level patent and citation
data, and discuss the key variables. We also distinguish between our clean and dirty
patenting criteria and present the descriptive statistics of our variables.

3.1 Sample of Firms and Firm-Level Patent Data

Our analysis draws firm-level data from the Worldscope database, which comprises
firm-level data for the largest global firms, and firm-level patent data from World
Patent Statistical Database (PATSTAT), which is the most comprehensive global
patent database available.
Sourced from Worldscope, our firm-level dataset pertaining to the period 1995–
2012 annually encompasses financial and accounting information of listed firms that
are headquartered in the United States. We search the database for both active and
delisted firms to avoid survivorship bias and also select firms with usable International
Securities Identifying Number (ISIN) and industry codes. From this sample of firms,
we eliminate firms whose ISIN is not traceable. We then eliminate firms with negative
total assets, market capitalization or common cash dividend paid.1 Further, we drop
firms that have less than 5 consecutive firm-year observations between 1995 and
2012 across all the firm-level variables used in our regressions—year-end market
capitalisation, total assets, total debt, net sales, capital and advertising expenditure,
and earnings before interest and tax. This gives us a sample of 2526 firms with their
headquarters in the United States. For these firms, we then obtain stock-price data
from the Datastream database.

1 Worldscope database provides accounting data for the world’s largest companies. In our firm-level
dataset from Worldscope, we noticed that a few firms had negative total assets, market capitalization
or common cash dividend paid. Since these values cannot be negative, we eliminated those firms
since the discrepancies in their registered values flow from the database itself. Including such firms
would be erroneous and misguide our study.
230 A. Dechezleprêtre et al.

We obtain our firm-level patent data from the PATSTAT database maintained by
the European Patent Office. For each patent, we observe the application date, the
publication date and, if the patent was ever granted by the United States Patent and
Trademark office (USPTO), the grant publication date. PATSTAT also provides cita-
tion counts corresponding to each patent for a given year. From the patent dataset
we disaggregate patents into ‘clean’, ‘dirty’ and ‘other’ technology classes. Previous
work by the OECD Environmental Directorate2 informs our choice of patent classi-
fication codes for determining a patent as clean. Specifically, the patent classification
codes for clean innovation encompasses patenting activities related to renewable and
non-fossil sources of energy generation (hydro, solar, wind, geothermal, biomass,
marine, and energy from waste), technologies having the potential to mitigate car-
bon emissions and/or improve energy storage, in particular, energy conservation in
buildings, innovations in electric and hybrid automobiles, and combustion technolo-
gies with carbon mitigation potential (e.g., combined heat and power). Our choice of
patent classification codes for dirty innovation is informed by seminal papers such as
Noailly and Smeets (2015) and Aghion et al. (2016). The patent classification codes
for dirty innovation pertains to fossil-based electricity generation and transportation
technologies (combustion engines, combustion apparatus and furnaces, steam engine
plants, gas turbine plants, and steam generation).3
To match our firm-level dataset with firm-level patent dataset we adopt the match-
ing algorithm developed by Bureau van Dijk, which is based on matching both name
and geographical location (country, address, etc.). This algorithm ensures that each
firm in our sample is matched with its corresponding patents and citations from
the patent dataset. Our final matched dataset is also subjected to rigorous manual
checking to further ensure the credibility of our dataset.
Since our study examines, over time, the value accorded by the capital markets
to low-carbon (‘clean’) and fossil fuel (‘dirty’) innovation in the United States, we
choose the patent publication date when constructing our key innovation variables.
Our choice is guided by the reasonable assumption that the capital market participants
would be aware of the new patents by the patent publication date.

3.2 Limitations of Using Patent Measures as a Proxy for

Innovation

A large recent empirical literature uses patent measures to proxy for technological
change at the firm level. Although a majority of the most economically significant
innovations are patented (Khan & Dernis, 2006), Frietsch and Schmoch (2006) note
that patents are only one among several appropriability mechanisms to protect an
innovation, including deliberately making complex specifications and employing

2See www.oecd/environment/innovation.
3The patent classification codes used to extract clean and dirty patents from the database are
presented in Tables 8 and 9 in the Internet Appendix 1.
Does the Market Value Clean Innovation? Evidence from US … 231

secrecy. Additionally, the practice of filing a patent and its valuation varies markedly
across different jurisdictions, industries and over time (Griliches, 1998). Therefore,
patent figures may not entirely capture a firm’s level of innovativeness, making them
a noisy proxy for a firm’s technology production.
While previous empirical studies using patent counts to indicate innovation have
largely ignored differences across industries, our work accounts for this concern by
employing adjusted patent citations and industry fixed effects in the econometric
models. It is widely accepted that citations of a firm’s patents indicate the techno-
logical and economic significance of the innovation (Hall et al., 2005; Hirshleifer et
al., 2013). We follow Hirshleifer et al. (2013), Pandit et al. (2011), and Gu (2005),
to construct our measures of citation productivity. In following them, we ensure
that our citation productivity measures are observable to the investors while making
their investment decisions.4 Further, our measures of citation productivity consisting
in citation of the year, technological fields, and grant year account for the cita-
tion propensity. Further, to alleviate the problem of comparing innovations across
different industries, we include industry fixed effects in our econometric models
(Griliches, 1998). Finally, since we have a large dataset (2526 firms), we invoke the
law of large numbers to partly address the issue of heterogeneity in patent values,
besides employing citation measures (Griliches, 1998).

3.3 Key Variables of Interest

In this sub-section, we present the variable descriptions of our dependent variable

(Tobin’s Q), our innovation productivity measures, (RDBE, Pat/Book, Cit/Book,
Cit_clean/Book, Cit_dirty/Book, Cit_other/Book) and the firm traits
employed in our study (Total_assets, Debt/Assets, Salesgrowth_2yr, EBIT ,
invBE, CEME, ADME, taxRDBE). The list of all the variable definitions can be
found in Table 1.

3.3.1 Dependent Variable

To investigate whether capital markets recognize the value of clean innovation, we test
the contemporaneous association between our innovation variables with the Tobin’s
Q, controlling for firm traits.
The variable construction of the dependent variable, Tobin’s Q, is in line with
prior literature and is defined as the market value of the firm to its replacement cost:

4 This is crucial for a meaningful analysis as we seek to investigate, from a market information
assimilation perspective whether there is an incentive for firms to pursue strategies of clean envi-
ronmentally supportive innovation, as opposed to carbon-emitting dirty innovation activities in the
United States.
232 A. Dechezleprêtre et al.

Table 1 Variable definitions

Variable Definition
Measures of firm value
Tobin’s Q Market value of the firm to the book value of tangible assets
(Total_assets − Book + Market_Value)/(Total_assets)
Total_assets (millions of $) Total assets represents the sum of total current assets, long term
receivables, investment in unconsolidated subsidiaries, other
investments, net property plant and equipment and other assets
Market value Total market value of the company based on year end price and
number of shares outstanding converted to U.S. dollars using the
year end exchange rate
Book (millions of $) Book value of equity
Measure of R&D productivity
RDBE Research and Development expense divided by Book
Measures of innovation productivity
Pat/Book Number of US patents of the firm, in any patent category, divided
by Book
Cit/Book The numerator is the number of citations received in year t by US
patent k, granted in year t − j (j = 1–5) scaled by the average
number of citations received in year t by all patents of the same
subcategory granted in year t − j (j = 1–5). This number is summed
over the total number of patents granted in year t − j to firm i. The
numerator is divided by the book value of equity
Cit*/Book As per Cit/Book but US patent category is *: clean, dirty and other
technologies
Firm traits
Debt/Assets Total debt (sum of long and short term debt; in millions of $) over
Total_assets (millions of $)
NetSales
Sales growth2ye Past two-year sales growth (( NetSalesi,t−2
i,t
− 1) ∗ 100); Net Sales in
millions of U.S. $
EBIT (millions of $) Earnings before Interest and Taxes (pretax income plus interest
expense on debt minus interest capitalized)
invBE Inverse of Book
CEME Capital expenditure (funds used to acquire fixed assets other than
those associated with acquisitions) to Market Value of Equity
ADME Advertising expenditure to Market Value of Equity
taxRDBE Tax shelter associated with R&D expenditure
Regulation
EPS Environmental Policy Stringency Index (Botta & Kozluk, 2014);
This index takes the value from 0 (least stringent) to 6 (most
stringent) and is a country-specific stringency measure
Does the Market Value Clean Innovation? Evidence from US … 233

Total assets − Book valueequity + Market Value

Tobin s Q = (1)
Total assets
Consistent with the interpretation of Tobin’s Q in prior literature (Hall et al., 2005),
we refer to Tobin’s Q as the ‘market value’ of the firm. In our econometric models
we use the natural logarithm of Tobin’s Q as the dependent variable and we account
for the past two-year sales growth and R&D expenses, since recent sales growth and
R&D expenses are positively correlated with the firm’s valuation (Guenster et al.,
2011).

3.3.2 Innovation Productivity Variables

Our key measures of a firm’s innovation productivity are associated with its patent
filings and related citations, in a given year. These measures are inspired by prior
literature (Deng et al., 1999; Gu, 2005) and explain how successfully a firm obtains
patents and whether these innovations are economically significant.
We employ patents over book value of equity (Pat/Book) as a proxy for a firm’s
innovativeness, which is defined as the ratio of firm i’s patents (published by the
USPTO) in year t scaled by the book value of equity (Deng et al., 1999):
Pati,t Patentsi,t
= (2)
Booki,t Book valueequityi,t

However, consistent with our line of reasoning, we acknowledge that patents are a
noisy measure of R&D success and, therefore, we use a citation measure to account
for patents’ commercial value. Our citation measure, inspired by Gu (2005), ensures
that the citations count is observable to capital markets’ participants when making
their investment decisions. Specifically, we construct the citation measure as:
5 Nt−j t−j
Citi,t j=1 k=1 Citationsi,k
= (3)
Booki,t Book valueequityi,t

t−j
In the above equation, Citationsi,k denotes the citation count in year t for patent
k granted in year t − j, scaled by the average citation count in the same year for all
patents belonging to the same subcategory and granted in year t − j. Nt−j represents
the total number of patents granted to firm i in the year t − j. To construct the
firm-level adjusted patent citations (numerator), we take citations related to patents
granted in the previous five years since in most industries a 5-year technology cycle
is observed with respect to the benefits of R&D spending. Our measure of adjusted
patent citations, in line with Gu (2005), Hirshleifer et al. (2013) and Pandit et al.
(2011) accounts for the average citation propensity at the citation year, technological
field, and grant year levels.
Since we seek to investigate whether US firms have sufficient incentives to invest
in clean technologies, we disaggregate our citation productivity measures into three
234 A. Dechezleprêtre et al.

categories: clean, dirty, and other. For example, the clean citation productivity vari-
able, Cit_clean/Book, corresponds to the adjusted clean patent citations over book
value of equity.
The prior literature has shown the existence of a positive association between
R&D spending and Tobin’s Q, and, therefore, we control for R&D spending using
the variable RDBE. It is defined as a firm’s R&D spending scaled by its book value
of equity (Chan et al., 2001),
R&Di,t
RDBEi,t = (4)
Book valueequityi,t

This measure of R&D spending accounts for the proportion of a firm’s book value
of equity spent on Research and Development and indicates a firm’s propensity to
produce novel innovations.
In this work, we primarily seek to study whether capital markets value the impor-
tance of clean innovation, measured by Cit_clean/Book, relative to dirty innovation,
measured by Cit_dirty/Book, controlling for aggregate R&D and patent productivity,
and firm traits.

3.3.3 Firm Traits

In our econometric models, we control firm traits that can influence the capital market
valuation of a firm. Our adopted set of firm traits are inspired by Hirshleifer et al.
(2013) and Guenster et al. (2011) and control for the firm’s size, level of risk, recent
sales growth, inverse of book value of equity, capital expenditure, advertisement
expenditure, earnings before interest and taxes, and tax shelter associated with R&D
expenditure.
To control for firm size, we use the natural logarithm of total assets (Total_assets),
and we use the debt-to-asset ratio to measure a firm’s level of risk (Debt/Assets).
To account for recent sales growth, we calculate the past two-year sales growth for
each firm. We account for capital and advertisement expenditure (CEME and ADME,
respectively) as prior literature finds an association between these variables and a
firm’s operating performance (Pandit et al., 2011).
Further, we include industry fixed effects in all our econometric models. To gen-
erate industry indicator variables, we use the 48 Fama-French industry classification
codes obtained from Kenneth R. French’s website.5

5 https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html.
Does the Market Value Clean Innovation? Evidence from US … 235

Fig. 1 Environmental Policy Stringency score for the US during the period 1995–2012. The figure
shows, over time, the Environmental Policy Stringency score (Botta & Kozluk, 2014) for the United
States during the period 1995–2012

3.4 Descriptive Statistics

Figure 1 shows the evolution of environmental policy stringency in the United States
during 1995–2012. This is the economy-wide environmental policy stringency (EPS)
index constructed in Botta and Koźluk (2014).6 The figure shows a sharp tightening
in the EPS in the US starting in 2005.
This leads us to hypothesize that there is a positive association between our clean
innovation citation productivity variable with the contemporaneous Tobin’s Q and
that there is either a negative or no association between dirty innovation citation
productivity variable with contemporaneous Tobin’s Q, particularly starting in 2005
when environmental policy became much stricter.
We present the descriptive statistics of the variables used in this study in Table 2.
For our sample of firms, we observe that, on an average, US firms have allocated only
8.9% of the book value of equity to R&D during 1995–2012. We also note that clean
and dirty citation productivity is a small fraction of total citation productivity. The
clean citation productivity accounts for only 4.44% of total citation productivity and
dirty citation productivity accounts for only 0.54% of the total citation productivity
during 1995–2012.7

6 We have used the EPS index constructed in Botta and Kozluk (2014) since this index highly and
significantly correlates with alternative measures of EPS used in the literature.
7 On an average, US firms allocate 8.9% (0.0888 ∗ 100) of the book value of equity to R&D during

1995–2012. The clean and dirty citation productivity account for 4.44% ((0.0107/0.2410) ∗ 100)
and 0.54% ((0.0013/0.2410) ∗ 100) of total citation productivity during 1995–2012.
236 A. Dechezleprêtre et al.

Table 2 Summary statistics

Full sample
Variable N Mean SD
Dependent variable
Tobin’s Q 35,408 2.2240 3.5560
Innovation productivity variables
RDBE 35,408 0.0888 0.3790
Pat/Book 35,428 0.0394 0.3260
Cit/Book 35,428 0.2410 2.6110
Cit_clean/Book 35,428 0.0107 0.3270
Cit_dirty/Book 35,428 0.0013 0.0359
Cit_other/Book 35,428 0.2290 2.5630
Firm traits
Total_assets 39,307 5,652 23,739
Debt/Assets 39,086 22.200 21.310
Sales growth2yr 33,781 22.040 59.990
EBIT 37,880 52,895 248,903
invBE 45,468 −0.0024 9.1480
CEME 36,960 13.330 340.40
ADME 45,468 0.3900 6.6210
taxRDBE 45,468 0.0088 0.2240
(a) Panel A: Summary statistics of variables during the period 1995–2012
Subperiod: 1995–2005 Subperiod: 2006–2012
Variable N Mean SD N Mean SD
Dependent variable
Tobin’s Q 19,660 0.6118006 0.6394763 15,742 0.4803632 0.533674
Innovation productivity variables
RDBE 19,665 0.0831245 0.3527756 15,743 0.0958653 0.409849
Pat/Book 19,682 0.042123 0.3260576 15,746 0.0360227 0.3258815
Cit/Book 19,682 0.2267093 2.895265 15,746 0.2589683 2.205336
Cit_clean/Book 19,682 0.0113527 0.4002653 15,746 0.0098721 0.201913
Cit_dirty/Book 19,682 0.0016531 0.0457812 15,746 0.0009423 0.0167454
Cit_other/Book 19,682 0.2137035 2.840315 15,746 0.2481539 2.16698
Firm traits
Total_assets 22,435 4318.448 18260.69 16,872 7426.025 29394.78
Debt/Assets 22,250 22.56543 21.06775 16,836 21.72419 21.62235
Sales growth2yr 17,310 26.37418 64.11276 16,471 17.48939 54.95599
EBIT 21,444 39544.04 179081 16,436 70314.51 316874.4
invBE 27,786 0.0052971 11.38873 17,682 −0.0145058 3.374483
CEME 20,377 12.29706 198.8624 16,583 14.59129 457.8233
ADME 27,786 0.3084896 6.383881 17,682 0.518257 6.976617
taxRDBE 27,786 0.0084626 0.1403147 17,682 0.0094046 0.3133699
Panel B: Summary statistics of variables during the sub periods 1995–2005 and 2006–2012
Notes The table presents summary statistics for Tobin’s Q, innovation productivity variables (RDBE,
Pat/Book, Cit_clean/Book, Cit_dirty/Book, Cit_other/Book), and variables controlling for firm traits
(Guenster et al., 2011; Hirshleifer et al., 2013) during the period 1995–2012, in Panel A, and during
the sub periods 1995–2005 (Subperiod 1) and 2006–2012 (Subperiod 2), in Panel B
Does the Market Value Clean Innovation? Evidence from US … 237

4 Methodology

From the perspective of an efficient markets argument which posits that stock prices
incorporate all the available information (Fama, 1991), the capital market provides a
summative signal of the complex combination of factors which incentivize innovation
either in radically new clean technologies or in improved dirty technologies (Chan
et al., 2001; Dechezleprêtre et al., 2020; Hall et al., 2005; Hirshleifer et al., 2013).
Innovation productivity variables comprising an R&D measure, a patent measure and
a citation measure in generating returns over a number of years meet our need for a
forward-looking and market-based measure to determine the association between a
firm’s environmental and financial performance (Grandi et al., 2009). Additionally,
our rationale for employing a forward-looking and market-based measure also relies
on recent theoretical and empirical literature which proves that managers’ decisions
on R&D investments are influenced by the stock market reactions to innovations
(Benner, 2007; Munari et al., 2005).
To determine the association between the market value of a firm and its clean and
dirty innovation productivity variables, we decompose firms’ market value into their
tangible and intangible assets. Our approach in decomposing the market value of
firms is in line with prior literature (Dechezleprêtre et al., 2020; Hall et al., 2005).
Prior studies focused on different aspects of firm valuation, such as monopoly power,
research and development investment, advertising, and brand equity. We contribute
to this literature by investigating the role of clean and dirty innovation productivity
on market value.
Following prior literature, we decompose the market value, MV, of the firm i at a
time t into tangible assets, VT , and intangible assets, VI ,

MVi,t = VT ,i,t + VI ,i,t (5)

Dividing both sides of the above equation by VT we obtain,

MVi,t VI ,i,t
=1+ (6)
VT ,i,t VT ,i,t

Since investors do not observe the value of tangible assets, VT , we measure the value
of tangible assets, VT , as the replacement cost of tangible assets. To estimate the
replacement cost of tangible assets we use accounting-based values for the firm’s
assets. This approximation leads to an equation of Tobin’s Q, qi,t ,

MVi,t VI ,i,t
qi,t = =1+ (7)
VT ,i,t VT ,i,t

No consensus exists in the literature concerning a specific functional form for an

equation to estimate Tobin’s Q. We, therefore, estimate a semi-log specification
following Guenster et al. (2011):
238 A. Dechezleprêtre et al.

ln(qi,t ) = α + βX +  (8)

To estimate the association between clean and dirty innovation productivity vari-
ables and the Tobin’s Q, we estimate our semi-log specifications, where the depen-
dent variable is the natural logarithm of Tobin’s Q and X is a matrix embedding the
explanatory and control variables. Our chief contribution consists in eliciting capi-
tal market evaluations associated with the disaggregated Deng et al. (1999) and Gu
(2005) innovation productivity measure to account for ‘clean’ and ‘dirty’ innovation
production. We estimate all our models using the Fama-Macbeth two stage (Fama
& MacBeth, 1973) estimator.

5 Empirical Analysis

To examine whether the capital markets recognize the value of environment-friendly

innovation relative to fossil-based innovation, we estimate the models that are in the
vein of Hirshleifer et al. (2013) and Guenster et al. (2011). Besides estimating the
models by running the annual Fama and MacBeth (1973) cross-sectional regressions
for the period 1995–2012, we also conduct sub-period analysis of the most heavily
parameterised baseline model. We present a discussion of our findings in Sects.
5.1–5.3.

5.1 Association Between Tobin’s Q and Innovation

Productivity Variables

We first analyze the association between our aggregated innovation productivity

variables and the Tobin’s Q. To determine this association, we estimate the various
specifications of the model below:

Pati,t Citi,t 
ln(Tobin s Qi,t ) = α + β1 RDBEi,t + β2 + β3 + γk Zk + i,t (9)
Booki,t Booki,t
k

where Zk is a control variable. Our set of controls is described in the Data and Variable
section.
We report the time-series average slopes and corresponding standard errors of the
regressors for each of these models in Table 3. We find, for all the specifications, a
positive and highly statistically significant correlation between our R&D productivity
variable (RDBE) and the Tobin’s Q, on average. This corroborates the finding in prior
literature that the capital markets positively value R&D investments of a firm (Hall et
al., 2005). We get incoherent results for the association between patent productivity
variable (Pat/Book) and Tobin’s Q. Among the specifications with R&D and patent
productivity variables (Model 1; column 1 in Table 3) and these variables along with
Does the Market Value Clean Innovation? Evidence from US … 239

Table 3 Tobin’s Q as a function of aggregated innovation productivity variables and firm traits,
estimated using Fama-MacBeth regressions
(1) (2) (3) (4)
Intercept 0.628*** 0.577*** 0.761*** 0.758***
(0.129) (0.131) (0.149) (0.135)
RDBE 0.107*** 0.103*** 0.208*** 0.196***
(0.0266) (0.0236) (0.0273) (0.0267)
Pat/Book 0.112*** 0.0294 0.127*** 0.0390
(0.0213) (0.0568) (0.0218) (0.0359)
Cit/Book 0.0234* 0.0234***
(0.0126) (0.00772)
ln(Total_assets) −0.0167* −0.0159*
(0.00874) (0.00865)
Debt/Assets −0.00389*** −0.00392***
(0.000713) (0.000707)
Sales growth2yr 0.00140*** 0.00141***
(0.000158) (0.000158)
EBIT 1.39e−07*** 1.37e−07***
(2.71e−08) (2.69e−08)
invBE 0.273*** 0.273***
(0.0842) (0.0788)
CEME −0.0100*** −0.00997***
(0.00182) (0.00180)
ADME −0.672*** −0.670***
(0.0861) (0.0860)
taxRDBE 0.303** 0.309**
(0.124) (0.126)
Industry FE YES YES YES YES
Firm-level NO NO YES YES
controls
Observations 18,556 18,556 15,846 15,846
Avg. R-squared 0.171 0.177 0.408 0.410
Notes The table presents the regression results of various specifications of the model
logQit =α+γ1 RDBEit +γ2 Pat/Bookit +γ3 Cit/Bookit + γ4 ln(Total_Assets) + γ5 Debt/Assets +
γ6 Salesgrowth2yr + γ7 EBIT + γ8 invBEit + γ9 CEMEit + γ10 ADMEit + γ11 taxRDBEit +
 48
j=2 βj Industryj + it

that are estimated using Fama-MacBeth method. These models test whether the knowledge creation
process acts as a continuum from R&D to patents and citations. In our specifications we use RDBE
as a proxy for R&D productivity; Pat/Book as a proxy for patent productivity; Cit/Book as a proxy
for citation productivity. Our dependent variable is the natural logarithm of Tobin’s Q and we
report standard errors in parentheses. We use the following significance stars ∗∗∗ p < 0.01, ∗∗ p <
0.05, ∗ p < 0.1
240 A. Dechezleprêtre et al.

all the firm controls (Model 3), we find that the market values patent productivity.
However, when we add citation productivity variable to these specifications (Models
2 and 4) we note that there is a positive association between patent productivity
variable and Tobin’s Q, though the result is not statistically significant. Further,
we find in these specifications that the citation productivity variable is positively
and statistically significant correlated with Tobin’s Q. This result further provides
evidence that citations of a firm’s patents indicate the technological and economic
significance of the innovation (Hall et al., 2005; Hirshleifer et al., 2013).

5.2 Do US Firms Have Sufficient Incentive to Produce

Environment-Friendly Technologies?

To determine if there is a clean innovation premium, we disaggregate our citation

productivity variable into clean, dirty, and other components. To investigate if there is
premium associated with producing environment-friendly technologies, we estimate
various specifications of the model

Cit_cleani,t Cit_dirtyi,t
ln(Tobin s Qi,t ) = α + β1 RDBEi,t + β2 + β3 (10)
Booki,t Booki,t
Cit_otheri,t Pati,t 
+ β4 + β5 + γk Zk + i,t
Booki,t Booki,t
k

where Zk is a control variable.

We first estimate models with disaggregated citation productivity measures along
with aggregated R&D and patent productivity variables (Table 4). For all the models
reported in Table 4 we find a clean innovation premium that is highly statistically
significant. We find that the capital market accords a 1.20% higher valuation to the
firms producing environment-friendly innovation and decreases the market value of
firms producing fossil-based technologies to the tune of 0.45%, on average. Further-
more, the difference between clean and dirty citation productivity variables for all
the specifications is statistically different from zero at the 5% significance level.8 We
report the t-test results for the most heavily parameterized model (Model 4, Table 4)
in Table 13 in the Internet Appendix 1.
We further adopt firm-level market-value model to check if our results are invari-
ant. The market-value model, predominantly employed in the valuation of R&D
investments literature (Dechezleprêtre et al., 2020; Hall et al., 2005), assumes that
the stock market assigns value to the firm based on both its tangible and intangible
assets. This model is presented below

8 Tables 15 and 16 of Internet Appendix 2, using a Fama-Macbeth regression specification, report

consistent results with market value (equity market-to-book ratio; MTB) and future operating profit
(free cashflow over total assets; Cash_ta) as response variables.
Does the Market Value Clean Innovation? Evidence from US … 241

Table 4 Tobin’s Q as a function of disaggregated innovation citation productivity variables, esti-

mated using Fama-MacBeth regressions
(1) (2) (3) (4)
Intercept 0.450*** 0.496*** 0.601*** 0.570***
(0.135) (0.125) (0.129) (0.130)
RDBE 0.136*** 0.0988*** 0.108*** 0.106***
(0.0288) (0.0237) (0.0272) (0.0240)
Cit_clean/Book 0.199*** 0.180*** 0.169** 0.170**
(0.0574) (0.0563) (0.0630) (0.0720)
Cit_dirty/Book −0.774* −0.805* −0.800* −0.857*
(0.411) (0.407) (0.413) (0.426)
Cit_other/Book 0.0230*** 0.0212
(0.00635) (0.0131)
Pat/Book 0.106*** 0.0323
(0.0237) (0.0640)
Industry FE YES YES YES YES
Firm-level NO NO NO NO
controls
Observations 18,556 18,556 18,556 18,556
avg. R-squared 0.174 0.178 0.175 0.180
Notes The table presents the regression results of various specifications of the model
logQit = α + 
γ1 RDBEit + γ2 Cit_clean/Bookit + γ3 Cit_dirty/Bookit + γ4 Cit_other/Bookit + γ5
Pat/Bookit + 48 j=2 βj Industryj + it

that are estimated using Fama-MacBeth method. These models test whether the knowledge creation
process acts as a continuum from R&D to patents and clean citations. In our specifications we use
RDBE as a proxy for R&D productivity; Pat/Book as a proxy for patent productivity; Cit/Book as a
proxy for citation productivity. Our dependent variable is the natural logarithm of Tobin’s Q and we
report standard errors in parentheses. We use the following significance stars ∗∗∗ p < 0.01, ∗∗ p <
0.05, ∗ p < 0.1

 
Vi,t Ki,t
logQi,t = log = logqt + log 1 + γ + i,t (11)
Ai,t Ai,t

where Qi,t stands for Tobin’s Q. From the above model, one can estimate the average
effect of a unit currency invested in knowledge stocks on the firm’s market value.
Following this model, we proxy for KAi,ti,t with R&D expense over book value of
equity (RDBE), patents over book value of equity (Pat/Book) and adjusted patent
citations over book value of equity (Cit/Book), to be consistent across our models,
and estimate the models using non-linear least square methods. Further, we also
reconstruct our productivity variables (RDBE, Pat/Book, Cit/Book) in line with the
firm-level market-value model (i.e. we use book value of total assets as denominator
instead of book value of equity). We run baseline regressions with the redefined
variables, controlling for time and industry effects and report that our results are
242 A. Dechezleprêtre et al.

invariant to this alternative ‘book value of assets’ denominator. The results are now
reported in Tables 17 and 18 in the Internet Appendix.9
We then add our set of control variables, controlling for firm traits, to all these
models (Table 5). These set of controls include a subset of firm traits that are inspired
by the accounting-based asset pricing literature (Hirshleifer et al., 2013) and include
variables that account for firm size, level of risk and recent sales growth of the
firm (Guenster et al., 2011). For these heavily parameterized models, we again find
that the market recognizes the value of producing clean technologies and accords
1.20% higher valuation to the firms. We also report a negative association between
fossil-based innovation and Tobin’s Q for all the models; however, these results are
not statistically significant. For these heavily parameterized models we once again
find that the difference between the coefficient estimates of clean and dirty citation
productivity variables is statistically different from zero. We report the t-test results
for Model 4 of Table 5 in Table 14 in the Internet Appendix 1.

5.3 Association Between Disaggregated Innovation

Productivity Variables and Tobin’s Q: Tests for
Time-Varying Capital Market Response

We further seek to investigate whether clean innovation premium holds over time. For
this we estimate the most heavily parameterized model (column 4) reported in Tables
4 and 5 for two sub-periods. Sub-period 1 runs from 1995–2005 and sub-period 2
runs from 2006–2012.
We report the time series average slopes and the corresponding standard errors
for the regressors of Model 4 (column 4) of Table 4 for the two sub-periods in Table
6. We note an increase in association between clean citation productivity variable
and Tobin’s Q. However, only during the 2006–2012 period do we find a statistically
and economically significant positive association between clean citation productivity
variable and Tobin’s Q. A possible interpretation for this finding is that the increase in
environmental policy stringency led the market to positively value clean innovation,
which can help firms comply with the new regulations. In this sub-period, we find that
the market value of firms producing environment-friendly technologies increases by
1.3%. We also find that the negative association between dirty citation productivity
variable and Tobin’s Q decreases as we move from sub-period 1 to 2.10

9 We have also conducted a test which splits the “other patents” category into innovation related to
emerging technologies (which is distinct to the clean and dirty technology innovation categories)
and the rest. When we include emerging technology variants of innovation productivity variables
in our regressions, (1) it shows a positive relation with Tobin’s Q and, critically, (2) its inclusion
doesn’t compromise the main result obtained in a clean innovation premium. This new finding is
now reported in Table 19 in the Internet Appendix 3. The patent classification codes used to extract
emerging patents from the database is presented in Table 10 in the Internet Appendix 1.
10 We further estimate the above Model for three sub-periods, 1995–2000, 2000–2005 and 2005–

2012, respectively, and find similar results. Please see Table 11 in the Internet Appendix 1.
Does the Market Value Clean Innovation? Evidence from US … 243

Table 5 Tobin’s Q as a function of disaggregated innovation citation productivity variables and

firm traits, estimated using Fama-MacBeth regressions
(1) (2) (3) (4)
Intercept 0.841*** 0.778*** 0.753*** 0.723***
(0.161) (0.167) (0.150) (0.146)
RDBE 0.243*** 0.207*** 0.210*** 0.199***
(0.0279) (0.0249) (0.0269) (0.0261)
Cit_clean/Book 0.211*** 0.188** 0.180** 0.184***
(0.0648) (0.0666) (0.0610) (0.0614)
Cit_dirty/Book −0.173 −0.207 −0.208 −0.220
(0.175) (0.178) (0.179) (0.180)
Cit_other/Book 0.0235*** 0.0233***
(0.00498) (0.00791)
Pat/Book 0.127*** 0.0353
(0.0240) (0.0335)
ln(Total_assets) −0.0169* −0.0156* −0.0158* −0.0151
(0.00875) (0.00859) (0.00875) (0.00867)
Debt/Assets −0.00381*** −0.00390*** −0.00388*** −0.00391***
(0.000709) (0.000716) (0.000713) (0.000712)
Sales growth2yr 0.00140*** 0.00140*** 0.00140*** 0.00141***
(0.000162) (0.000164) (0.000163) (0.000163)
EBIT 1.39e−07*** 1.37e−07*** 1.38e−07*** 1.36e−07***
(2.75e−08) (2.68e−08) (2.71e−08) (2.68e−08)
invBE 0.305*** 0.271*** 0.276*** 0.280***
(0.0870) (0.0802) (0.0841) (0.0786)
CEME −0.0100*** −0.00996*** −0.00999*** −0.00995***
(0.00181) (0.00180) (0.00182) (0.00180)
ADME −0.671*** −0.668*** −0.671*** −0.670***
(0.0854) (0.0854) (0.0860) (0.0860)
taxRDBE 0.291** 0.305** 0.302** 0.308**
(0.120) (0.126) (0.124) (0.127)
Industry FE YES YES YES YES
Firm-level controls YES YES YES YES
Observations 15,846 15,846 15,846 15,846
avg. R-squared 0.409 0.411 0.410 0.412
Notes The table presents the regression results of various specifications of the model
logQit = α + γ1 RDBEit + γ2 Cit_clean/Bookit + γ3 Cit_dirty/Bookit + γ4 Cit_other/Bookit + γ5
Pat/Bookit + γ6 ln(Total_Assets) + γ7 Debt/Assets+γ8 Salesgrowth2yr +γ9 EBIT + γ10 invBEit +

γ11 CEMEit + γ12 ADMEit + γ13 taxRDBEit + 48 j=2 βj Industryj + it

that are estimated using Fama-MacBeth method. These models test whether the knowledge creation
process acts as a continuum from R&D to patents and clean citations. In our specifications we use
RDBE as a proxy for R&D productivity; Pat/Book as a proxy for patent productivity; Cit/Book as a
proxy for citation productivity. Our dependent variable is the natural logarithm of Tobin’s Q and we
report standard errors in parentheses. We use the following significance stars ∗∗∗ p < 0.01, ∗∗ p <
0.05, ∗ p < 0.1
244 A. Dechezleprêtre et al.

Table 6 Tobin’s Q as a function of disaggregated innovation citation productivity variables, esti-

mated using Fama-MacBeth regressions on different sub periods
(1) (2)
Variables Subperiod 1 Subperiod 2
Intercept 0.713∗∗∗ 0.396∗∗
(0.184) (0.119)
RDBE 0.0997∗∗∗ 0.124∗∗
(0.0303) (0.0375)
Cit_clean/Book 0.101 0.268∗∗∗
(0.106) (0.0620)
Cit_dirty/Book −1.380∗∗ 0.00274
(0.617) (0.327)
Cit_other/Book 0.0112 0.0340∗∗∗
(0.0209) (0.00547)
Pat/Book 0.122 −0.0968∗∗∗
(0.0959) (0.0207)
Industry FE YES YES
Firm-level controls NO NO
Observations 10,652 8,975
avg. R-squared 0.197 0.156
Notes The table presents the regression results of the most heavily parameterised model of Table 3,
logQit = α + 
γ1 RDBEit + γ2 Cit_clean/Bookit + γ3 Cit_dirty/Bookit + γ4 Cit_other/Bookit + γ5
Pat/Bookit + 48 j=2 βj Industryj + it

that is estimated using Fama-MacBeth method. This Model tests whether the knowledge creation
process acts as a continuum from R&D to patents and clean citations. We estimate the above
regression model for two sub periods. Subperiod 1 is from 1995–2005 and Subperiod 2 is from
2006–2012. In our specifications we use RDBE as a proxy for R&D productivity; Pat/Book as a
proxy for patent productivity; Cit/Book as a proxy for citation productivity. Our dependent variable
is the natural logarithm of Tobin’s Q and we report standard errors in parentheses. We use the
following significance stars ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1

We further estimate this model along with all the controls (Model 4 in
Table 5) for the two sub-periods and find that, as observed earlier, there is a positive
association between clean citation productivity variable and Tobin’s Q. However,
it is statistically and economically significant only in sub-period 2 with increase in
market value to the tune of 1.2% approximately. We find a negative and a statistically
significant correlation between dirty citation productivity variable with Tobin’s Q for
sub-period 1. In this sub-period, the market value of the firms producing fossil-based
technologies decreases to the tune of 0.6% (Table 7).11

11We further estimate the above Model for three sub-periods, 1995–2000, 2000–2005 and 2005–
2012, respectively, and find similar results. Please see Table 12 in the Internet Appendix 1.
Does the Market Value Clean Innovation? Evidence from US … 245

Table 7 Tobin’s Q as a function of disaggregated innovation citation productivity variables and

firm traits, estimated using Fama-MacBeth regressions on different sub periods
(1) (2)
Variables Subperiod 1 Subperiod 2
Intercept 0.776∗∗ 0.763∗∗∗
(0.242) (0.166)
RDBE 0.168∗∗∗ 0.238∗∗∗
(0.0335) (0.0389)
Cit_clean/Book 0.183 0.184∗∗∗
(0.107) (0.0426)
Cit_dirty/Book −0.530∗∗ 0.179
(0.192) (0.274)
Cit_other/Book 0.0266∗ 0.0191∗∗
(0.0137) (0.00554)
Pat/Book 0.0671 −0.00563
(0.0525) (0.0348)
ln(Total_assets) −0.00650 −0.0261∗∗
(0.0138) (0.00814)
Debt/Assets −0.00563∗∗∗ −0.00171∗
(0.000744) (0.000710)
Sales growth2yr 0.00143∗∗∗ 0.00138∗∗∗
(0.000273) (0.000155)
EBIT 1.93e−07∗∗∗ 6.25e−08∗∗∗
(3.74e−08) (1.06e−08)
invBE 0.235∗∗ 0.339∗∗
(0.0969) (0.135)
CEME −0.0130∗∗∗ −0.00606∗∗∗
(0.00272) (0.00117)
ADME −0.710∗∗∗ −0.618∗∗∗
(0.138) (0.0944)
taxRDBE 0.480∗∗ 0.0870
(0.207) (0.0613)
Industry FE YES YES
Firm-level controls YES YES
Observations 8,297 7,549
Avg. R-squared 0.440 0.376
Notes The table presents the regression results of the most heavily parameterised model of Table 4,
logQit = α + γ1 RDBEit + γ2 Cit_clean/Bookit + γ3 Cit_dirty/Bookit + γ4 Cit_other/Bookit +
γ5 Pat/Bookit + γ6 ln(Total_Assets) + γ7 Debt/Assets + γ8 Salesgrowth2yr + γ9 EBIT +

γ10 invBEit + γ11 CEMEit + γ12 ADMEit + γ13 taxRDBEit + 48 j=2 βj Industryj + it

that is estimated using Fama-MacBeth method. This Model tests whether the knowledge creation
process acts as a continuum from R&D to patents and clean citations. We estimate the above
regression model for three sub periods. Subperiod 1 is from 1995–2005 and Subperiod 2 is from
2006–2012. In our specifications we use RDBE as a proxy for R&D productivity; Pat/Book as a
proxy for patent productivity; Cit/Book as a proxy for citation productivity. Our dependent variable
is the natural logarithm of Tobin’s Q and we report standard errors in parentheses. We use the
following significance stars ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1
246 A. Dechezleprêtre et al.

6 Conclusion and Discussion

This study asks whether there is an adequate capital market incentive to decarbonize
the US economy by implementing a change in the mix of technologies used to
produce and consume energy. Available evidence suggests that certain policies can
incentivize low-carbon innovation while discouraging fossil-based innovation (Calel
and Dechezleprêtre, 2016; Popp, 2002). In this study, we investigate whether the
cumulative impact of all existing factors (environmental policies, energy subsidies,
innovation policies, innovation capabilities etc.) incentivize or disincentivize ‘clean’
innovation.
In our study we examine, over time, the value accorded by the capital markets
to low-carbon (‘clean’) and fossil fuel (‘dirty’) innovation in the United States. We
address this question using a patent dataset from the US patent office which pertains
to 2526 US listed firms across 18 years, from 1995 to 2012. Informed by the seminal
literature that accords a market evaluation to firm-level eco-efficiency (e.g., Guenster
et al. (2011)) and knowledge stock (e.g., Hirshleifer et al. (2013)), we disaggregate
innovation measurements (e.g., Deng et al. (1999) and Gu (2005)) of US firms’
knowledge stock into clean, dirty and other innovation. We, then, elicit their market
evaluations over time.
We first analyse the association between our aggregated innovation productivity
variables with the Tobin’s Q to check if our results are consistent with prior literature
(Dechezleprêtre et al., 2020; Hall et al., 2005) and further to verify the validity of our
data and empirical set-up. Then, to determine if there is a clean innovation premium,
we disaggregate our innovation measurement into clean, dirty, and other components.
Our chief contribution consists in eliciting capital market evaluations associated with
the disaggregated Deng et al. (1999) and Gu (2005) innovation productivity measure
to account for ‘clean’ and ‘dirty’ innovation production. To sum up, we report that
‘clean’ innovation is typically associated with an economically important and in a
positive manner with Tobin’s Q, while the capital market ascribes no (or a negative)
market value influence to ‘dirty’ innovation.
We do not claim to establish the underlying mechanism that can account for a
clean or a dirty innovation premium. This is not the purpose of our study. Instead, we
wish to establish whether a clean or a dirty innovation premium exists in our data.
The raise in the number and stringency of climate change mitigation policies makes
it plausible that clean technologies enjoy greater growth opportunities than dirty
technologies. This hypothesis is consistent with the finding that the positive clean
innovation premium is statistically significant only in the post-2005 period which
corresponds to a sharp strengthening of environmental policy stringency in the US.
Nevertheless, there can be several other explanations that drive the existence of a
clean innovation premium. First, clean technologies are in general more novel than
dirty technologies and the lifecycle of technologies and industry growth could also
explain the clean innovation premium. Second, environment-friendly firms may also
have to invest more in the future than firms producing dirty patents in order to realize
the value of their patent stock. Third, markets could recognize that it may be more
Does the Market Value Clean Innovation? Evidence from US … 247

difficult to obtain patents on environment-friendly innovation than on fossil-based

innovation. We leave to future research to discern why such an empirical finding is
evident.
Internet Appendices 1–3 (Supplementary Material) for “Does the Market Value
Clean Innovation? Evidence from US Listed Firms”
Internet Appendix 1 See Tables 8, 9, 10, 11, 12, 13 and 14.

Table 8 Clean Patent classification codes

Patent code Definition
Y02E REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS,
RELATED TO ENERGY GENERATION, TRANSMISSION OR
DISTRIBUTION
Y02E10/00 Energy generation through renewable energy sources
Y02E20/00 Combustion technologies with mitigation potential
Y02E30/00 Energy generation of nuclear origin
Y02E40/00 Technologies for an efficient electrical power generation,
transmission or distribution
Y02E50/00 Technologies for the production of fuel of non-fossil origin
Y02E60/00 Enabling technologies or technologies with a potential or indirect
contribution to GHG emissions mitigation
Y02E70/00 Other energy conversion or management systems reducing GHG
emissions
Y02C CAPTURE, STORAGE, SEQUESTRATION OR DISPOSAL OF
GREENHOUSE GASES [GHG]
Y02C10/00 CO2 capture or storage
Y02C20/00 Capture or disposal of greenhouse gases [GHG] other than CO2
Y02T CLIMATE CHANGE MITIGATION TECHNOLOGIES
RELATED TO TRANSPORTATION
Y02T10/00 Road transport of goods or passengers
Y02T30/00 Transportation of goods or passengers via railways
Y02T50/00 Aeronautics or air transport
Y02T70/00 Maritime or waterways transport
Y02T90/00 Enabling technologies or technologies with a potential or indirect
contribution to GHG emissions mitigation
Y02B CLIMATE CHANGE MITIGATION TECHNOLOGIES
RELATED TO BUILDINGS, e.g., HOUSING, HOUSE
APPLIANCES OR RELATED END-USER APPLICATIONS
Y02B10/00 Integration of renewable energy sources in buildings
Y02B20/00 Energy efficient lighting technologies
Y02B30/00 Energy efficient heating, ventilation or air conditioning [HVAC]
(continued)
248 A. Dechezleprêtre et al.

Table 8 (continued)
Patent code Definition
Y02B40/00 Technologies aiming at improving the efficiency of home
appliances
Y02B50/00 Energy efficient technologies in elevators, escalators and moving
walkways
Y02B70/00 Technologies for an efficient end-user side electric power
management and consumption
Y02B80/00 Architectural or constructional elements improving the thermal
performance of buildings
Y02B90/00 Enabling technologies or technologies with a potential or indirect
contribution to GHG emissions mitigation

Table 9 Dirty Patent classification codes

Patent code Definition
C10J PRODUCTION OF PRODUCER GAS, WATER-GAS,
SYNTHESIS GAS FROM SOLID CARBONACEOUS
MATERIAL, OR MIXTURES CONTAINING THESE
GASES; CARBURETTING AIR OR OTHER GASES
F01K STEAM ENGINE PLANTS; STEAM ACCUMULATORS;
ENGINE PLANTS NOT OTHERWISE PROVIDED FOR;
ENGINES USING SPECIAL WORKING FLUIDS OR
CYCLES
F02C GAS-TURBINE PLANTS; AIR INTAKES FOR
JET-PROPULSION PLANTS; CONTROLLING FUEL
SUPPLY IN AIR-BREATHING JET-PROPULSION PLANTS
F02G HOT GAS OR COMBUSTION-PRODUCT
POSITIVE-DISPLACEMENT ENGINE PLANTS; USE OF
WASTE HEAT OF COMBUSTION ENGINES; NOT
OTHERWISE PROVIDED FOR
F22 STEAM GENERATION
F23 COMBUSTION APPARATUS; COMBUSTION PROCESSES
F27 FURNACES; KILNS; OVENS; RETORTS
F02B INTERNAL-COMBUSTION PISTON ENGINES;
COMBUSTION ENGINES IN GENERAL
Does the Market Value Clean Innovation? Evidence from US … 249

Table 10 Emerging Technologies Patent classification codes

Patent code Definition
Nanotechnology
B82 NANOTECHNOLOGY
GMO
C12N/15 Mutation or genetic engineering; DNA or RNA concerning genetic
engineering, vectors, e.g., plasmids, or their isolation, preparation or
purification; Use of hosts therefor
3D
H04N/13 Stereoscopic video systems; Multi-view video systems; Details thereof
Wireless
H04W WIRELESS COMMUNICATION NETWORKS
Robots
B25J MANIPULATORS; CHAMBERS PROVIDED WITH
MANIPULATION DEVICES
IT
G06 (excl G06Q) COMPUTING; CALCULATING; COUNTING
G10L SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION;
SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO
CODING OR DECODING
Biotechnology
C07G COMPOUNDS OF UNKNOWN CONSTITUTION
C07K PEPTIDES
C12M APPARATUS FOR ENZYMOLOGY OR MICROBIOLOGY
C12N MICROORGANISMS OR ENZYMES; COMPOSITIONS
THEREOF
C12P FERMENTATION OR ENZYME-USING PROCESSES TO
SYNTHESISE A DESIRED CHEMICAL COMPOUND OR
COMPOSITION OR TO SEPARATE OPTICAL ISOMERS FROM A
RACEMIC MIXTURE
C12Q MEASURING OR TESTING PROCESSES INVOLVING
ENZYMES, NUCLEIC ACIDS OR MICROORGANISMS;
COMPOSITIONS OR TEST PAPERS THEREFOR; PROCESSES
OF PREPARING SUCH COMPOSITIONS;
CONDITION-RESPONSIVE CONTROL IN MICROBIOLOGICAL
OR ENZYMOLOGICAL PROCESSES
C12R MICROORGANISMS
250 A. Dechezleprêtre et al.

Table 11 Tobin’s Q as a function of disaggregated innovation citation productivity variables,

estimated using Fama-MacBeth regressions on different sub periods
(1) (2) (3)
VARIABLES Subperiod 1 Subperiod 2 Subperiod 3
Intercept 0.519 0.823** 0.444**
(0.336) (0.278) (0.129)
RDBE 0.155** 0.0275 0.124**
(0.0465) (0.0215) (0.0375)
Cit_clean/Book −0.0686 0.252 0.268***
(0.162) (0.150) (0.0620)
Cit_dirty/Book −2.764* −0.325 0.00274
(1.070) (0.236) (0.327)
Cit_other/Book 0.00437 0.0177 0.0340***
(0.0452) (0.0187) (0.00547)
Pat/Book 0.255 0.0165 −0.0968***
(0.198) (0.0561) (0.0207)
Industry FE YES YES YES
Firm-level controls NO NO NO
Observations 4,789 4,792 8,975
avg. R-squared 0.224 0.173 0.156
Notes The table presents the regression results of the most heavily parameterised model of Table 3,
logQit = α + γ1
RDBEit + γ2 Cit_clean/Bookit + γ3 Cit_dirty/Bookit + γ4 Cit_other/Bookit +
γ5 Pat/Bookit + 48
j=2 βj Industryj + it

that is estimated using Fama-MacBeth method. This model tests whether the knowledge creation
process acts as a continuum from R&D to patents and clean citations. We estimate the above
regression model for three sub periods. Subperiod 1 is from 1995–2000, Subperiod 2 is from 2000–
2005, and Subperiod 3 is from 2005–2012. In our specifications we use RDBE as a proxy for
R&D productivity; Pat/Book as a proxy for patent productivity; Cit/Book as a proxy for citation
productivity. Our dependent variable is the natural logarithm of Tobin’s Q and we report standard
errors in parentheses. We use the following significance stars ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1
Does the Market Value Clean Innovation? Evidence from US … 251

Table 12 Tobin’s Q as a function of disaggregated innovation citation productivity variables,

estimated using Fama-MacBeth regressions on different sub periods
(1) (2) (3)
Variables Subperiod 1 Subperiod 2 Subperiod 3
Intercept 0.739* 0.644 0.767***
(0.187) (0.466) (0.0965)
RDBE 0.219*** 0.102** 0.252***
(0.00663) (0.0306) (0.0364)
Cit_clean/Book 0.152 0.225 0.170***
(0.222) (0.159) (0.0397)
Cit_dirty/Book −0.974* −0.503** 0.240
(0.246) (0.117) (0.245)
Cit_other/Book 0.0473 0.0170 0.0183***
(0.0326) (0.0161) (0.00486)
Pat/Book 0.0987 0.0604 −0.00415
(0.149) (0.0537) (0.0301)
ln(Total_assets) 0.0103 −0.00793 −0.0290***
(0.00485) (0.0233) (0.00764)
Debt/Assets −0.00745** −0.00479*** −0.00204**
(0.00163) (0.000665) (0.000698)
Sales growth2yr 0.00220* 0.00108*** 0.00132***
(0.000521) (0.000231) (0.000148)
EBIT 2.37e−07 1.93e−07*** 6.20e−08***
(9.38e−08) (3.43e−08) (9.18e−09)
invBE 0.269 0.273* 0.290*
(0.234) (0.112) (0.127)
CEME −0.0112 −0.0134** −0.00727***
(0.00412) (0.00455) (0.00158)
ADME −0.617** −0.627** −0.716***
(0.0658) (0.200) (0.128)
taxRDBE 0.143 0.681 0.137
(0.0954) (0.353) (0.0729)
Industry FE YES YES YES
Firm-level controls YES YES YES
Observations 2,804 4,481 8,561
Avg. R-squared 0.480 0.411 0.387
Notes The table presents the regression results of the most heavily parameterised model of Table 4,
logQit = α + γ1 RDBEit + γ2 Cit_clean/Bookit + γ3 Cit_dirty/Bookit + γ4 Cit_other/Bookit + γ5
Pat/Bookit + γ6 ln(Total_Assets) + γ7 Debt/Assets + γ8 Salesgrowth2yr + γ9 EBIT + γ10 inv

BEit + γ11 CEMEit + γ12 ADMEit + γ13 taxRDBEit + 48 j=2 βj Industryj + it ,

that is estimated using Fama-MacBeth method. This Model tests whether the knowledge creation
process acts as a continuum from R&D to patents and clean citations. We estimate the above
regression model for three sub periods. Subperiod 1 is from 1995–2000, Subperiod 2 is from 2000–
2005, and Subperiod 3 is from 2005–2012. In our specifications we use RDBE as a proxy for
R&D productivity; Pat/Book as a proxy for patent productivity; Cit/Book as a proxy for citation
productivity. Our dependent variable is the natural logarithm of Tobin’s Q and we report standard
errors in parentheses. We use the following significance stars ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1
Table 13 The table reports the linear hypothesis for the coefficients from the most heavily parameterised model reported in Table 4
252

t test for Fama-MacBeth F stat Prob > F

estimation
Cit_clean/Book − 7.53 0.0138
Cit_dirty/Book = 0
Cit_clean/Book = 0 5.59 0.0302
Cit_dirty/Book = 0 4.05 0.0604
(a) Panel A: Test for linear hypotheses after estimation for the most heavily parameterised model reported in Table 4 (column 4 in Table 4) during 1995–2012
Subperiod: 1995–2000 Subperiod: 2000–2005 Subperiod: 2005–2012
t test for Fama-MacBeth F stat Prob > F F stat Prob > F F stat Prob > F
estimation
Cit_clean/Book − 8.67 0.0422 2.87 0.1653 0.94 0.3657
Cit_dirty/Book = 0
Cit_clean/Book = 0 0.18 0.6944 2.85 0.1666 18.71 0.0035
Cit_dirty/Book = 0 6.67 0.0611 1.90 0.2402 0.00 0.9936
Panel B: Test for linear hypotheses after estimation for the most heavily parameterised model reported in Table 4 (column 4 in Table 4) during Subperiod 1995–2000,
Subperiod 2000–2005, and Subperiod 2005–2012
Subperiod: 1995–2005 Subperiod: 2006–2012
t Test for Fama-MacBeth F stat Prob > F F stat Prob > F
estimation
Cit_clean/Book − 7.34 0.0220 1.00 0.3553
Cit_dirty/Book = 0
Cit_clean/Book = 0 0.91 0.3631 15.59 0.0075
Cit_dirty/Book = 0 5.01 0.0491 0.01 0.9307
Panel C: Test for linear hypotheses after estimation for the most heavily parameterised model reported in Table 4 (column 4 in Table 4) during Subperiod 1995–2005 and
Subperiod 2006–2012
Notes The table presents the tests for linear hypotheses after estimation of the model
48
logQit = α + γ1 RDBEit + γ2 Cit_clean/Bookit + γ3 Cit_dirty/Bookit + γ4 Cit_other/Bookit + γ5 Pat/Bookit + j=2 βj Industryj + it
that is estimated using Fama-MacBeth method. The model tests whether the knowledge creation process acts as a continuum from R&D to patents and clean
citations. We conduct these tests for the whole sample and for various subperiods. In our specifications we use RDBE as a proxy for R&D productivity; Pat/Book
A. Dechezleprêtre et al.

as a proxy for patent productivity; Cit/Book as a proxy for citation productivity. Our dependent variable is the natural logarithm of Tobin’s Q
Table 14 The table reports the linear hypothesis for the coefficients from the most heavily parameterised model reported in Table 5
t Test for Fama-MacBeth F stat Prob > F
estimation
Cit_clean/Book − 4.37 0.0539
Cit_dirty/Book = 0
Cit_clean/Book = 0 8.95 0.0091
Cit_dirty/Book = 0 1.49 0.2411
(a) Panel A: Test for linear hypotheses after estimation for the most heavily parameterised model reported in Table 5 (column 4 in Table 5) during 1995–2012
Subperiod: 1995–2000 Subperiod: 2000–2005 Subperiod: 2005–2012
t Test for Fama-MacBeth F stat Prob > F F stat Prob > F F stat Prob > F
estimation
Cit_clean/Book − 8.06 0.1049 8.69 0.0421 0.10 0.7659
Cit_dirty/Book = 0
Cit_clean/Book = 0 0.47 0.5651 2.00 0.2302 18.26 0.0037
Cit_dirty/Book = 0 15.75 0.0580 18.56 0.0126 0.96 0.3595
(b) Panel B: Test for linear hypotheses after estimation for the most heavily parameterised model reported in Table 5 (column 4 in Table 5) during Subperiod 1995–2000,
Subperiod 2000–2005, and Subperiod 2005–2012
Subperiod: 1995–2005 Subperiod: 2006–2012
t Test for Fama-MacBeth F stat Prob > F F stat Prob > F
estimation
Cit_clean/Book − 8.35 0.0202 0.00 0.9850
Does the Market Value Clean Innovation? Evidence from US …

Cit_dirty/Book = 0
Cit_clean/Book = 0 2.92 0.1261 18.69 0.0050
Cit_dirty/Book = 0 7.63 0.0246 0.43 0.5364
Panel C: Test for linear hypotheses after estimation for the most heavily parameterised model reported in Table 5 (column 4 in Table 5) during Subperiod 1995–2005 and
Subperiod 2006–2012
Notes The table presents the tests for linear hypotheses after estimation of the model
logQit = α + γ1 RDBEit + γ2 Cit_clean/Bookit + γ3 Cit_dirty/Bookit + γ4 Cit_other/Bookit+ γ5 Pat/Bookit + γ6 ln(Total_Assets) + γ7 Debt/Assets +
γ8 Salesgrowth2yr + γ9 EBIT + γ10 invBEit + γ11 CEMEit + γ12 ADMEit + γ13 taxRDBEit + 48 j=2 βj Industryj + it

that is estimated using Fama-MacBeth method. The model tests whether the knowledge creation process acts as a continuum from R&D to patents and clean
citations. We conduct these tests for the whole sample and various subperiods. In our specifications we use RDBE as a proxy for R&D productivity; Pat/Book
253

as a proxy for patent productivity; Cit/Book as a proxy for citation productivity. Our dependent variable is the natural logarithm of Tobin’s Q
254 A. Dechezleprêtre et al.

Internet Appendix 2: Alternative dependent variables: Equity Market-to-Book

ratio (MTB) and Free Cashflow-to-Total assets ratio (Cash_ta)
See Tables 15 and 16.

Table 15 Equity Market-to-Book ratio as a function of disaggregated innovation citation produc-

tivity variables, estimated using Fama-MacBeth regressions
(1) (2) (3) (4)
Intercept 131.9∗∗∗ 130.7∗∗∗ 152.5∗∗∗ 152.2∗∗∗
(36.41) (37.02) (37.03) (38.26)
RDBE 57.73∗∗∗ 52.73∗∗∗ 55.64∗∗∗ 54.32∗∗∗
(7.124) (6.513) (6.461) (6.302)
Cit_clean/Book 28.96∗∗∗ 25.59∗∗∗ 25.15∗∗∗ 25.05∗∗∗
(7.447) (6.593) (7.072) (8.325)
Cit_dirty/Book −67.96 −74.16 −70.84 −77.04
(46.70) (46.12) (46.47) (47.55)
Cit_other/Book 3.084∗∗ 3.339∗
(1.068) (1.726)
Pat/Book 10.70∗∗ −1.580
(4.019) (8.098)
Industry FE YES YES YES YES
Firm-level NO NO NO NO
controls
Observations 18,558 18,558 18,558 18,558
Avg. R-squared 0.204 0.210 0.207 0.212
Notes The table presents the regression results of various specifications of the model
logMTBit = α +γ1 RDBEit + γ2 Cit_clean/Bookit + γ3 Cit_dirty/Bookit + γ4 Cit_other/Bookit +
γ5 Pat/Bookit + 48
j=2 βj Industryj + it

that are estimated using Fama-MacBeth method. These models test whether the knowledge creation
process acts as a continuum from R&D to patents and clean citations. In our specifications we use
RDBE as a proxy for R&D productivity; Pat/Book as a proxy for patent productivity; Cit/Book as a
proxy for citation productivity. Our dependent variable is the natural logarithm of Tobin’s Q and we
report standard errors in parentheses. We use the following significance stars ∗∗∗ p < 0.01, ∗∗ p <
0.05, ∗ p < 0.1
Does the Market Value Clean Innovation? Evidence from US … 255

Table 16 Free Cashflow-to-Total assets ratio as a function of disaggregated innovation citation

productivity variables, estimated using Fama-MacBeth regressions
(1) (2) (3) (4)
Intercept 3.262∗∗∗ 3.400∗∗∗ 3.427∗∗∗ 3.491∗∗∗
(0.213) (0.247) (0.195) (0.224)
RDBE 0.284∗∗∗ 0.244∗∗∗ 0.256∗∗∗ 0.260∗∗∗
(0.0625) (0.0636) (0.0569) (0.0639)
Cit_clean/Book 0.516∗∗∗ 0.486∗∗∗ 0.465∗∗∗ 0.450∗∗∗
(0.131) (0.133) (0.124) (0.115)
Cit_dirty/Book −1.446 −1.471 −1.472 −1.532
(0.974) (0.973) (0.981) (0.986)
Cit_other/Book 0.0354∗∗ 0.0289
(0.0139) (0.0203)
Pat/Book 0.140∗∗ 0.0658
(0.0538) (0.0921)
Industry FE YES YES YES YES
Firm-level NO NO NO NO
controls
Observations 17,897 17,897 17,897 17,897
Avg. R-squared 0.250 0.252 0.251 0.253
Notes The table presents the regression results of various specifications of the model
logCash_tai,t+1 = α + γ1 RDBEit + γ2 Cit_clean/Bookit + γ3 Cit_dirty/Bookit +

γ4 Cit_other/Bookit + γ5 Pat/Bookit + 48
j=2 βj Industryj + it

that are estimated using Fama-MacBeth method. These models test whether the knowledge creation
process acts as a continuum from R&D to patents and clean citations. In our specifications we use
RDBE as a proxy for R&D productivity; Pat/Book as a proxy for patent productivity; Cit/Book as a
proxy for citation productivity. Our dependent variable is the natural logarithm of Tobin’s Q and we
report standard errors in parentheses. We use the following significance stars ∗∗∗ p < 0.01, ∗∗ p <
0.05, ∗ p < 0.1

Internet Appendix 3: Replacement cost

See Tables 17, 18, 19 and 20
256 A. Dechezleprêtre et al.

Table 17 Tobin’s Q as a function of disaggregated innovation citation productivity variables,

estimated using non-linear least squares method
(1) (2) (3) (4)
Intercept 0.100 0.0997 0.0997 0.0997
(0.212) (0.212) (0.212) (0.212)
RDBE 0.334∗∗∗ 0.203∗∗ 0.243∗∗ 0.200∗
(0.0958) (0.0772) (0.0869) (0.0789)
Cit_clean/Book 0.340∗ 0.298∗ 0.281∗ 0.296∗
(0.135) (0.122) (0.120) (0.122)
Cit_dirty/Book −0.716∗∗ −0.737∗∗∗ −0.762∗∗ −0.745∗∗∗
(0.251) (0.217) (0.249) (0.225)
Cit_other/Book 0.0827∗∗ 0.0798∗
(0.0320) (0.0373)
Pat/Book 0.334∗ 0.0228
(0.146) (0.138)
Time FE YES YES YES YES
Industry FE YES YES YES YES
Firm-level controls NO NO NO NO
Observations 18,556 18,556 18,556 18,556
Adjusted R2 0.161 0.164 0.162 0.164
Notes The table presents the regression results of various specifications of the model
logQit = α + log(1 + γ1 RDBEit + γ2 Cit_clean/Bookit + γ3 Cit_dirty/Bookit +
γ4 Cit_other/Book
 it + γ5 48
Pat/Bookit + 2012
i=1996 κi yeari + j=2 βj Industryj ) + it

that are estimated using non-linear least squares method and are in the vein of the models reported
in Hall et al. (2005). These models test whether the knowledge creation process acts as a contin-
uum from R&D to patents and clean citations. In our specifications we use RDBE as a proxy for
R&D productivity; Pat/Book as a proxy for patent productivity; Cit/Book as a proxy for citation
productivity. Our dependent variable is the natural logarithm of Tobin’s Q and we report clustered
standard errors in parentheses. We use the following significance stars ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗
p < 0.001
Does the Market Value Clean Innovation? Evidence from US … 257

Table 18 Tobin’s Q as a function of disaggregated innovation citation productivity variables,

estimated using non-linear least squares method
(1) (2) (3) (4)
Intercept 0.105 0.106 0.105 0.106
(0.210) (0.210) (0.210) (0.210)
RDTA 2.839∗∗∗ 2.343∗∗∗ 2.482∗∗∗ 2.326∗∗∗
(0.655) (0.565) (0.590) (0.563)
Cit_clean/Total_assets 0.597∗ 0.453∗ 0.386∗ 0.422∗
(0.258) (0.177) (0.167) (0.168)
Cit_dirty/Total_assets −0.926∗∗∗ −0.937∗∗∗ −0.935∗∗∗ −0.938∗∗∗
(0.215) (0.210) (0.210) (0.210)
Cit_other/Total_assets 0.247∗∗ 0.216∗∗
(0.0779) (0.0775)
Pat/Total_assets 1.104∗∗ 0.249
(0.387) (0.313)
Time FE YES YES YES YES
Industry FE YES YES YES YES
Firm-level controls NO NO NO NO
Observations 18,556 18,556 18,556 18,556
Adjusted R2 0.191 0.197 0.195 0.197
Notes The table presents the regression results of various specifications of the model
logQit = α + log(1 + γ1 RDTAit + γ2 Cit_clean/Total_assets
 it + γ3 Cit_dirty/Total_assets
48 it + γ4
Cit_other/Total_assetsit + γ5 Pat/Total_assetsit + 2012
i=1996 κi yeari + j=2 βj Industryj ) + it

that are estimated using non-linear least squares method and are in the vein of the Models reported
in Hall et al. (2005). These models test whether the knowledge creation process acts as a continuum
from R&D to patents and clean citations. In our specifications we use RDTA as a proxy for R&D
productivity; Pat/Total_assets as a proxy for patent productivity; Cit/Total_assets as a proxy for
citation productivity. Our dependent variable is the natural logarithm of Tobin’s Q and we report
clustered standard errors in parentheses. We use the following significance stars ∗ p < 0.05, ∗∗ p <
0.01, ∗∗∗ p < 0.001
258 A. Dechezleprêtre et al.

Table 19 Tobin’s Q as a function of disaggregated innovation citation productivity variables,

including emerging technology variants of innovation productivity, estimated using Fama-MacBeth
regressions
(1) (2)
Intercept 0.570∗∗∗ 0.627∗∗∗
(0.130) (0.125)
RDBE 0.106∗∗∗ 0.103∗∗∗
(0.0240) (0.0233)
Cit_clean/Book 0.170∗∗ 0.173∗∗
(0.0720) (0.0701)
Cit_dirty/Book −0.857∗ −0.896∗∗
(0.426) (0.422)
Cit_emtech/Book 0.123∗∗
(0.0430)
Cit_other/Book 0.0212 0.0186
(0.0131) (0.0133)
Pat/Book 0.0323 0.00675
(0.0640) (0.0669)
Industry FE YES YES
Firm-level controls NO NO
Observations 18,556 18,556
Adjusted R2 0.180 0.182
Notes The table presents the regression results of various specifications of the model
logQit = α + γ1 RDBEit + γ2 Cit_clean/Book
 it + γ3 Cit_dirty/Bookit + γ4 Cit_emtech/Bookit +
γ5 Cit_other/Bookit + γ6 Pat/Bookit + 48
j=2 βj Industryj + it

that are estimated using Fama-MacBeth method. These models test whether the knowledge creation
process acts as a continuum from R&D to patents and clean citations. In our specifications we use
RDBE as a proxy for R&D productivity; Pat/Book as a proxy for patent productivity; Cit/Book as a
proxy for citation productivity. Our dependent variable is the natural logarithm of Tobin’s Q and we
report standard errors in parentheses. We use the following significance stars ∗∗∗ p < 0.01, ∗∗ p <
0.05, ∗ p < 0.1
Does the Market Value Clean Innovation? Evidence from US … 259

Table 20 Tobin’s Q as a function of disaggregated innovation citation productivity variables,

industry sectors and the interaction between the citation productivity variables and drugs industry
sector, estimated using Fama-MacBeth regressions
(1) (2) (3)
Intercept 1.006∗∗∗ 0.570∗∗∗ 0.594∗∗∗
(0.0536) (0.130) (0.124)
RDBE 0.180∗∗∗ 0.106∗∗∗ 0.105∗∗∗
(0.0525) (0.024) (0.0239)
Cit_clean/Book −935.7 0.170∗∗ 0.173∗∗
(865.3) (0.072) (0.0730)
Cit_dirty/Book 199.2∗∗ −0.857∗ −0.865∗
(93.33) (0.426) (0.425)
Cit_other/Book 0.0274 0.021 0.0211
(0.0241) (0.013) (0.0130)
Pat/Book −0.207 0.032 0.0317
(0.122) (0.064) (0.0640)
Industry = Drugs 0.419∗∗∗
(0.131)
Cit_clean/Book*Drugs −967.9
(917.8)
Cit_dirty/Book*Drugs 191.8∗∗
(86.57)
Industry FE NO YES YES
Firm-level controls NO NO NO
Observations 1,516 18,556 18,556
Adjusted R2 0.095 0.180 0.182
Notes The table presents the regression results of various specifications of the model
logQit = α + γ1 RDBEit + γ2 Cit_clean/Bookit + γ3 Cit_dirty/Bookit + γ4 Cit_other/Bookit +
γ5 Pat/Book
 it + μ1 Cit_clean/Bookit ∗ Drugs + μ2 Cit_dirty/Bookit ∗ Drugs +
+ 48 j=2 βj Industryj + it

that are estimated using Fama-MacBeth method. These models test whether the knowledge creation
process acts as a continuum from R&D to patents and clean citations. We choose the Pharmaceutical
Products (henceforth Drugs) industry and focus on its interaction with our clean and dirty innovation
citation productivity variables. The sample of firms in the first regression model (column 1) of the
Table belong to the Drugs industry sector. In the other regression models (columns 2 and 3) of
the Table the sample of firms is the whole sample of firms in our data set. In our specifications
we use RDBE as a proxy for R&D productivity; Pat/Book as a proxy for patent productivity;
Cit/Book as a proxy for citation productivity. Our dependent variable is the natural logarithm of
Tobin’s Q and we report standard errors in parentheses. We use the following significance stars
∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1
260 A. Dechezleprêtre et al.

References

Aghion, P., Dechezleprêtre, A., Hemous, D., Martin, R., & Van Reenen, J. (2016). Carbon taxes,
path dependency, and directed technical change: Evidence from the auto industry. Journal of
Political Economy, 124(1), 1–51.
Benner, M. J. (2007). The incumbent discount: Stock market categories and response to radical
technological change. Academy of Management Review, 32(3), 703–720.
Botta, E., & Koźluk, T. (2014). Measuring environmental policy stringency in OECD countries:
A composite index approach. OECD Economics Department Working Papers, No. 1177. Paris:
OECD Publishing. http://dx.doi.org/10.1787/5jxrjnc45gvg-en
Bowman, E. H., & Haire, M. (1975). A strategic posture toward corporate social responsibility.
California Management Review, 18(2), 49–58.
Calel, R., & Dechezleprêtre, A. (2016). Environmental policy and directed technological change:
Evidence from the European carbon market. Review of Economics and Statistics, 98(1), 173–191.
Chan, L. K., Lakonishok, J., & Sougiannis, T. (2001). The stock market valuation of research and
development expenditures. The Journal of Finance, 56(6), 2431–2456.
Dechezleprêtre, A., Muckley, C. B., & Neelakantan, P. (2020). Is firm-level clean or dirty innovation
valued more? The European Journal of Finance, 27(1–2), 31–61.
Deng, Z., Lev, B., & Narin, F. (1999). Science and technology as predictors of stock performance.
Financial Analysts Journal, 55(3), 20–32.
Dowell, G., Hart, S., & Yeung, B. (2000). Do corporate global environmental standards create or
destroy market value? Management Science, 46(8), 1059–1074.
Fama, E. (1991). Efficient capital markets: II. Journal of Finance, 46(5), 1575–1617.
Fama, E., & MacBeth, J. (1973). Risk, return and equilibrium: Empirical tests. Journal of Political
Economy, 81(3), 607.
Fisher-Vanden, K., & Thorburn, K. S. (2011). Voluntary corporate environmental initiatives and
shareholder wealth. Journal of Environmental Economics and Management, 62(3), 430–445.
Frietsch, R., & Schmoch, U. (2006). Technological structures and performance as reflected by patent
indicators. National systems of innovation in comparison (pp. 89–105). Springer.
Grandi, A., Hall, B. H., & Oriani, R. (2009). R&D and financial investors (pp. 143–165). Chel-
tenham, UK: Evaluation and Performance Measurement of Research and Development, Edward
Elgar.
Griliches, Z. (1998). Patent statistics as economic indicators: A survey. R&D and productivity: The
econometric evidence (pp. 287–343). University of Chicago Press.
Gu, F. (2005). Innovation, future earnings, and market efficiency. Journal of Accounting, Auditing
& Finance, 20(4), 385–418.
Guenster, N., Bauer, R., Derwall, J., & Koedijk, K. (2011). The economic value of corporate eco-
efficiency. European Financial Management, 17(4), 679–704.
Hall, B. H., Jaffe, A., & Trajtenberg, M. (2005). Market value and patent citations. RAND Journal
of Economics, 16–38.
Hart, S. L. (1995). A natural-resource-based view of the firm. Academy of Management Review,
20(4), 986–1014.
Hart, S. L., & Ahuja, G. (1996). Does it pay to be green? An empirical examination of the relationship
between emission reduction and firm performance. Business Strategy and the Environment, 5(1),
30–37.
Heinkel, R., Kraus, A., & Zechner, J. (2001). The effect of green investment on corporate behavior.
Journal of Financial and Quantitative Analysis, 431–449.
Hirshleifer, D., Hsu, P.-H., & Li, D. (2013). Innovative efficiency and stock returns. Journal of
Financial Economics, 107(3), 632–654.
IPCC. (2014). Climate change 2014: Synthesis report. Contribution of Working Groups I, II and III
to the Fifth Assessment Report of the intergovernmental panel on Climate Change (Vol. 151, p.
151). Geneva, Switzerland: IPCC.
Does the Market Value Clean Innovation? Evidence from US … 261

Karpoff, J. M., Lott, J. R., Jr., & Wehrly, E. W. (2005). The reputational penalties for environmental
violations: Empirical evidence. The Journal of Law and Economics, 48(2), 653–675.
Khan, M., & Dernis, H. (2006). Global overview of innovative activities from the patent indicators
perspective. OECD Science, Technology and Industry Working Papers, No. 2006/03. Paris: OECD
Publishing. https://doi.org/10.1787/674714465672
Klassen, R. D., & McLaughlin, C. P. (1996). The impact of environmental management on firm
performance. Management Science, 42(8), 1199–1214.
Makower, J. (1994). The E-factor: The bottom-line approach to environmentally responsible busi-
ness. Plume Books.
Munari, F., Oriani, R., & Sobrero, M. (2005). Do owner identity and financial markets affect R&D
investments? In Academy of Management: An analysis of European firms.
Noailly, J., & Smeets, R. (2015). Directing technical change from fossil-fuel to renewable energy
innovation: An application using firm-level patent data. Journal of Environmental Economics and
Management, 72, 15–37.
Nyborg, K., & Zhang, T. (2013). Is corporate social responsibility associated with lower wages?
Environmental and Resource Economics, 55(1), 107–117.
Palmer, K., Oates, W. E., & Portney, P. R. (1995). Tightening environmental standards: The benefit-
cost or the no-cost paradigm? Journal of Economic Perspectives, 9(4), 119–132.
Pandit, S., Wasley, C. E., & Zach, T. (2011). The effect of research and development (R&D) inputs
and outputs on the relation between the uncertainty of future operating performance and R&D
expenditures. Journal of Accounting, Auditing & Finance, 26(1), 121–144.
Popp, D. (2002). Induced innovation and energy prices. American Economic Review, 92(1), 160–
180.
Porter, M., & Van der Linde, C. (1995). Green and competitive: Ending the stalemate. Harvard
Business Review, 73(5), 120–134.
Reinhardt, F. L. (1999). Bringing the environment down to earth. Harvard Business Review, 77(4),
149–149.
Russo, M. V., & Fouts, P. A. (1997). A resource-based perspective on corporate environmental
performance and profitability. Academy of Management Journal, 40(3), 534–559.
Schmidheiny, S., & Timberlake, L. (1992). Changing course: A global business perspective on
development and the environment, (Vol. 1). MIT Press.
Shane, P. B., & Spicer, B. H. (1983). Market response to environmental information produced
outside the firm. Accounting Review, 521–538.
Turban, D. B., & Greening, D. W. (1997). Corporate social performance and organizational attrac-
tiveness to prospective employees. Academy of Management Journal, 40(3), 658–672.
The Power Grid: From a Technical
to a Finance Issue. Who Bears
the Financial Risk?

André B. Dorsman and Kees van Montfort

1 Introduction

Coping with the need to electrify and reach economic development most coun-
tries ended up with the structure of the industry which dominated the 20th century:
large, vertically integrated companies, with either statutory monopolies or significant
market power. However, the market is facing three developments that will determine
the structure of the future electricity industry worldwide no matter how regulators
or market forces might think, namely:
1. electricity cannot be stored in an economic efficient way;
2. marginal costs of electricity is nearly zero in case of solar and wind energy; and
3. realized supply can deviate from the expected supply, which can lead to negative
prices.
Although the industry is very busy trying to solve the storage problem, it is not
possible to store electricity now and in the near future in an economically accept-
able way. For most commodities, demand represents the willingness and ability to
pay for the services. This simple approach is complicated in electricity by several
considerations, but most importantly, by the need for balancing: demand and supply
at any time should be in balance, and the difference comes at a price: positive or
negative. This means that energy data about the user, his electricity consumption or
generation and the differences in time and location between them become a key and
valuable economic commodity too.

A. B. Dorsman (B)
VU University Amsterdam, Amsterdam, The Netherlands
K. van Montfort
Nyenrode Business Universiteit, Breukelen, The Netherlands
Amsterdam University of Applied Sciences, Amsterdam, The Netherlands
K. van Montfort
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 263
A. B. Dorsman et al. (eds.), Applied Operations Research and Financial Modelling
in Energy, https://doi.org/10.1007/978-3-030-84981-8_12
264 A. B. Dorsman and K. van Montfort

To keep the grid in balance, the TSO requires suppliers and demanders of energy
to give their bid and ask prices and volume one day ahead. The realized quantities of
renewable energy can deviate from the expected quantities one day earlier. In case it
is sunnier and windier than expected, the supply of renewable energy is higher than
expected and—due to the storage problem—can lead to negative prices.
The second development is that the upfront investments for a solar or wind park
are very high, but when these parks are running, the marginal costs are very low,
thereby pushing the suppliers of fossil fuels out of the market. Nearly zero marginal
costs have already begun transforming the electricity industry. They undermine a core
feature of the 20th-century industry: the overwhelming dominance of the wholesale
market, in turn, driven by the price of coal and gas. The small scale of production of
solar and wind energy is breaking up the market and deviates more and more from
a wholesale market.
The last development is the most characteristic one for the electricity market. The
value of future trades between parties is the product of price and quantity. Producers
of wind and solar energy don’t know precisely one day ahead what they shall deliver.
The promised quantity (one day ahead) can deviate from the realized quantity on
the delivery day. In other words, there is not only a price risk, but also a quantity
risk. Note that the producers of fossil fuel are (nearly) not confronted with this
problem. They can exactly deliver the promised supply. The switch from fossil fuel
to renewables means that the deviation between promised and realized supply—the
quantity risk—will grow and makes the task of the Transmission System Operator
(TSO) more difficult.
Commodities like oil, gas, and coal have to be transported from producer to
consumer. This is not the case with electricity. The grid knows a certain frequency;
in Europe 50 Hz in USA 60 Hz. Producing or consuming electricity means that the
voltage on the grid changes. The transmission system operator (TSO) has the task of
keeping the system in balance and maintaining the frequency on the electricity grid
on the level of 50 Hz, which means that demand and supply have to be the same at
every moment. The imbalance is the difference between the registered transactions
between trading parties per chosen time slot and the balance of the measured values
of a portfolio of the Balancing Responsible Party (BRP) per chosen time slot.1 In
case there is a producer in location A and a consumer in location B, electrons will not
go from A to B. However, when you produce or consume electricity, a future cash
flow will be created. To avoid that the grid will not be in balance (and the system
goes down) BRP’s have to give their price and quantity upfront. In other words, the
electricity market is a futures market.
ESMA (2014, 2020) reduced the number of exceptions of the guidelines of the
financial instruments and since Jan 3, 2018 also energy companies have to fulfill the
requirements of MiFID II. The guidelines of MiFID II apply for all EU countries.

1 There are several “markets” for balancing. There is a wholesale market between BRP’s. Further-
more, a bidding ladder organised by a TSO to support balancing. Thirdly there is an imbalance
account between the TSO and BRP’s which also could be considered to have the characteristics of
a market.
The Power Grid: From a Technical to a Finance Issue. Who Bears … 265

The consequences for the electricity market are substantially. However, the electricity
markets are not in all EU markets organized in the same way. For example, France has
a vertical integrated electricity market. In other EU countries the electricity market
knows more entities witch specific task in the market. The Netherlands is the most
unbundled market in the EU. In the EU we see a tendency of unbundling in the
electricity market. The consequences of MiFID II for the Dutch electricity market is
therefore important.2
The importance of the clearing is also shown in the Brexit-deal. This deal was
closed in December 2020. This deal contains appointments of an equal level playing
field for companies in the UK and the EU. However, most of the relevant issues for
the financial sector has to be further negotiated in 2021. Dobber (2021) writes that in
December 2020 there is only one of the approximately forty issues solved, namely:
the clearing.
Creating future cash flows means creating financial risk on both sides. This article
will focus on the question of who bears the financial risk on these future cash flows.
The electricity market is a commodity market, which means that there is a price risk
and a quantity risk. The uncertainty of the quantity deviates among the commodities.
The quantity risk of oil and gas is very limited. The difference between realized and
expected supply is nearly zero. Farmers sell their harvest against a fixed price and
deliver in the future. Weather conditions will influence the size of the harvest. In case
the size is less than expected, the farmer has to buy/sell futures contracts to eliminate
the expected position on the expiration moment. The difference between the market
for potatoes and the electricity market is the time interval. For potatoes, you see a
couple of weeks before the expiration moment, whether the size of the harvest will
be higher or lower than expected. The time interval for the electricity market is much
smaller and can be limited to an hour or even fifteen minutes. Another difference is
that in case not 100 but 99 potatoes are delivered, the market will not be harmed. The
condition for the electricity market is that at any moment, demand and supply have
been equal. Deviations in the supply have to be adjusted in one way or the other. In
other words, the quantity risk of electricity produced by wind and solar is high and
also for a short period unforeseen.
The TSO has a core position in the electricity market. She is responsible that at
any moment, the supply equals the demand. In case one of the participants has to
stop the activities, the TSO has to take steps to guarantee the delivery security of
electricity. Due to this task, the TSO—as a clearinghouse—bears a large financial
risk position. To decrease this position, the TSO has to check whether the margins
of the BRP’s are high enough. Moody’s wrote in her report of May 8, 2018 (page 8):
“(The Dutch TSO) TenneT’s A3 rating incorporates a two-notch uplift from its stand-
alone credit quality taking into account its ownership by the Dutch government and
the strategic importance to national energy policy”. In other words, Moody’s wrote
that she is fully aware that the Dutch government bears the risk of TenneT in case of
financial problems. A better understanding of the financial risks for TSO’s and other
participants in the electricity market is important.

2 Note that MiFID II is also important for CO-rights. However, that is not a topic for this study.
266 A. B. Dorsman and K. van Montfort

The contribution of this article to the literature is threefold. Firstly, as far as

the authors know there is not earlier described the financial risk positions in the
electricity market. Secondly, ESMA forces electricity companies to fulfill at the
requirements of MiFID II, which means that electricity companies have to deal
with their futures contracts like other entities, like banks and insurance companies.
Thirdly, the electricity market was traditionally a market for technicians. However,
the electricity market switches from a technical to a finance issue. Technicians are
underestimating the financial risk for the several parties on the electricity market
and especially for TenneT, that acts like a spider in the electricity web. We look at
the electricity market from a finance perspective. From that perspective TenneT is a
clearinghouse and the BRP’s are clearing members.3
The structure of the paper is as follows. In Sect. 2 we discuss the clearing of
stock markets. Section 3 contains a description of the price and quantity risk in the
electricity market. Understanding of the balancing market is necessary to get insight
in the financial risk positions of the parties in the electricity market. A description
of the balancing of the grid and its effects on the electricity price and risks will be
given in Sect. 4. The fact that on every moment, demand and supply have to be
equal can lead to extreme price effects. The main risk component is the counterparty
risk. Risk can be reduced by taking opposite risk positions. Therefore, we close the
contracts. However, when the counterparty fails to fulfill his obligations, the risk is
still back again. The options market has a system to reduce the counterparty risk to
(nearly) zero. Section 4 contains also this mechanism on the options exchange and
compares it to the situation on the electricity market to get a better understanding
of the counterparty risk. Section 5 will focus on the empirical part of our study.
Section 6 concludes and will also contain some remarks for further study.

2 Clearing

As already mentioned, ESMA forces electricity companies to deal with financial

instruments like banks and insurance companies. The same is true for the Dutch
TSO TenneT. The financial relations between TenneT and the BRP’s are futures and
the risks of those futures have to be covered. There is a lot of experience in the
stock market in this area. Therefore, we look in this section at the clearing for stock
markets.
At the end of the last century, Japanese banks became into trouble, and the Japanese
financial market was facing a credit crunch. A direct result was that during the finan-
cial crisis in 2007–2008, the FED, US central bank, and ECB in Europe supported
the banks to avoid a credit crunch. Fujao (2000) mentioned two factors behind the
financial crisis in Japan at the end of the previous century, namely (1) the crash on
the stock market and real estate market in the nineties of the last century and (2) the
loss of confidence in the accounting and auditing system in Japan. As an example

3 In Sect. 2 we define clearinghouse and clearing member.

The Power Grid: From a Technical to a Finance Issue. Who Bears … 267

of this last factor, Fujao refers to the Yamaichi Securities, which was hiding a loss
that was more than one-half of its equity capital. The regulators in Japan were not
able to discover this “misadministration”. In our view, it was also possible that the
regulator doesn’t want that Yamaichi Securities showed her weak position to keep
the confidence in the financial system. In other words, in the financial phrase, trust
is important.
Transactions on the Over-the counter-markets (OTC-markets) are normally bilat-
eral. There is always the possibility that at the end of the contract period, the coun-
terparty can’t or will not pay his debt. That is the so-called counterparty risk. In
2008 Lehman Brothers became bankrupt. Lehman Brothers was an important player
in the OTC market. Market participants that due to the OTC-contract, had a reduced
risk position saw that the risk position was revived by the bankruptcy of Lehman
Brothers. The first working day after the fall of Lehman Brothers the OTC market
was overloaded with contracts by those players. Also, nowadays, the counterparty
risk can be substantial. Fifteen financial institutions lost money in the fraud case of
Agritrade International (Singapore). Agritrade is a trade company in palm oil and
coal. Due to the lower growth rate in China in 2019, many coal mines in the USA
became bankrupt. ING was probably one of these victims of Agritrade and had to
take a large loss (See Wee, 2020, and Financieele Dagblad, Feb 29, 2020).
A regulated future market has normally a clearinghouse to reduce the counterparty
risk. Also, on the spot market a clearinghouse normally exists. At the end of the
19th century, the NYSE (New York Stock Exchange) had no clearinghouse. In that
time, many stocks were dual-listed on the NYSE and the CSE (Consolidated Stock
Exchange). The CSE had a clearinghouse. Bernstein et al. (2019) examined the effect
of the establishment of a clearinghouse at NYSE. They conclude that introducing
a clearinghouse reduced the volatility in the returns of the stocks listed on NYSE.
In other words, a clearinghouse attributes to financial stability. Acharya and Bisin
(2014) show that the absence of a clearinghouse creates the possibility of excess risk-
taking (see also Bliss and Steigerwald, 2006, and Kroszer, 2006). A political reaction
to the financial crisis of 2007/2008 was more regulations for the financial sector. One
of the effects was that the clearing in Denmark, Finland, and Sweden was no longer
an individual activity but was centralized by one clearinghouse, namely European
Multilateral Clearing Facility (EMCF). Menkveld et al. (2013) researched the effect
of this transition from a national clearing system to a centralized clearing system for
these three countries. At first, they mentioned that the transition was made in two
stages. In the first stage, the clearing was on a voluntary basis, while in the second
stage, the clearing by EMCF was mandatory. Menkveld et al. mentioned that on a
voluntary base, only 1% of the market participants used EMCF, which means that the
market participants don’t believe that a better clearing system is creating value for
them. Furthermore, they found that due to stringent collateral requirements, central
clearing reduces volatility.
The creation of a central counterparty reduces the counterparty risk and creates
value for the market participants. However, when a “systemic” institution as the
central counterparty goes in default, the market structure becomes “under tension”.
In 1974 a clearinghouse, the Caisse de Liquidation des affaires en Marchandises
268 A. B. Dorsman and K. van Montfort

(CLAM), in Paris became in such a situation. Bignon and Vuillemey (2020) analyzed
this failure. They focused on two points, namely: (1) identifying the risk manage-
ment failures that caused the default and (2) the creating of new agency problems
when supervisors have discretionary powers over debt structure for a failed entity.
Analyzing the first point Bignon and Vuillemey (2020) identified two factors.
Firstly, the CCP members (clearing members) had crowded exposures from finan-
cially vulnerable clients. In other words, weak clients had open positions and the
clearing members guaranteed these positions, and the clearinghouse guaranteed the
clearing members and bore unacceptable risks.
Secondly, the clearinghouse was not able to prevent the growth of large positions
of a clearing member. In other words, the risk positions of the clearing members
increased, and the clearinghouse had no means to stop this process.
In the CLAM case, the managers didn’t release the information that she was in
default when the largest clearing member could no longer respond to the margin
calls. In fact, they hoped that the sugar price should improve and that the losses
should decrease. The benefits of this action are for the equity holders of CLAM,
while the risk of a further price decrease of sugar is for the debtholders of CLAM.
Due to the postponement of releasing the information of the actual situation of the
financial position of CLAM there is a risk-switch from shareholders to debtholders
of CLAM.
Kupiec (1989) found no empirical evidence of a relationship between margin
requirements and the volatility of the S&P 500 index portfolio’s excess returns.
Hardouvelis and Theodossiou (2002) looked at the influence of a change in the
initial margin requirements on the stock volatility across bull and bear markets. They
conclude that it is a prudential rule for setting margins to lower them in sharply
declining markets in order to enhance liquidity and avoid a depyramiding effect in
stock prices but subsequently, rise them and keep them at a higher level in order to
prevent a future pyramiding effect. In the literature the pyramiding effect is described
as the process that optimistic investors can borrow stocks with a high debt level. In
a bullish market investors are optimistic and may make use of leverage when the
(initial) margins are low. The opposite is that when the market turns to be more
bearish, the (initial) margin requirements are becoming more severe and therefore
damaging the liquidity of the market.
Deng and Oren (2006), after looking at the roles of several electricity derivatives in
mitigating market risk and structuring hedging strategies for several market partic-
ipants, emphasize the importance of standardization of these contracts and avoid
exotic contract forms that can meet specific needs for hedging and speculation.
As we will discuss later more in detail, TenneT has the role of the clearinghouse
in the electricity market in the Netherlands and is responsible that at every moment,
demand and supply on the grid are the same. In case a PV-party (clearing member)
has a shortage or a surplus, TenneT will at the expense and the risk of that party buy
or sell electricity. For that reason, TenneT can buy and sell electricity in the intraday
market. Due to the energy switch from fossil fuel to renewables, the instability of
the grid has been increased and the intraday market has become more important.
Comparing to the stock market there is an additional problem for the electricity
The Power Grid: From a Technical to a Finance Issue. Who Bears … 269

market. On the stock market we know exactly the quantity of the underlying value
of a contract. However, there is a quantity risk in electricity contracts as we will see
in the next section.
The electricity market is a concrete market and deviates from abstract markets,
like the stock market and the bond market. The lack of storing possibilities and the
necessity that at every moment the demand and supply has to be the same, makes
that outages can have immediately consequences on the (volatility of the) electricity
price. Margin requirements are therefore very important for this market. Due to
the outages—which are not uncommon in the electricity market—it is not easy to
predict the price and quantity risk, which we will discuss in the next section. See
also Montfort et al. (2020).

3 The Introduction of the Price and Quantity Risk

in the Electricity Market

The liberalization of the electricity market introduced a variable price depending

on the demand and supply of the market. In other words, we had an introduction to
price risk. In case the consumer produces his own electricity, he is price-indifferent. A
separation between consumer and producer creates price risk for both market parties.
A higher price is good for the producer, but bad for the consumer and vice versa.
How more market parties between the (end) producer and the (end) consumer, the
more market parties are facing price risk. However, due to the switch from fossil
fuel to renewables, we have not only a price risk but also an increasing quantity
risk. A TSO system expects parties to submit their supply and demand for a given
time frame for the next day. The real supply of the next day can deviate from the
promised supply. For example, if it was less sunny and/or windy than the day before
was expected, less electricity produced with solar and wind energy will be delivered
than the day before was expected. Larger deviations between expected and realized
supplies will cause an unstable balancing mechanism. As mentioned, compared to
solar and wind energy electricity produced with fossil fuel can be estimated very
accurately. The switch from fossil fuel to renewables (solar and wind) doesn’t only
mean a substantial increase in the quantity risk.
The small producers of wind- and solar energy will not reduce their supply when
the price decreases; neither will they increase their supply when the electricity price
increases. In fact, the relationship is exactly the opposite. These small producers will
increase (decrease) the supply when the electricity price decreases (increases), which
makes a new price equilibrium—the price where demand and supply are equal—more
difficult. In other words, the access of small producers of renewable energy on the
market reduces the stability of the price equilibrium. On the other hand, the entrance
of many small producers in a market with a small number of large fossil producers
makes it less possible that a small number of market participants can set the price.
270 A. B. Dorsman and K. van Montfort

As stated before, the issue at stake in this article is whether or not the TSO bears
ultimate financial responsibility for demand and supply being equal at every moment
of the day. If that would be the case and in case demand and supply by BRP on day
t deviate from their promised demand and supply on day t-1, the TSO will buy/sell
electricity on day t on the expense and risk of the BRP’s. In other words, at first, the
BRP’s bear the financial risk. However, in case a BRP can’t fulfill his obligations,
it is the TSO that bears the financial risk. The better the financial positions of the
BRP’s are, the less the financial risk for the TSO is.
Another point that is important for the financial strength of a system is the number
of participants. When there are many participants in the system, the system can easily
survive the financial collapse of one of the participants. However, when the number
of participants is limited, impact of the collapse of one of the participants will be
larger. The Dutch electricity market knows four dominating companies that generate
electricity on a large scale, namely: EON, Engie, Eneco, and Nuon/Vattenfall. In
fact, these producers are system-players. If one of these players fails to fulfill her
obligations, the security of supply could be at stake with physical and financial
consequences. In that case, the Dutch TSO (and her owner, the Dutch state) can be
compelled to take over the position of the failing company.
Quantifying the risk components
The value of a contract (V) depends on price (p) and quantity (q), V = p q. Value
risk is the difference between realized value and expected value.

VR = pr qr − pf qe

where
VR = value risk
pr = realized price
qr = realized quantity
pf = fixed price
qe = estimated quantity.
The above expression of VR can be rewritten in:
   
VR = pr qr − pf qe = pr qr − pr qe + pr qe − pf qe = pr qr − qe + qe pr − pf

In words: the value risk is the sum of the realized price pr times the quantity risk
(qr −qe ) plus the estimated quantity qe times the price risk (pr −pf ). In case there is
no quantity risk (qr = qe ), the value risk consists only of the price risk times the
quantity. Due to the deviations between the realized quantity on the next day and
the estimated quantity one day earlier, the value risk has a second component: the
realized price times the quantity risk. In the empirical part of this paper we will focus
on quantity risk.
The Power Grid: From a Technical to a Finance Issue. Who Bears … 271

4 Balancing in the Electricity Market and Counterparty

Risks

Balancing is crucial for power markets to function properly and financially adequate.
Electricity requires physical electricity balance management. When more power is
being produced than consumed, the created imbalance affects the stability of the
power system as a whole. As earlier said, supply and demand need to be in balance
at any time. This balance only materializes when production equals consumption.
This balance is crucial in preventing outages and the disturbance of system frequency
of 50 ± 0.1 Hz in Europe and 60 ± 0.1 Hz in the USA. Demand and supply need to
be matched as long as electricity can’t be stored in an economically viable way.
In our view, balance management is a physical power system operation and a
financial service vital for ensuring security of supply through making equal the real-
time balancing of mutual financial obligations of the markets parties involved in the
balancing of power demand and supply. If the system is out of balance ultimately,
power stability and quality will deteriorate, which could ultimately cause power
blackouts.
In vertically integrated electricity systems, it is relatively easy to maintain the
power system in balance. However, unbundling in European energy markets since the
beginning of this century has separated the various functions connecting demand and
supply. This requires delving into (financial) balance management and its financial
consequences.
We consider the balancing market as an arrangement to be considered the last in a
sequence of electricity markets, after year-ahead, month-ahead, day-ahead, and intra-
day markets. Balancing encompasses both financial transactions (the power market)
and physical exchanges (the power system). The electricity system is a peer-to-peer
transaction-based system (see Nobel, 2016). We will concentrate in this article on the
risks of balancing. As long as the producer and the consumer are directly connected to
each other and can settle their deals directly with each other, there is no problem. As
long there is in one service area a system for balancing, which guarantees the tuning
and settlement between parties in that one area, there still is no problem. The problem
will arise the moment there are no settlements with connected areas that really reflect
the transactions and calculations between production and consumption. In such a
connected area, there could be more production than consumption, and the surplus
will, according to the laws of physics, search for the path of the least resistance. Such
a surplus can be balanced by the neighbouring balancing authority or market party if
technically and financially, the neighbours can cope with the imbalance. If not, there
is a serious system problem with direct physical and financial consequences.
Balancing market parties, all free in their choice to produce electricity or to
consume electricity, has financial consequences. Those market parties enter into
energy contracts. This results not only in need of a physical balance obligation
but also in future financial contracts (and financial risk positions). The execution
of contractual electricity includes a financial settlement. This process of balancing
requires financial security.
272 A. B. Dorsman and K. van Montfort

In most ‘ordinary’ economic sectors an imbalance in the contractual obligations—

when supply and consumption do not match in time—normally has no direct conse-
quences for third parties. Imbalance at the electricity market directly affects third
parties because of the power failures in the electricity system of a region, a count(r)y
or a continent.4 Making sure that all electricity contracts can be met both physically
and financially is crucial.
Balancing in a decentralized sustainable and unbundled power system raises three
key questions:
1. Who is responsible for the imbalance and the associated costs?
2. Who pays which costs incurred?
3. Who receives which payments?
In the new decentralized and sustainable world of electricity with no longer “one”
strong (fossil) machine it is needed to answer the question who is going to pay for
what. The electricity market is developing from an engineering issue to a financial
issue, while the financial structure is not yet clear. To mention one point, who will be
the leading regulator? Up to now, in the Netherlands, the ACM (Authority Consumer
Markets) is the regulator of the energy sector. This is based on the twentieth century
idea that the relation between producer and consumer is a physical one, like the food
industry. However, the switch from a technical issue to a finance issue has also the
consequence that the AFM (Authority Financial Markets) also has to play a role in
this area and, in our view, has to be the leading regulating authority.
Official future markets exist mainly because of their task in helping to reduce
counter-party risk. This is done by placing an intermediate institution, the clearing-
house, between buyer and seller. The clearinghouse will guarantee both buyer and
seller that the transaction will take place according to the contract (see also Kolb,
2000). To make sure that this will not lead to a too big risk for the clearinghouse itself,
both buyer and seller will have to deposit an initial margin with the clearinghouse.
Only after both parties have done so, the transaction will be officially registered in
the market. At that moment, the buyer will have a long position and the seller a short
position.
After the transaction has taken place, the price of the underlying value will start to
fluctuate. These fluctuations can either be favorable or negative for the buyer as well
as the seller of the contract. A positive development for the buyer implies a negative
price change for the seller and vice versa. When the price development is unfavorable

4 A dispute between Serbia and Kosovo has disrupted the electric power grid for most of the
Continent, making certain kinds of clocks—many of those on ovens, in heating systems and on
radios—run up to six minutes slow. The slowdown began in mid-January 2018, and since then
clocks in 25 countries, from Poland to Portugal and Denmark to Turkey have lost time. The
fluctuation in the power supply is infinitesimally small—not nearly enough to make a meaningful
difference for most powered devices—and if it was a brief disturbance, the effect on clocks might
be too little to worry about. This disruption was caused by a discussion about the settlement of
113 GWh according to ENTSO-E, the association of European transport system operators, see
https://www.entsoe.eu/news/2018/03/06/press-release-continuing-frequency-deviation-in-the-con
tinental-european-power-system-originating-in-serbia-kosovo-political-solution-urgently-needed-
in-addition-to-technical/.
The Power Grid: From a Technical to a Finance Issue. Who Bears … 273

for the buyer, it is possible that a point is reached where the initial margin is no longer
sufficient. The buyer will then have to pay an additional margin. The level of margin
at which buyer or seller will have to pay an additional margin is called maintenance
margin. The additional amount that buyer or seller needs to pay is called the variation
margin itself. Normally we see that the maintenance margin will be approximately
75–85% of the initial margin. This implies that a reduction of the value of the initial
margin with 15–25% will lead to a margin call by the clearinghouse. We will normally
see that this margin call will be such that the addition to the margin, the variation
margin, will at least restore the original situation. When in this case (unfavorable
development for buyer) the buyer will or can’t deposit an additional margin, the
clearinghouse will close the position on his/her behalf using the money in the margin
account. The clearinghouse will, of course play a similar role if the price development
is unfavorable to the seller. The clearinghouse will settle daily and mark the margin
accounts to market. This is done so as to make sure that additional margin will be paid
in time. This will help ensure the proper functioning of the official forward market. It
is also possible that—when price developments are very favorable for one party—the
clearinghouse will allow him or her to reduce the margin account by extracting part
of the sum initially deposited.
On the Dutch option market, the market maker has an obligation to give a bid-ask
spread. He will buy against the bid price and sell against the ask price. The market
maker deals exclusively with his own account and risk. (He may not accept orders
from others). To be sure that the market maker will pay his debts, he needs the ful
financial support of a clearing member, mostly an affiliate of a bank. The clearing
member is mostly also financing the financial position of the market maker. The
clearing member guarantees to the clearinghouse the market maker. If the clearing
member is not willing to guarantee the market maker any longer, he has to inform
the clearinghouse and the market maker is not allowed to deal on the option market
any longer.5
The circumstances on the Dutch electricity market are comparable with the Dutch
options market. The TSO ultimately has the functional role of the clearinghouse. All
cash flows related to electricity ultimately, ends up at the TSO. In the balancing mech-
anism the TSO is the counterparty of each other BRP. To reduce her risk, she accepts
a limited number of clearing members, called BRP’s. Every BRP is fully financially
responsible for the imbalance position of her clients. In other words, the BRP guaran-
tees her client (supplier or demander) to a certain extent. The supplier or demander on
the electricity market is like the market maker on the options exchange. The compar-
ison is not 100%. The market maker on the options exchange tries to get a zero position
in the underlying value. He is a trader, not an investor, while at the electricity market,
the supplier or demander is not a trader, but a (short/long) investor. However, from the
risk perspective, the situation on the options and electricity markets is comparable.

5In 1998 one of the clearing members of the Dutch option market, ING Clearing, was facing
solvency problems. ING clearing was an affiliate of the Dutch Insurance bank ING. Although it
was a 100% affiliate of ING, ING didn’t guarantee ING clearing. Due to the financial problems of
ING Clearing the survival of the whole Dutch options market was at stake.
274 A. B. Dorsman and K. van Montfort

Option market

Buyer=> clearing member =>clearinghouse => clearing member => Seller

Electricity market

Demander => BRP => TSO (TenneT) => BRP => supplier

Scheme 1 Relations on the option market and the electricity market

In Scheme 1, we see in the top segment the order flow in the options market. The
order of the buyer will be sent to the market (clearinghouse). The same is true for
the seller’s side. However, you have to pay upfront the margin. Nevertheless, we see
that in some cases, the clearinghouse can come into serious troubles. In the FSB
discussion paper (FSB, 2018) three defaults are mentioned after World War II. “In
1974, the Caisse de Liquidation in Paris, due to default on margin calls when sugar-
futures prices fell sharply; in 1983, the Kuala Lumpur Commodities Clearing House,
when half a dozen large brokers defaulted following a crash in palm-oil futures; and
most seriously, the Hong Kong Futures Exchange clearing house, in the wake of
1987’s global stock market crash.” The supervisory authorities don’t always take the
right steps and sometimes make the situation worse (See, for example, Dorsman &
Buckley, 2001). After these crises, the golden rule on the financial markets is now:
No margin means no deal.
In the down segment of Scheme 1, we see the order relation on the electricity
market. The BRP’s brings the orders of the demanders and the suppliers to the
market (TenneT). In contrast with the options market, there is no golden rule. This
means a high-risk position for TenneT.
In summary, the core role in these risk systems are the clearinghouse in the options
market and the TSO in the electricity market. Both institutions, clearinghouse, and
TSO, try to reduce their counterparty risk by fully guaranteeing the clearing members
and BRP’s respectively. However, what will happen when a large clearing member
or BRP is facing insolvency problems? In the Dutch electricity markets, we have
four “system” players (BRP’s), namely Eneco, Vattenfall/Nuon, Engie, and EON.
When one of them came into financial troubles, the TSO (read: the 100% owner of
TSO the state, or in other words the taxpayer) has to pay.
For the financial markets the requirements of margining are clear. The European
Commodity Clearing has published a report ECC (2019) and writes on page 19 that
the clearinghouse has to create a default fund that has sufficient resources to cover
losses arising out of a default of two clearing members to whom the ECC has the
largest exposure to. Also the Dutch Central Bank (DNB) mentioned in (DNB, 2018,
p. 17) this requirement for the default fund for the Clearinghouse. In case we apply
this rule for the electricity market in the Netherlands, TenneT has to create a default
fund that is large enough to cover the damage in case of the collapse of two of the
The Power Grid: From a Technical to a Finance Issue. Who Bears … 275

four system BRP’s. Such a fund will be extreme large and therefore very expensive.
However, not having such a default fund means that the consequences of the risks
for TenneT will not be reduced.

5 Empirical Quantification of Risk

First, in this section, we quantify the quantity risk (qr −qe ): the difference between the
realized quantity and the expected quantity of electricity. For that, we use an open data
source of TenneT (https://www.tennet.org/english/operational_management/export_
data.aspx). As we already mentioned, TenneT is a large European electricity transmis-
sion system operator (TSO) with activities in the Netherlands and in Germany. They
strive to ensure a reliable and uninterrupted supply of electricity in a high-voltage grid
for some 41 million people. Within the open data source of TenneT we choose the
dataset Imbalance, which reports on a quarterly base the following information over
the year 2018: imbalance quantity (kWh), estimated quantity (kWh), realized quan-
tity (kWh), fixed price (e), realized price (e), upward incidence reserve (yes/no),
etc. In Table 1, several descriptive statistics of the imbalance quantity, the absolute
value of the imbalance quantity, the realized quantity, and the ratio of the absolute
imbalance quantity and the realized quantity are presented.
From the first column of the above table, it follows that during 2018 the quarterly
difference between the realized and estimated imbalance quantity of TenneT was, on
average only −998 kWh and in total about −35 gWh. This means that on average
the quarterly quantity shortages were rather small, e.g., less than 1%. However,
the maximum and minimum quarterly imbalance quantities range from −355,233 to

Table 1 Descriptive statistics of the imbalance quantity, the absolute value of the imbalance
quantity, the realized quantity, and the ratio of the absolute imbalance quantity and the realized
quantity (on a 15-min base; over the year 2018; in The Netherlands)
Imbalance Negative Positive Absolute Realized Ratio of
quantity imbalance imbalance imbalance quantity (kWh) absolute
(kWh) quantity quantity quantity (kWh) imbalance
(kWh) (kWh) quantity and
realized
quantity
Mean −998 −40,868 38,709 39,786 148,969 0.269
Median 125 −33,142 31,910 32,545 140,786 0.235
Standard 51,489 31,789 29,145 32,695 53,154 0.190
deviation
Maximum 230,784 0 230,784 355,233 540,735 0.988
Minimum −355,233 −355,233 0 0 30,179 0.000
Number of 35,040 17,556 17,484 35,040 35,040 35,040
quarters
Sum of −35,000,000 −715,000,000 680,000,000 1,390,000,000 5,220,000,000 0.266
2018
276 A. B. Dorsman and K. van Montfort

230,784 kWh per quarter. These values are extremely high compared to the quarterly
realized quantities. Furthermore, it is not so important what the mean of the imbalance
quantity is because negative imbalances and positive imbalances cancel each other
out However, the TSO is the magnitude of the imbalance significant and not facing
too much or too little electricity.
In the columns 2 and 3 we present information in case of a shortage of electricity
(column 2) or a surplus of electricity (column 3). We see that in 17,556 quarters (of 15-
min) there is a shortage and in 17,484 quarters a surplus of electricity. Furthermore,
the absolute value of the sum of the negative imbalance quantities (715,000,000 kWh)
is greater than the absolute value of the sum of the positive imbalance quantities
(680,000,000 kWh). As a consequence, the annual electricity shortage is 34,000,000
kWh (or about 0.7% of the sum of the realized quantity in 2018).
The last column of Table 1 indicates the ratio between the absolute values of the
imbalance quantities and the realized quantities (on a quarterly base). It turns out
that these ratios are, on average, about 0.269 with a maximum and a minimum of
0.988 and 0.000. In other words, on average the difference between the quarterly
realized and estimated quantity is about 27% of the realized quantity. The value of
the standard deviation indicates that this value does not vary that much over the year.
Furthermore, during the year, several times, the emergency power has to be
switched on by TenneT. Such a step can be caused for various reasons and is often an
“extended period” then deployed to a quarter. The cause of these persistent problems
may be different, occasionally (e.g., failure or malfunction of units or components) or
correlated forecast errors, for example, lack of wind production in multiple locations.
Often the market solves such incidents, but if that is not enough, TenneT intervenes
with emergency power. These are products that were bought and used by contractors
making withhold volumes then available (with the appropriate pricing). At that time,
TenneT sends a general signal to all parties that TenneT “needs support” (everyone
can contribute and create imbalances the right direction) and the financial part is
generally attractive.
In 2018 during 195 quarters, TenneT had to charge the emergency power (0.5% of
all the quarters). In total, there were 19 emergency periods of one or more quarters.
Often in the first quarter of such an emergency period the imbalance ratios were
extremely high with values above 0.50. In addition, TenneT indicates that these
values of the imbalance ratios often go hand in hand with technical problems of the
power supply.
Figure 1, shows the mean imbalances (kWh) per hour over the day (i.e., 24 h)
with data for the year 2018. Roughly speaking, during the off-peak hours (8.00 am–
8.00 pm) there are no shortages of electricity. In this period of the day, the electricity
prices and the electricity volumes are relatively low. On the opposite, during peak
hours (8.00 pm–8.00 am) the electricity prices, and the electricity volumes are rela-
tively high. During that period, the imbalances take positive values. In other words,
between 8.00 am–8.00 pm, the expected electricity volumes are estimated too low,
where between 8.00 pm–8.00 am, the expected electricity volumes are estimated too
high. Exceptions are around 4.00–9.00 pm.
The Power Grid: From a Technical to a Finance Issue. Who Bears … 277

10000

5000

2:00 PM

9:00 PM
12:00 AM

7:00 AM

1:00 PM

3:00 PM
4:00 PM
5:00 PM
6:00 PM
7:00 PM
8:00 PM

10:00 PM
11:00 PM
1:00 AM
2:00 AM
3:00 AM
4:00 AM
5:00 AM
6:00 AM

8:00 AM
9:00 AM
10:00 AM
11:00 AM
12:00 AM
-5000

-10000

-15000

Fig. 1 Mean imbalances (kWh) per hour over the day (mean values in peaked line; standard errors
in smoothed line)

An interesting question is whether the distribution of the negative imbalances

deviates from the distribution of the positive imbalances. In Fig. 2, we present both
distributions during the day. The top graph is the graph of the positive quarterly

60000

40000

20000

0
1am
2am
3am
4am
5am
6am
7am
8am
9am
10am
11am
12am
1pm
2pm
3pm
4pm
5pm
6pm
12pm

7pm
8pm
9pm
10pm
11pm

-20000

-40000

-60000

Fig. 2 Mean imbalance (kWh) per quarter over the day (highest line: only positive imbalances;
second line: all imbalances; lowest line: only negative imbalances; vertical axe: kWh; horizontal
axe: quarter of the day over 24 h)
278 A. B. Dorsman and K. van Montfort

imbalances and the down graph of the negative quarterly imbalances. The graph in
the midst contains the distribution of all quarterly imbalances over the day. In fact,
this graph represents the differences between the quarterly surplus and quarterly
shortages. From the graph with all imbalances, it follows that during the morning
between 6–8 am, during the end of the afternoon between 4–5 pm, and during the end
of the day between 10–12 pm, the values of the quarterly imbalances vary extremely,
i.e., very positive values and very negative values. In these timeslots, the electricity
consumption is relatively high, and it is difficult to forecast this consumption on a
quarterly base.
After the financial crisis of 2007/2008, the policy of the European Union (EU) is
that an insolvency problem of a bank may no longer harm the tax-payer. It is logical
to have the same policy for the electricity markets.

6 Conclusion

Trust is essential for financial markets. In Sect. 2, we discussed the clearing of the
positions of market participants. No trust meant at the end of the ninetieth century
a credit crunch in the Japanese market. The default of Lehman Brothers, a large
market player in the OTC market, recreated financial positions that were assumed to
be hedged. In case of a crisis, the political reaction is mostly: more rules. That was
also the case in the Nordic countries after the financial crises of 2007/2008. Market
participants were not against the new rules but believed that it was better for others
than for themselves.
The creation of a clearinghouse that asks initial and additional margins will reduce
the counterparty risk and will increase the liquidity of the market. However, when
the survival of a clearinghouse is at stake, the impact on the market is larger and a
system risk threats. In other words, creating a system of clearinghouse and clearing
members is positive for the market conditions, but you have to avoid a situation
where you replace the counterparty risk for individual market participants through the
counterparty risk of the clearinghouse. Initial and additional margins of a substantial
level are a must to avoid a system default.
In this chapter we postulate that the electricity market is a financial market with
the TSO as clearinghouse and the BRP’s as clearing members. ESMA requires from
participants, and especially clearinghouses, in the financial market that they reduce
their risk position. One of the requirements of ESMA is that the margins are on a
certain level. At this moment the margin requirements of TenneT for the BRP’s are
very limited and nearly zero.
In the Netherlands, there are four system players: EON, Engie, Eneco, and
Nuon/Vattenfall. A collapse of one of these system players can harm the system.
Therefore, it is necessary that—like the banking sector—the margins of system
players have to higher than for non-system players. The requirements for system
banks are that they have a minimum limit of equity. This limit depends on the assets.
How riskier an asset is, how more equity as a buffer is required. In the electricity
The Power Grid: From a Technical to a Finance Issue. Who Bears … 279

market, the underlying value—electricity—is unique and has a certain risk level. A
link with risk weighted assets is therefore not necessary. We propose a link with
the size of the BRP. Larger BRP’s have to give a larger collateral to TenneT than
smaller BRP’s. In the case of a system player we propose to increase this margin
substantially. It is not easy to rank companies by size. TenneT also has difficulty with
this, see netcode (network code) 10.8.6 Further study is necessary to define the size
of the BRP’s and to determine the margin x.
In the electricity market in the Netherlands the distinguished entities fulfill sepa-
rate tasks. In other words, there is not a vertical integrated market (like France). The
requirements of ESMA are in a non-vertical integrated market more important than
for vertical integrated markets. However, we see in Europe a tendency to disintegra-
tion which means that the developments in the Dutch electricity market are important
for other European electricity markets.

References

Acharya, V., & Bisin, (2014). Counterparty risk externality: Centralized versus over-the-counter
markets, Journal of Economic Theory, 149, 153–182.
Authority Consumers and Markets (2016). Netcode elektriciteit (Netcode electricity).
Bernstein, A., Hughson, E., & Weidenmier, M. (2019). Counterparty risk and the establishment of
the New York Stock Exchange clearinghouse. Journal of Political Economy, 127(2), 689–729.
Bignon, V., & Vuillemey, G. (2020). The failure of a clearinghouse: Empirical evidence, Review of
Finance (pp. 99–128).
Bliss, R. R., & Steigerwald, R. S. (2006). Derivatives clearing and settlement: A comparison of
central counterparties and alternative structures. Economic perspectives. Federal Reserve Bank
of Chicago.
Deng, S. J., & Oren, S. S. (2006). Electricity derivatives and risk management, Energy, 940–953.
DNB (2018). De CCP als spil in het financieel systeem (The Central clearing party—the
clearinghouse—as pivot in the financial system).
Dobber, J. (2021). De Londense City is vergeten in de nieuwe brexitdeal. (They forgot the London
city in the new Brexit deal). Het Financieele Dagblad (The Financial Newspaper), 11 Jan 2021.
Dorsman, A. B., & Buckley, A. (2001). The Amsterdam Options Exchange in 1998: How the
Supervisory authorities turned a problem into a crisis. European Management Journal, 19(3),
286–290.
ECC (2019). ECC margining. 26 Aug 2019. version 1.8.6.
ESMA (2014). Consultation paper MIFID II/MIFIR.
ESMA (2020). MiFID II: C6 energy derivative contracts and the EMIR requirements.
Financieele Dagblad (2020). ING slachtoffer fraude Singaporees bedrijf (ING victim Singaporean
company, 29 Feb.
FSB (2018). Discussion paper, Financial resources to support CCP resolution and the treatment of
CCP equity in resolution. 15 Nov 2018. http://www.fsb.org/wp-content/uploads/P151118-2.pdf
Fujao, M. (2000). Recapitalizing Japan’s Banks: The functions and problems of financial
revitalization act and bank recapitalization act. Keio Business Review, 38, 1–16.

6Decision of the Authority for Consumers and Markets (ACM) of 21 April 2016, reference
ACM/DE/2016/202151, laying down the conditions as referred to in Article 31 of the Electricity
Act 1998 (Network Code Electricity).
280 A. B. Dorsman and K. van Montfort

Hardouvelis, G. A., & Theodossiou, P. (2002). The asymmetric relation between initial margin
requirements and stock market volatility across bull and bear markets. The Review of Financial
Studies, 15(5), 1525–1559.
Kolb, R. W. (2000). Futures, Options, & Swaps, Blackwell, publishers.
Kroszer, R. S. (2006). Central counterparty clearing: History, innovation and regulation. Economic
perspectives. Federal Reserve Bank of Chicago.
Kupiec, P. H. (1989). Initial margin requirements and stock returns volatility: Another look. Journal
of Financial Services Research, 3, 287–301.
Menkveld, A. J., Pagnotta, E., & Zoican, M. (2013). Does Central Clearing Affect Price Stability?
Evidence from Nordic Equity Markets (27 Oct 2015). Available at SSRN: https://ssrn.com/abs
tract=2350762 or https://doi.org/10.2139/ssrn.2350762
Montfort, K. V., Thorsdottir, H., & Bohnsack, R. (2020). Optimizing resource usage in an unobtru-
sive way through smart integration: The case of electric vehicle charging in Amsterdam, In A.
Dorsman, Ö. Arslan-Ayaydin, & ThewissenJ (Eds.),Regulations in the energy industry. Springer.
Moody’s Investor services (2018). Tennet Holding, available on the website of TenneT (8 May).
Nobel, F.A. (2016). On balancing market design, thesis, Technical University Eindhoven.
Wee, D. (2020). ING alleges father-son fraud after commodity firm collapses, Bloomberg. 27 Feb
2020 and updated 28 Feb 2020.