Jesco Humpola

Gas Network Optimization

by MINLP

λογος
 Gas Network Optimization
by MINLP

vorgelegt von
Dipl.-Math. Jesco Humpola
aus Melle

von der Fakultät II – Mathematik und Naturwissenschaften

der Technischen Universität Berlin
zur Erlangung des akademischen Grades

Doktor der Naturwissenschaften

– Dr. rer. nat. –

genehmigte Dissertation

Promotionsausschuss:

Vorsitzender: Prof. Dr. Peter Bank

Gutachter: Prof. Dr. Dr. h.c. mult. Martin Grötschel
Gutachter: Prof. Dr. Thorsten Koch
Gutachter: Prof. Andrea Lodi, Ph.D.

Tag der wissenschaftlichen Aussprache: 06. November 2014

Berlin 2014
D 83
Bildnachweis Einband Vorderseite:
Mit freundlicher Genehmigung von Open Grid Europe GmbH, Essen.

Bibliografische Information der Deutschen Nationalbibliothek:

Die Deutsche Nationalbibliothek verzeichnet diese Publikation in der

Deutschen Nationalbibliografie; detaillierte bibliografische Daten sind im
Internet über http://dnb.d-nb.de abrufbar.

©Copyright Logos Verlag Berlin GmbH 2017

Alle Rechte Vorbehalten.

ISBN: 978-3-8325-4505-5

Logos Verlag Berlin GmbH

Comeniushof, Gubener Str. 47,
10243 Berlin
Tel.: +49 (0)30 42 85 10 90
Fax: +49 (0)30 42 85 10 92
INTERNET: http://www.logos-verlag.de
Abstract

One quarter of Europe’s energy demand is provided by natural gas distributed

through a vast pipeline network covering the whole of Europe. At a cost of 1 million
Euros per kilometer the extension of the European pipeline network is already a
multi billion Euro business. The challenging question is how to expand and operate
the network in order to facilitate the transportation of specified gas quantities at
minimum cost. This task can be formulated as a mathematical optimization problem
that reflects to real-world instances of enormous size and complexity. The aim of this
thesis is the development of novel theory and optimization algorithms which make it
possible to solve these problems.
Gas network topology optimization problems can be modeled as nonlinear mixed-
integer programs (MINLPs). Such an MINLP gives rise to a so-called active trans-
mission problem (ATP), a continuous nonlinear non-convex feasibility problem which
emerges from the MINLP model by fixing all integral variables. The key to solving
the ATP as well as the overall gas network topology optimization problem and the
main contribution of this thesis is a novel domain relaxation of the variable bounds
and constraints in combination with a penalization in the objective function. In case
the domain relaxation does not yield a primal feasible solution for the ATP we offer
novel sufficient conditions for proving the infeasibility of the ATP. These conditions
can be expressed in the form of an MILP, i.e., the infeasibility of a non-convex NLP
can be certified by solving an MILP. These results provide an efficient bounding
procedure in a branch-and-bound algorithm.
If the gas network consists only of pipes and valves, the ATP turns into a passive
transmission problem (PTP). Although its constraints are non-convex, its domain
relaxation can be proven to be convex. Consequently, the feasibility of the PTP can
be checked directly in an efficient way. Another advantage of the passive case is that
the solution of the domain relaxation gives rise to a cutting plane for the overall
topology optimization problem that expresses the infeasibility of the PTP. This cut is
obtained by a Benders argument from the Lagrange function of the domain relaxation
augmented by a specially tailored pc-regularization. These cuts provide tight lower
bounds for the passive gas network topology optimization problem.

iii
 The domain relaxation does not only provide certificates of infeasibility and
cutting planes, it can also be used to construct feasible primal solutions. We make
use of parametric sensitivity analysis in order to identify binary variables to be
switched based on dual information. This approach allows for the first time to
compute directly MINLP solutions for large-scale gas network topology optimization
problems.
All the research in this thesis has been realized within the collaborative research
project “Forschungskooperation Netzoptimierung (ForNe)”. The developed software
is in use by the cooperation partner Open Grid Europe GmbH.
Parts of this thesis have been published in book chapters, journal articles and
technical reports. An overview of the topics and solution approaches within the
research project is given by Martin et al. (2011) and Fügenschuh et al. (2013).
Gas network operation approaches and solution methods are described in detail by
Pfetsch et al. (2014) and with a special focus on topology optimization in Fügenschuh
et al. (2011). The primal heuristic presented in this thesis is published by Humpola
et al. (2015a). The method for pruning nodes of the branch-and-bound tree for an
approximation of the original problem is described in Humpola and Fügenschuh
(2015) and Humpola et al. (2015b) and Humpola and Serrano (2017). The Benders
like inequality is introduced by Humpola et al. (2016).

iv
Zusammenfassung

Ein Viertel des europäischen Energiebedarfs wird durch Gas gedeckt, das durch
ein europaweites Pipelinesystem verteilt wird. Aufgrund von Ausbaukosten von
1 Mio. Euro pro Kilometer ist der Netzausbau ein Milliardenunterfangen. Die größte
Herausforderung besteht darin zu entscheiden, wie das Netzwerk kostengünstig
ausgebaut und genutzt werden kann, um notwendige Gasmengen zu transportieren.
Diese Aufgabe kann mit Hilfe eines mathematischen Optimierungsproblems formuliert
werden, wobei anwendungsnahe Instanzen eine enorme Größe und Komplexität
aufweisen. Ziel der vorliegenden Arbeit ist die Entwicklung neuer mathematischer
Theorien und damit einhergehender Optimierungsalgorithmen, die es ermöglichen,
derartige Probleme zu lösen.
Die Optimierung der Topologie eines Gasnetzwerks kann mit Hilfe eines nichtlinea-
ren gemischt-ganzzahligen Programms (MINLP) modelliert werden. Durch Fixierung
aller ganzzahligen Variablen ergibt sich ein kontinuierliches Zulässigkeitsproblem, das
als aktives Transmissionsproblem (ATP) bezeichnet wird. Die zentrale Methode um
dieses ATP zu lösen, ist eine neuartige Relaxierung, welche Variablenschranken und
einige Nebenbedingungen relaxiert und in der Zielfunktion bestraft. Diese Relaxierung
bildet den Kern der in dieser Arbeit vorgestellten Theorie und ermöglicht so die
effiziente Lösung der Topologieoptimierung eines Gasnetzwerkes. Für den Fall, dass
die Relaxierung keine Primallösung für das ATP liefert, ist es uns gelungen, hinrei-
chende Bedingungen für die Unzulässigkeit des ATP zu formulieren, die durch ein
MILP dargestellt werden. Kurz gefasst kann die Unzulässigkeit eines nicht-konvexen
NLP durch Lösung eines MILP bewiesen werden. Beide Methoden liefern effiziente
Schranken in einem branch-and-bound Lösungsverfahren.
Besteht ein Gasnetzwerk nur aus Rohren und Schiebern, dann wird das ATP als
passives Transmissionsproblem (PTP) bezeichnet. Obwohl die Nebenbedingungen
des PTP nicht konvex sind, konnten wir zeigen, dass seine Relaxierung konvex ist.
Daher kann die Unzulässigkeit des PTP direkt auf effiziente Weise geprüft werden.
Außerdem können mit Hilfe der Relaxierung in diesem speziellen Fall Schnittebenen
für das Topologieoptimierungsproblem aufgestellt werden. Diese repräsentieren die

vii
Unzulässigkeit des PTP und folgen aus der Lagrange Funktion der Relaxierung
zusammen mit einer speziellen Erweiterung, der sogenannten pc-Regularisierung.
Abgesehen von den genannten Klassifizierungen kann die Relaxierung auch genutzt
werden, um primale Lösungen zu konstruieren. Hier nutzen wir die parametrische
Sensitivitätsanalyse, um mit Hilfe dualer Informationen Binärvariablen des ATP zu
identifizieren, deren Werte angepasst werden müssen. Dieser Ansatz erlaubt es zum
ersten Mal, direkt MINLP Lösungen für das Topologieoptimierungsproblem realer
Gasnetzwerke zu berechnen.
Die Resultate dieser Arbeit wurden im Rahmen des Forschungsprojekts “For-
schungskooperation Netzoptimierung (ForNe)” erarbeitet. Die entwickelte Software
wird vom Kooperationspartner Open Grid Europe GmbH aktiv genutzt.
Teile dieser Arbeit sind in Buchkapiteln, Journalen und technischen Berichten
publiziert. Eine Übersicht über die Themen und Lösungsansätze im ForNe-Projekt
veröffentlichten Martin u. a. (2011) und Fügenschuh u. a. (2013). Für Lösungsme-
thoden für die operative Nutzung von Gasnetzwerken verweisen wir auf Pfetsch u. a.
(2014). Ansätze für eine Topologieoptimierung wurden von Fügenschuh u. a. (2011)
beschrieben. Die in dieser Arbeit präsentierte primale Heuristik ist publiziert von
Humpola u. a. (2015a). Die genannte Methode, um Knoten innerhalb des branch-and-
bound Baums abzuschneiden, wurde für eine Approximation des Topologieproblems
von Humpola und Fügenschuh (2015) und Humpola u. a. (2015b) und Humpola und
Serrano (2017) beschrieben. Ein Bericht über die Ungleichungen nach Benders ist in
Humpola u. a. (2016) nachzulesen.

viii
Acknowledgements

First of all, I would like to thank my supervisor Prof. Dr. Dr. h.c. mult. Martin
Grötschel for enabling me to work in such a fascinating field of work, making it
possible for me to apply my theoretical advances to real life scenario.
In addition I wish to thank my second supervisor Prof. Dr. Thorsten Koch. He
encouraged me to attend diverse mathematical conferences, resulting in profitable
discussions, novel ideas and a meeting with Prof. Andrea Lodi, who later agreed to
be one of my supervisors. Moreover, from very early on, he put his trust in me and
encouraged me to follow my own ideas.
Furthermore I would also like to thank Prof. Andrea Lodi for inviting me to
Bologna, having long and fruitful discussions on further applications of the research
of this thesis and for including my defense into his busy schedule.
I have to thank Prof. Dr. Ralf Borndörfer for the constructive discussions we had
over the years and all valuable insights and ideas he shared with me. Thank you for
supporting and motivating me and for providing me with the last finishing touches
and suggestions for this thesis.
I am especially obliged to Prof. Dr. Armin Fügenschuh, Dr. Thomas Lehmann
and Dr. Nam Dung Hoang for careful reading this thesis and for their diverse criticism
and suggestions. I have to emphasize Prof. Dr. Armin Fügenschuh for his willingness
to support my publication ideas, enhancing them and to make our final results so
engaging to read.
I am very grateful to my colleague Felipe Serrano for his essential comments on
Chapter 6 which resulted in the rewriting and rethinking of the whole chapter.
Thanks must go to my colleagues of the working group “Energy” at ZIB, Dr.
Benjamin Hiller, Ralf Lenz, Jonas Schweiger, Robert Schwarz. I always appreciated
our vital discussions and the pleasant working atmosphere. I also like to thank them
for commenting on parts of this thesis.
I would also like to thank our industry partners, in particular Dr. Klaus Spreck-
elsen, Dr. Lars Huke and his team from Open Grid Europe GmbH. They made it
possible to test the algorithms developed in this thesis within real-life applications
by providing the data.

xi
 I wish to thank all my former and fellow colleagues of the ForNe project team
and the SCIP group, especially Prof. Dr. Marc E. Pfetsch for all the hours of fertile
discussions we had.
Thank you to Dr. Radoslava Mirkov for all the advice, friendship and inspring
conservations we had over the past years. I am happy I came to Berlin for my PhD.
Also I have to give many thanks to Zara, Daniel, Nadine and my mother for all
the proof reading and rephrasing of my thesis.
I have to express my biggest gratitude to my family, especially to my parents for
supporting me in my studies. Because of you I became the person I am today!
Last but not least, I want to thank Nadine for all her patience and her constant
trust in me.

xii
Contents

1 Introduction: Gas Network Optimization 1

1.1 Optimization Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Previous Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Solution Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Computational Study . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5 Outline of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2 Solving Mixed-Integer Nonlinear Optimization Problems 13

2.1 Definition of MINLP . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Details on SCIP for Solving MINLP . . . . . . . . . . . . . . . . . . 14
2.3 Convex and Non-Convex MINLP . . . . . . . . . . . . . . . . . . . . 17
2.4 Necessary Conditions for Local Optimality . . . . . . . . . . . . . . . 19
2.5 A Specially Tailored Adaptation of SCIP . . . . . . . . . . . . . . . . 20

3 An MINLP Model for Gas Network Topology Optimization 23

3.1 Technical Background . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 An MINLP Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3 Complexity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.4 The Passive and Active Transmission Problem . . . . . . . . . . . . 41
3.5 Computational Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4 Efficiently Solving the Passive Transmission Problem 51

4.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.2 Relaxation of Domains . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2.1 Existence of a Solution . . . . . . . . . . . . . . . . . . . . . . 54
4.2.2 Characterization of the Feasible Region . . . . . . . . . . . . 58
4.2.3 Interpretation of Lagrange Multipliers . . . . . . . . . . . . . 64
4.3 Relaxation of Flow Conservation Constraints . . . . . . . . . . . . . 71
4.3.1 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.3.2 Existence of a Solution . . . . . . . . . . . . . . . . . . . . . . 73
4.3.3 Characterization of the Feasible Region . . . . . . . . . . . . 78

xv
 4.3.4 Interpretation of Lagrange Multipliers . . . . . . . . . . . . . 85
4.4 Relaxation of Potential-Flow-Coupling Constraints . . . . . . . . . . 88
4.4.1 Conditions of the KKT System . . . . . . . . . . . . . . . . . 88
4.4.2 Different KKT Points . . . . . . . . . . . . . . . . . . . . . . 90
4.5 Solving the Passive Transmission Problem . . . . . . . . . . . . . . . 91
4.6 Integration and Computational Results . . . . . . . . . . . . . . . . . 95

5 An Improved Benders Cut for the Topology Optimization Problem 103

5.1 Valid Inequalities for the Passive Transmission Problem . . . . . . . 104
5.1.1 A Nonlinear Inequality . . . . . . . . . . . . . . . . . . . . . . 106
5.1.2 A Linear Inequality . . . . . . . . . . . . . . . . . . . . . . . 111
5.1.3 Feasibility Characterization by a Linear Inequality . . . . . . 114
5.1.4 A Linear Inequality derived from the Lagrange Function of the
Domain Relaxation . . . . . . . . . . . . . . . . . . . . . . . . 124
5.2 A Valid Inequality for the Topology Optimization Problem . . . . . 126
5.3 Integration and Computational Results . . . . . . . . . . . . . . . . . 133

6 Sufficient Conditions for Infeasibility of the Active

Transmission Problem 141
6.1 Non-Convex Relaxations for the Active Transmission Problem . . . . 142
6.1.1 Relaxation of Domains . . . . . . . . . . . . . . . . . . . . . . 143
6.1.2 Relaxation of Flow Conservation Constraints . . . . . . . . . 150
6.2 Detecting Infeasibility of the Active Transmission Problem by MILP 155
6.3 Interpretation of the Infeasibility Detection MILP . . . . . . . . . . 170
6.4 Integration and Computational Results . . . . . . . . . . . . . . . . . 175

7 A Primal Heuristic based on Dual Information 181

7.1 A Relaxation of the MINLP . . . . . . . . . . . . . . . . . . . . . . . 182
7.2 A Primal Heuristic for MINLP with Indicator Constraints . . . . . . 184
7.2.1 Theoretical Motivation . . . . . . . . . . . . . . . . . . . . . . 184
7.2.2 The Basic Dual Value MINLP Heuristic . . . . . . . . . . . . 187
7.2.3 Embedding the Heuristic in a Branch-and-Prune Search . . . 188
7.3 A Specialization to the Topology Optimization Problem . . . . . . . 188
7.3.1 The Relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . 189
7.3.2 Handling Different Modes of Active Devices . . . . . . . . . . 190
7.3.3 Handling Loop Extensions . . . . . . . . . . . . . . . . . . . . 194
7.4 Implementation Details . . . . . . . . . . . . . . . . . . . . . . . . . 196
7.5 Computational Results . . . . . . . . . . . . . . . . . . . . . . . . . . 198

xvi
8 Conclusions 203

A Tables 207

Bibliography 227

xvii
Chapter 1

Introduction: Gas Network

Optimization

Natural gas is a nontoxic, odorless, transparent, and flammable gas that originates
from underground deposits. Today natural gas is mainly used for heating private
houses and office buildings, for the generation of electrical power, as fuel for vehicles,
and for several reactions in chemical process engineering. Natural gas usage represents
one quarter of the world’s energy demand (BGR 2013). It must be transported from
the deposits to the customers, sometimes over distances of thousands of kilometers.
For very long distances it is more economic to cool the gas down to −160° so that
it becomes liquid and can be transported by ships (see Cerbe 2008). For shorter
distances or for the delivery to the end customers large pipeline systems are used.
Existing gas networks have usually grown over time. In Germany, the high-pressure
pipeline system was built by gas supply companies. It has a size of approximately
35 000 km (FNBGas 2013).
Historically these companies were both gas traders and gas network operators.
They purchased gas from other suppliers and operated the necessary infrastructure
to transport the gas from those suppliers to their own customers. During the
liberalization of the German gas market these business functions were separated by
regulatory authorities (GasNZV 2005). Nowadays, there are companies that trade gas
and others whose sole task is the operation of gas networks for the transportation of
gas. One of the requirements set by the regulatory authorities is that every trader can
use the network infrastructure to transport gas. Open access to these gas networks
has to be granted to all the trading companies free of any discrimination. This means
that the gas supplies and demands cannot be fully controlled by the network operator.
Therefore the network operator is required to have a high degree of operational
flexibility. The majority of network management is carried out manually with the
aid of simulation software. There is a need to develop a more automated process

1
2 Chapter 1 Introduction: Gas Network Optimization

in order to cope with anticipated challenges associated with the addition of more
traders accessing the network. Here we have developed mathematical optimization
methods to improve the network operation and to enhance the cost effectiveness of
investments in the infrastructure.

1.1 Optimization Tasks

The physics of a gas transport network, which is used for the transportation of gas,
can roughly be described as follows: Most of the network elements are pipelines. A
gas flow through a pipeline (pipe) is induced if the gas has different pressures at
the end nodes of the pipe. Usually gas pipelines can withstand nominal pressures of
16 bar up to 100 bar. Typically at long transport distances of about 100 km to 150 km
the gas pressure gets too low which is technically not feasible. In situations where
lack of pressure is an issue compressors are used to increase the pressure again. High
gas pressures can also be problematical in parts of the network. Therefore in order to
protect the network it can be necessary to reduce the pressure by using control valves.
This is particularly important when the network includes older pipelines which have
a lower pressure limit. If parts of the network need to be deactivated, then valves
allow to split the network into physically independent subnetworks.
A gas network is mathematically modeled by a directed graph, refer to Korte and
Vygen (2007) for the notations in graph theory. This graph consists of nodes together
with connecting arcs. The end points of an arc are nodes. Each node corresponds
to a geographical position. Each arc models a network element. We use a directed
graph in order to distinguish between gas flow in the direction of an arc and in the
opposite direction. A network element can either be a pipeline, a valve, a compressor
or a control valve. All network elements determine a specific relation between the
flow through the element and the pressure at the end nodes. Pipelines are called
passive network elements, while valves, compressors and control valves are called
active network elements. Pipelines have a unique relation between the flow through
a pipeline and the pressures at its end nodes. This is different for active elements.
Their physical behavior can be influenced by the network operator, for instance, a
valve can be open or closed. An open valve means that the pressures at the end
nodes are equal, while closed means that there is no gas flow through the valve.
Compressors and control valves can be open or closed. This means that each active
network element has different operation modes or rather configurations available. An
accurate description of every network element is given in Section 3.1.
1.1 Optimization Tasks 3

∃ q, π

a4 s. t. q1 = 1,
−2.1 +1
q2 − q1 = 0,
v4 v5
q2 + q3 = 2.1,
a3 q3 + q4 = 2.1,
q4 = 1,
+2.1 v3
44 q1 |q1 | − (π2 − π1 ) = 0,
44 q2 |q2 | − (π3 − π2 ) = 0,
a2
44 q3 |q3 | − (π3 − π4 ) = 0,
−1
44 q4 |q4 | − (π5 − π4 ) = 0,
v1 a1 v2 π1 = 0,
(a) Test network. 0 ≤ π2 , π3 , π4 , π5 ≤ 92,
π1 , π2 , π3 , π4 , π5 , q1 , q2 , q3 , q4 ∈ R.

(b) Nomination validation problem for the net-

work shown in Figure 1.1a (left picture).

Figure 1.1: An example of a test network and a model of the corresponding nomination validation problem
as used in this thesis. The instance has two entries at node v3 with flow amount +2.1 and v5 with flow
amount +1. Two exits are at node v1 with flow amount −1 and v4 with flow amount −2.1. Node v2 is
a transmission node. These node flows imply the arc flows 1 for a1 and a2 and a4 and 1.1 for a3 . Every
arc is a pipeline which means a unique relation between the arc flow and the pressures at the end nodes.
The pressure at node v1 is fixed to zero which implies that any other node pressure is fixed. Hence the
pressure at node v5 is also fixed. The pressure bounds of the pipes imply that the pressure at node v5
violates its upper bound. So the nomination validation problem is infeasible for this instance.

We distinguish between two different optimization tasks. Given is a nomination

that specifies for each single node the amount of gas that enters or leaves the network
there. The network in combination with the nomination defines an instance. In
order to operate a gas transport network the task is to compute a configuration
of valves, control valves and compressors, a pressure for each node, and a flow for
each element. This computation requires that the flow specified by the nomination
is transported through the network and all technical and physical as well as legal
constraints are fulfilled. The arising problem is a feasibility problem which is called
nomination validation problem of the specified instance. Figure 1.1 shows an example.
A simple tree network together with a nomination is shown in 1.1a and the associated
nomination validation problem is infeasible. This means that it is not possible to
transport gas through the network according to the specified nomination. A model
of this problem as used in this thesis is shown in 1.1b and will be explained in detail
in Chapter 3. Due to the infeasibility of the nomination the network needs to be
extended. In principal all network elements, i.e., pipelines, valves, control valves and
compressors, can be added to the gas network. The additional elements are called
extensions. The task is to compute cost-optimal extensions in order to transport the
gas defined by the nomination through the network. The arising problem is called
topology expansion problem. Its objective is to minimize the building costs of the
4 Chapter 1 Introduction: Gas Network Optimization

(a) Extended by a control valve. (b) Extended by a pipe in combina-

tion with a compressor.

Figure 1.2: Extended versions of the network shown in Figure 1.1a. The discussion in Figure 1.1 explains
that the nomination validation problem 1.1b is infeasible because the pressure at node v5 violates its
upper pressure bound. Decreasing the pressure in node v4 is a possible adaptation in order to reduce the
pressure in node v5 . This adjustment is achieved by the proposed extensions in 1.2a and 1.2b.

additional network elements. Note that operation costs of the network elements are
assumed to be a constant term in this objective which is not taken into account
for the optimization. They are caused by compressors which consume energy for
the compression of the gas or pipelines which have to be maintained regularly, for
example. Figure 1.2 shows two different suitable extensions for the network in
Figure 1.1a. We refer to Figure 1.3 for a visualization of the effect of extending a gas
network. More precisely the node pressures are shown before and after extending
the network. The networks shown so far are only small examples. A real-world gas
network with approximately 4000 arcs for which we want to solve the nomination
validation problem is shown Figure 1.4.
The building costs of a new network element depend on its type. For building a
pipeline the costs are mainly made up of the material costs, the construction costs
and costs of getting permission rights for the use of land. Construction costs for
compressors are different as a compressor is typically built at a single place. We
assume that the construction costs also contain a term representing the operation
costs of the specific network element.

1.2 Previous Work

A general survey over the application of optimization methods in the natural gas
industry is given by Zheng et al. (2010). The surveys by Shaw (1994) and Ríos-
Mercado and Borraz-Sánchez (2012) are closer to the problems we are studying. A
monograph outlining the earlier state-of-the-art is described in the book of Osiadacz
(1987).
1.2 Previous Work 5

max min
Figure 1.3: Element flow and node pressure corresponding to a realization of a nomination in a test
network consisting of pipelines only. The line width represents the flow value (the thicker the more flow),
while its color depicts the mean value of the node pressures at both end nodes. The nodes are depicted
by squares and the node colors represent the node pressures. In our test case we added two pipelines to
the network which results in a different flow and pressure distribution (see the lower picture). In both
pictures the maximum of the node pressures corresponds to the color red and the minimum to the color
blue.
6 Chapter 1 Introduction: Gas Network Optimization

Figure 1.4: The real-world network net7 provided by the cooperation partner Open Grid Europe GmbH.
The task is to solve the nomination validation problem for different instances of this network. Here
state-of-the-art solvers show a poor performance as discussed in Section 1.4.
1.2 Previous Work 7

Several models and solution approaches exist for the nomination validation
problem. One class of approaches is based on solving a mixed-integer linear program
(MILP). The main difficulty then is to get an adequate model for the nonlinearities.
For the stationary case, Möller (2004) used piecewise-linear approximations; see also
Martin et al. (2006). A similar approach using different linearizations was used by
Tomasgard et al. (2007) and Nørstebø et al. (2010).
Among the many approaches based on solving a nonlinear program (NLP), we
mention Percell and Ryan (1987), who use a gradient-descent based method and De
Wolf and Smeers (2000), who consider sequential linear programming approaches.
Based on a fine simulation model, Jeníček (1993) and Vostrý (1993) use subgradient-
based methods. Also sequential quadratic programming (Furey 1993; Ehrhardt and
Steinbach 2005; Ehrhardt and Steinbach 2004) and interior point methods (Steinbach
2007) have been considered. Bonnans et al. (2011) use interval analysis techniques to
analyze an approximated model of the Belgian network with 20 pipes. Approaches
that rely on network reduction techniques are discussed in Hamam and Brameller
(1971), Mallinson et al. (1993), Wu et al. (2000), and Ríos-Mercado et al. (2002).
Another widely used approach is dynamic programming. A survey for the work up
until 1998 is written by Carter (1998). The first application were so-called gun-barrel
networks, i.e., straight line networks with compressors (Wong and Larson 1968b;
Wong and Larson 1968a). The method was later extended to more complex network
topologies; see e.g., Lall and Percell (1990), Gilmour et al. (1989), and Zimmer
(1975). Further extensions of this approach were proposed by Borráz-Sánchez and
Ríos-Mercado (2004).
Not surprisingly, many purely heuristic approaches have been developed. We
mention applications of simulated annealing by Wright et al. (1998) (specifically for
the optimization of compressor operations) and Mahlke et al. (2007), tabu search by
Borraz-Sánchez and Ríos-Mercado (2004), general expert systems by Sun et al. (2000),
genetic algorithms by Li et al. (2011), and ant colony optimization by Chebouba
et al. (2009). Kim et al. (2000) and Ríos-Mercado et al. (2006) present a two stage
iterative heuristic to minimize fuel cost of the compressors.
For the topology expansion problem there exist several approaches to improve
the topology of a gas network which mainly consist of various heuristic and local
optimization methods. Boyd et al. (1994) apply a genetic algorithm to solve a
pipe-sizing problem for a network with 25 nodes and 25 pipes, each of which could
have one of six possible diameters. André et al. (2006) consider a similar problem
and present a heuristic method based on relaxations and local optima. Castillo
and Gonzáleza (1998) also apply a genetic algorithm for finding a tree topology
8 Chapter 1 Introduction: Gas Network Optimization

solution for a network problem with up to 21 nodes and 20 arcs. In addition to

pipes compressors can also be placed in the network. Mariani et al. (1997) describe
the design problem of a natural gas pipeline. They present a set of parameters to
evaluate the quality of the transportation system. Based on these ones they evaluate
a number of potential configurations to identify the best among them. Osiadacz
and Górecki (1995) formulate a network design problem for a given topology as a
nonlinear optimization problem, for which they iteratively compute a local optimum.
For a given topology the diameter of the pipes is a free design variable. Their method
is applied to a network with up to 108 pipes and 83 nodes. De Wolf and Smeers
(1996) also use a nonlinear formulation, which is then solved heuristically. For a
given topology with up to 30 arcs and nodes they can determine optimized diameters
of pipes. Further approaches are available from Hansen et al. (1991), Bonnans et al.
(2011), Babonneau et al. (2012), and Zhang and Zhu (1996). A method for a complete
redesign of a gas transport network is presented by Hübner (2009).

1.3 Solution Approach

Within the joint research project “Forschungskooperation Netzoptimierung (ForNe)”
in cooperation with our industry partner Open Grid Europe GmbH (OGE) we have
developed a method for the nomination validation as well as the topology expansion
problem. OGE is a company which actually operates and maintains the largest gas
transport network in Germany. Such a company operating a gas network is also
called transmission system operator (TSO).
The previous approach of OGE for solving the nomination validation problem
was as follows: Given a nomination experts used simulation software like Simone
(LIWACOM 2005) and PSIGanesi (Scheibe and Weimann 1999) to compute flows
for each element and pressures for each node such that the flow specified by the
nomination is transported through the network. This computation needs to be as
such that physical, technical and legal constraints are fulfilled. The software simulates
the gas physics and characteristics of the active and passive elements of the network.
Therefor a set of input parameters must be available which require expert knowledge.
For solving the topology expansion problem the gas network was manually extended
first, while the nomination validation problem on the extended network was then
solved in a second step. These two steps were iterated while the number and size of
extensions in the first step were consecutively reduced.
An aim of the research project ForNe was to develop mathematical optimization
techniques which aid these processes. The physics of gas networks are described by
1.3 Solution Approach 9

nonlinear equations. Additionally discrete decisions are necessary because active

elements have different configurations as they can, for instance, be open or closed.
A possible approach is to state the nomination validation as well as the topology
expansion problem as a mixed-integer nonlinear program (MINLP). A formal
definition of an MINLP is given in Section 2.1.
The general approach in the research project for solving the nomination validation
problem was to split the solution process into two steps. In a first step different
approaches and MINLP models were developed for computing reasonable discrete
decisions, i.e., configurations of compressors, valves and control valves, for a given
nomination. Here physical constraints are simplified. Once these decisions are
computed, an NLP is solved in a second step. This NLP models the physical and
technical constraints of all network elements in detail. A solution for this NLP
yields a result which is in precision comparable to the manual approach using the
simulation software. For more details and a precise description of this two stage
solution approach we refer to Koch et al. (2015). All these models focus on the
stationary case meaning transient gas flows are not considered. In addition it should
be noted that the gas flow through a pipeline is modeled independently of time. This
is due to uncertain gas consumption in the long run. Here especially the nominations
for the topology optimization problem can only be specified roughly and independent
of time.
Let us briefly summarize the first stage MINLP models and solution approaches
described in Koch et al. (2015). Geißler (2011) presents a solution method which
bases on MIP relaxations. All nonlinearities are modeled by discretization techniques
while the user can give a predefined maximal approximation error. Based on this
model, Morsi (2013) presents a network decomposition approach. Furthermore there
are presented two heuristic strategies. The first one is based on network reductions
(see Stangl 2014) and the second one on an MPEC solution approach. For the second
stage NLP model we refer to Schmidt et al. (2014).
The strategy that we follow in this thesis is to provide an MINLP model for the
first stage. We do not differ between the nomination validation and the topology
expansion problem. Instead we consider both problems as a topology optimization
problem and model this by an MINLP. Therefor we assume that the network
operator OGE provides a predefined set of additional network elements together
with installation positions. Then the topology optimization problem is to compute
a cost-optimal selection in order to transport the flow defined by the nomination
through the network. This means that the nomination validation problem is equal to
the topology optimization problem without objective function for a fixed selection.
10 Chapter 1 Introduction: Gas Network Optimization

Baron Antigone SCIP SCIP + heuristic

feas infeas no sol feas infeas no sol feas infeas no sol feas infeas no sol
- 3 27 - 28 2 1 - 29 18 - 12

Table 1.1: Computational study of the nomination validation problem for 30 nominations on the network
net7 shown in Figure 1.4. For all instances we set a time limit of 14 400 s. The computational results of
the last column “SCIP + heuristic” are obtained by a specially tailored heuristic presented in Chapter 7.
The last column is used only to demonstrate that at least 18 instances are feasible.

Note that a cost-optimal solution for the topology optimization problem is a global
optimal solution which means that any other feasible solution does not improve the
objective function value.

1.4 Computational Study

We focused on the nomination validation problem for the network shown in Figure 1.4.
It was provided by our cooperation partner OGE. We applied state-of-the-art MINLP
solvers to the mixed-integer nonlinear program that we introduce in Chapter 3 as a
model of the nomination validation problem. Our aim was to get an impression of
the solving performance of the solvers. For a computational study we consider 30
different feasible nominations. The arising MINLPs consist of 1478 binary and 21 957
continuous variables. There are 3757 nonlinear and 31 931 linear constraints. All
nonlinear constraints are of the same type. They consist of two different continuous
variables z1 and z2 and write as z1 |z1 | = z2 . Among the linear constraints there are
3406 “bigM”-constraints meaning that they are either active or inactive dependent
on the value of a binary variable.
For solving these instances we applied the state-of-the-art MINLP solvers SCIP
(Achterberg 2009; Vigerske 2012), Baron (Tawarmalani and Sahinidis 2005) and
Antigone (Misener and Floudas 2014). SCIP was used in combination with
Cplex 12.1 (CPLEX) as linear programming solver and IPOpt 3.10 (Wächter and
Biegler 2006) as nonlinear solver. The version of Baron that we used is 12.7.3, the
version of Antigone is 1.1.
A summary of our computational results is shown in Table 1.1. Detailed results
are shown in the right column of Table A.1 – A.3. It turns out that neither Baron
nor Antigone were able to compute any feasible solution within a time limit of
4 hours. Antigone detects 28 instances to be infeasible while at least 18 out of 30
instances are feasible. SCIP cannot compute any feasible solution for 29 instances.
The results of the last column of Table 1.1 are obtained by a specially tailored
heuristic which is described in Chapter 7.
1.5 Outline of the Thesis 11

Overall the performance of state-of-the-art MINLP solvers is not appropriate for

solving the topology optimization problem (3.2.1). Therefore the aim of this thesis
is to present different methods for improving the performance of SCIP for solving
(3.2.1) on real-world gas networks.

1.5 Outline of the Thesis

The outline of the following chapters is as follows: In Chapter 2 we give a brief sum-
mary of solving MINLP by branch-and-bound, separation, and spatial branching in
general. We refer to the MINLP solver SCIP which enables this approach. In Chap-
ter 3 we provide a detailed description of the topology optimization problem (3.2.1).
We also identify the passive transmission problem (3.4.1) and the active transmission
problem (3.4.2) as subproblems of the topology optimization problem (3.2.1). The
networks and corresponding nominations which we consider for our computational
studies are presented. We consider different networks which are either obtained from
literature, or were generated manually representing realistic networks, or are based
on real-world data. These networks differ according to size and types of network
elements that are included. We split the networks into different groups where we
distinguish between the elements that are contained. For extensions that are pipes we
differentiate whether the pipe would follow an existing pipe or not. In the first case
the extension is called a loop. Now we distinguish between three types of network:

1. networks that only consist of pipelines and valves,

2. networks that only consist of pipelines, loops and valves,

3. networks that contain active elements.

Note that this is a hierarchical classification. The latter group always contains the
former groups. The subsequent Chapters 4 – 6 focus on a specialized solution approach
for each of these network types. When using a gas network for our computational
studies we apply the solution methods presented for the smallest set in which it is
contained following the aforementioned classification.
In Chapter 4 we focus solely on the first type of network, namely those that only
contain pipes and valves. We describe and compare novel solution methods for solving
the passive transmission problem as part of the topology optimization problem (3.2.1)
to global optimality. The solution methods consist of solving convex relaxations of
the passive transmission problem. These methods are further integrated into the
solver SCIP that we use for solving (3.2.1). This integration allows the number of
12 Chapter 1 Introduction: Gas Network Optimization

globally solved instances within our given time frame to increase by 29 %. On average
the run time is reduced by 72 % for those instances already solvable by SCIP.
In Chapter 5 we focus on the second and more complicated type of network
namely those that additionally contain loops. We derive an improved Benders cut
(Geoffrion 1972) for the topology optimization problem from the dual solution of the
relaxations of the passive transmission problem. This cut represents the infeasibility
of the current passive transmission problem and allows us to speed up the overall
solution process when added to the problem formulation (3.2.1). Using this strategy
we are able to increase the number of globally solved instances by approximately
13 %. For those instances which are already solvable by SCIP the run time is reduced
by approximately 33 %.
In Chapter 6 we focus on the third type of network which consists of all types of
network elements. We introduce a special tailored algorithm for solving the active
transmission problem (3.4.2) which is part of the topology optimization problem
(3.2.1). This algorithm consists of a primal heuristic and sufficient conditions of
infeasibility for (3.4.2). We integrate this algorithm into the solver SCIP for solving
problem (3.2.1). Thereby the number of globally solved topology expansion instances
is increased by 20 %.
With the above described methods we are still not able to solve the real-world
instances used for the computational study in the previous Section 1.4. These
instances are feasibility instances. Therefor in Chapter 7 we introduce a primal
heuristic for the topology optimization problem which is integrated into the solver
SCIP. Here we make use of the dual solution of a relaxation of the active transmission
problem (3.4.2). It allows to solve approximately 60 % (18 out of 30) of the initially
discussed nomination validation instances on net7 in Section 1.1. Recall that only
approximately 3 % of these instances (1 out of 30) were solvable using state-of-the-art
MINLP solvers.
Chapter 8 contains concluding remarks and ideas for future research. Moreover
we present additional benefits resulting from our methods which are especially useful
for a TSO.
Chapter 2

Solving Mixed-Integer Nonlinear

Optimization Problems

In this chapter we roughly describe a solution method for an MINLP such as the
topology optimization problem (3.2.1). We start with a mathematical definition of
MINLP in Section 2.1. Then, in Section 2.2, we briefly explain the solution methods
of SCIP for solving an MINLP, see Achterberg (2009) and Vigerske (2012). For a
survey of different solution methods we refer to Belotti et al. (2012) and for different
MINLP solvers to Bussieck and Vigerske (2010), D’Ambrosio and Lodi (2011), and
D’Ambrosio and Lodi (2013). In Section 2.3 we explain the difference between convex
and non-convex MINLP and the impact on the aforementioned solution methods.
Afterwards we present the necessary conditions for optimality of a primal solution in
Section 2.4.
From the computational study in Section 1.4 we concluded that we have to
improve the solution approach for solving the topology optimization (3.2.1) problem
in several directions. As discussed in the previous chapter we will present different
results that allow us to prune nodes of the branch-and-bound tree and further a
primal heuristic. For an implementation of these methods SCIP provides a flexible
framework. In Section 2.5 we explain our adaptations of SCIP for solving the
topology optimization problem.

2.1 Definition of MINLP

Nonlinear optimization problems containing discrete and continuous variables are
called mixed-integer nonlinear programs (MINLPs). A general MINLP can be
formulated as

min{f (x) | x ∈ X} (2.1.1a)

13
14 Chapter 2 Solving Mixed-Integer Nonlinear Optimization Problems

with
X := {x ∈ [x, x] | g(x) ≤ 0, h(x) = 0, xi ∈ Z, i ∈ I}. (2.1.1b)

Here x, x ∈ Rn determine the lower and upper bounds on the variables, I ⊆ {1, . . . , n}
denotes the set of variables with integrality requirement, f : [x, x] → R is the objective
function, and g : [x, x] → Rm and h : [x, x] → R` are the constraint functions. The
set X is called feasible set of (2.1.1). We assume f (x), g(x) and h(x) to be at least
continuous. Further we assume that f (x) is linear. This is obtained by shifting
a nonlinear objective function to the constraints while setting f (x) = z for a new
continuous variable z. This goes along with replacing the objective “min f (x)” by
“min z”.
A point x ∈ X is called local optimum, if there exists an  > 0 such that for all
y ∈ X with kx − yk <  we have f (x) ≤ f (y). A local optimum whose objective
function value equals the optimal value is called a global optimum. Note that, due to
continuity of f (x) and g(x), there always exists a global optimum of (2.1.1), if its
optimal value is finite.
MINLP problems arise in many fields such as energy production and distribution,
logistics, engineering design, manufacturing, and chemical and biological sciences,
see Floudas (1995), Grossmann and Kravanja (1997), Tawarmalani and Sahinidis
(2002), Pintér (2006), and Ahadi-Oskui et al. (2010).

2.2 Details on SCIP for Solving MINLP

Below we describe the solution technique for solving MINLP (2.1.1) as implemented
in SCIP, see Achterberg (2004) and Vigerske (2012). By “solving” we mean to
compute a feasible solution for an approximation of a given instance of the problem
together with a computational proof of its global optimality. The solution is globally
optimal for an approximation of the problem and not for the original version due to
numerical reasons.
The solution process in SCIP is as follows: The MINLP (2.1.1) is first relaxed
to a mixed-integer linear program (MILP) and further to a linear program (LP).
For a detailed introduction into linear and integer programming and combinatorial
optimization see for example Nemhauser and Wolsey (1989). Recall that X ⊆ Rn
is the feasible set of (2.1.1). A linear outer approximation of the feasible set X is
computed such that
X ⊆ x ∈ [x, x]  Dx ≤ d
 
2.2 Details on SCIP for Solving MINLP 15

zb = za |za | zb = za |za | zb = za |za |

za za za

(a) Outer approximation. (b) Separation. (c) Spatial branching together

with separation.

Figure 2.1: Example of outer approximation, separation, and spatial branching when handling the non-
linear relation zb = za |za | between two continuous variables za and zb .

for a suitable matrix D ∈ Rm×n and vector d ∈ Rm . As an example for a linear

outer approximation consider the nonlinear constraint

za |za | = zb , (2.2.1)

for two continuous variables za , zb , see Fügenschuh et al. (2010). An outer approxi-
mation of (2.2.1) with four different linear inequalities is shown in Figure 2.1a, for
instance. The LP relaxation of MINLP (2.1.1) then writes as

min{f (x) | Dx ≤ d, x ∈ [x, x]}. (2.2.2)

Recall that f is a linear function. LP (2.2.2) is solved in practice by the dual simplex
algorithm (see Dantzig 1951). The obtained solution value defines a lower bound
on the optimal value of the original MINLP (2.1.1). In case this solution fulfills all
constraints of (2.1.1), it is a proven global optimal MINLP solution. However, this
rarely happens in practice. Hence either cutting planes are added to strengthen the
relaxation or a branching on a variable is performed as described below.
By adding cutting planes the linear relaxation is improved. For a suitable matrix
D0 ∈ Rm ×n having more rows than D, i.e., m0 > m, and d0 ∈ Rm , the linear
0 0

relaxation of the feasible set X is improved by

X ⊆ {x ∈ [x, x] | D0 x ≤ d0 } ⊂ {x ∈ [x, x] | Dx ≤ d}.

The feasible space which remains after adding all available cutting planes is visualized
in Figure 2.1b for the nonlinear function (2.2.1). Adding cutting planes always yields
a tighter LP relaxation of the original problem while the relaxation is still a linear
program. For more details on separation we refer to Nemhauser and Wolsey (1989).
A branching is either performed on an integer or on a continuous variable. In
both cases the domain of a variable is restricted and a new subproblem is created.
16 Chapter 2 Solving Mixed-Integer Nonlinear Optimization Problems

Branching on an integral variable xi means to subdivide the previous linear relaxation

into two parts

X ⊆ x ∈ [x, x]  Dx ≤ d, xi ≤ y ∪ x ∈ [x, x]  Dx ≤ d, xi ≥ y + 1 .
   

for an integral value y. When branching on a continuous variable we speak of spatial

branching. Spatial branching on the continuous variable xi of the solution x∗ to the
linear relaxation refers to subdividing the previous linear relaxation into two parts

X ⊆ x ∈ [x, x]  Dx ≤ d, xi ≤ x∗i ∪ x ∈ [x, x]  Dx ≤ d, xi ≥ x∗i .

   

For each part of the relaxation a subproblem is created and a tighter outer
approximation can be computed due to tighter variable bounds, see Figure 2.1c for
example. Spatial branching thus improves the relaxation, in particular, in places
where the functions cannot be properly approximated by cutting planes. A branching
tree is used for managing the different subproblems. The initial LP relaxation (2.2.2)
is associated with the root of the branching tree. Whenever a problem corresponding
to a node of this tree is split in subproblems, each of these is associated with a child
node.
Branching is pursued until all integral variables take integral values and the outer
approximation is “close enough” to the feasible region. If a subproblem is infeasible or
its optimal value is larger than the best available solution so far, then the subproblem
needs not to be investigated furthermore.
The methods above are summarized as branch-and-bound, separation and spatial
branching. This way, global bounds on the objective function can be computed and
the problem can be solved to global optimality up to a certain accuracy. For more
details on branch-and-bound for MILP refer to Nemhauser and Wolsey (1989).
SCIP provides different constraint handlers. They allow to deal with different
types of constraints of an MINLP. For instance, the nonlinear constraint (2.2.1) is
handled by the so called “cons_abspower” constraint handler as described above. A
constraint of the form x = 1 ⇒ g(z) ≤ 0, where g(z) is a linear function and x a
binary variable, is called indicator constraint and handled by the “cons_indicator”
constraint handler. Roughly spoken, the constraint is modeled by the constraint
handler as
g(z) ≤ M (1 − x),

where M ∈ R≥0 is a suitable constant which is called “bigM”. Additionally there

are some auxiliaries implemented to handle numerical intractability. We refer to
these indicator constraints because they are used for modeling the different modes
2.3 Convex and Non-Convex MINLP 17

of active elements in our model (3.2.1) for the topology optimization problem. The
“cons_nonlinear” constraint handler is the most general MINLP constraint handler
of SCIP for nonlinear constraints, see Vigerske (2012).
Apart from constraint handlers SCIP has several other functionalities for solving
MINLPs. There are for example primal heuristics, for computing primal feasible
solutions. Five heuristics are specifically tailored towards mixed-integer nonlin-
ear programming problems: Undercover, RENS, nonlinear versions of RINS and
Crossover, see Danna et al. (2004) and Berthold (2014). Further an NLP local search
heuristic (Vigerske 2012), and a nonlinear diving heuristic (Bonami and Gonçalves
2012). Other heuristics are MILP heuristics that are applied to the linear outer
approximation plus the integrality constraints. Propagators are used to strengthen
variable bounds, different branching strategies allow one to decide which variable
branching is performed, and node selection strategies are available, which select nodes
of the branching tree to be considered next. Again we refer to Vigerske (2012) and
Achterberg (2009) for details.

2.3 Convex and Non-Convex MINLP

In general, when solving an MINLP of the form (2.1.1) we differ between convex and
non-convex MINLP. The optimization problem (2.1.1) is convex if the constraint
functions as well as the objective function are convex. A constraint function gj :
[x, x] → R is convex if

gj (λx1 ) + gj ((1 − λ)x2 ) ≤ λgj (x1 ) + (1 − λ)gj (x2 ) ∀ λ ∈ [0, 1] (2.3.1)

holds for any two points x1 , x2 ∈ [x, x]. Consequently (2.1.1) is non-convex if any
constraint function gj or hj does not fulfill the condition (2.3.1). A convex MINLP
has the important property that for fixed integral variables xi , i ∈ I, every local
minimum of (2.1.1) is global (see Boyd and Vandenberghe 2004, Section 4.2.2). For
the special case of convex MINLP there exist several solvers, see Bonami et al. (2008),
Abhishek et al. (2010), IBM (CPLEX), and FICO (2009).
An example for a non-convex MINLP is given in Figure 2.2. The optimization
problem as shown in 2.2b is a continuous non-convex optimization problem because
the feasible set as shown in Figure 2.2a is not convex. We conclude from Figure 2.2a
that the non-convex feasible set can be divided into two parts, one convex part with
x ≤ 0 and another part with x ≥ 0. This additional information allows splitting
problem 2.2b into two problems, while the convex one yields a global optimal solution.
Note that the information about a smart split of the feasible set is typically not
18 Chapter 2 Solving Mixed-Integer Nonlinear Optimization Problems

min y
x s. t.
2
(x − 5)x|x| − (x − 4.5) − y ≤ 0,
x, y ∈ R

(b) Non-convex optimization problem.

(a) Non-convex feasible set.

Figure 2.2: Example of a non-convex continuous optimization problem. The feasible set as shown in the
left figure is non-convex. The objective function of the optimization problem shown in the right part is
linear.

available. Hence the solution methods for convex MINLP cannot be applied directly
for solving non-convex MINLP. For more details on solution methods for MINLP
we refer to Vigerske (2012). We briefly sketch two methods.
A method for solving a non-convex MINLP is to apply the branch-and-bound,
separation, and spatial branching technique described in Section 2.2 as implemented
in SCIP. Here branching on integer as well as spatial branching on continuous
variables needs to be performed, see Figure 2.1 for instance. This is different for
a convex MINLP. Because of the convexity of the constraints, there is no need
to apply spatial branching. Any tangential hyperplane of a convex function yields
a globally valid inequality for MINLP (2.1.1). Hence all nodes of the branching
tree only arise from branching on integral variables. Therefore non-convex MINLP
are difficult to solve in comparison to convex MINLP. This in turn means, the
more information is known about the solution space of an MINLP, the more the
solving performance can be improved. Concerning the example shown in Figure 2.2
this means that by splitting the problem 2.2b into two MINLP, a convex and a
non-convex one, possibly time consuming spatial branching can be avoided.
A different solution approach for convex MINLP is to apply an NLP based
branch-and-bound method. Here a convex NLP is solved instead of an LP at each
node of the branching tree. This way even separation is not necessary. This solution
method can be extended to non-convex MINLP by using convex underestimators for
non-convex functions. A convex underestimator for g(x) is a convex function g̃(x)
with g̃(x) ≤ g(x) on the domain of g. Note that g̃ is not required to be linear. Using
2.4 Necessary Conditions for Local Optimality 19

convex underestimators for non-convex MINLP, i.e., for g and h, goes along with
spatial branching, i.e., branching on continuous variables, see Vigerske (2012).

2.4 Necessary Conditions for Local Optimality

After fixing all integral variables of MINLP (2.1.1), a continuous nonlinear optimiza-
tion problem (NLP) remains. We write this abbreviated as

min{f (z) : gi (z) ≤ 0, hj (z) = 0, z ∈ Rn } (2.4.1)

where f , gi (i = 1, . . . , m), and hj (j = 1, . . . , `) are continuous functions and z

represents the remaining unfixed variables. If these functions are continuously
differentiable, then the Karush-Kuhn-Tucker conditions of this optimization problem
yield a necessary condition for a feasible solution of (2.4.1) to be locally optimal, see
Conn et al. (2000). These conditions are defined as follows:

Definition 2.4.1:
Let the functions f, g, h of the optimization problem (2.4.1) be continuously differen-
tiable. The conditions

∇z L(z, µ, λ) = 0, (2.4.2a)
gi (z) ≤ 0, ∀ i = 1, . . . , m, (2.4.2b)
hj (z) = 0, ∀ j = 1, . . . , `, (2.4.2c)
λi ≥ 0, ∀ i = 1, . . . , m, (2.4.2d)
λi gi (z) = 0, ∀ i = 1, . . . , m, (2.4.2e)

with the Lagrange function L : Rn × R` × Rm → R defined as

m `
L(z, µ, λ) = f (z) + λi gi (z) + µj hj (z) = 0
X X

i=1 j=1

are called Karush-Kuhn-Tucker conditions (KKT conditions as abbreviation)

of the continuous nonlinear optimization problem (2.4.1). Every vector (z ∗ , µ∗ , λ∗ )
which fulfills the KKT conditions is called KKT point. The components of µ∗ and
λ∗ are called Lagrange multipliers.
20 Chapter 2 Solving Mixed-Integer Nonlinear Optimization Problems

Theorem 2.4.2 (Karush (2014) and Kuhn and Tucker (1951)):

Let the functions f, g, h of the nonlinear optimization problem (2.4.1) be continuously
differentiable functions. Let z ∗ be a local optimum of (2.4.1). If

• either the functions g and h are linear

• or the gradients of those constraints, which are fulfilled with equality, are linearly
independent

then there exist Lagrange multipliers µ∗ , λ∗ such that the vector (z ∗ , µ∗ , λ∗ ) is a KKT
point of (2.4.1).

For a feasible solution z ∗ of (2.4.1) and Lagrange multipliers µ∗ , λ∗ such that

(z ∗ , µ∗ , λ∗ ) is a KKT point we call the multipliers the dual part of the KKT point,
and also dual solution. When solving (2.4.1) by IPOpt (Wächter and Biegler 2006)
for instance, the solver does not only return a primal feasible solution but also a dual
solution, and both together form a KKT point.

2.5 A Specially Tailored Adaptation of SCIP

We are going to solve the topology optimization problem (3.2.1) as stated in Chapter 3
by SCIP. As described in Section 2.2, SCIP solves this problem by branch-and-bound,
separation and spatial branching. However, its solving performance for real-world
instances is insufficient as demonstrated in Section 1.4.
The implementation framework of SCIP allows to influence the solution process
in different directions. In the following we describe important functionalities of this
MINLP solver that we are going to exploit in this thesis. We also shortly refer
to each chapter where this functionality is used in order to improve the solving
performance of (3.2.1).
SCIP allows tracking the branching tree. For every node of this tree the user has
the possibility to influence the solving behavior. So it is for instance possible to prune
a node manually. We only prune a node if we can guarantee that the corresponding
problem of this node is infeasible. Therefor we call different routines (as described in
Chapter 4 and Chapter 6) in order to detect infeasibility of the current node. One
of these routines for example consists of solving a convex relaxation of the current
node. As described in Section 2.3 every local optimum of a convex optimization
problem is a global optimum. Using the solver IPOpt which only computes a local
optimum we are able to solve globally the problem associated with the current node
of the branching tree. If we detect infeasibility, then we prune the node. If these
2.5 A Specially Tailored Adaptation of SCIP 21

routines yield a primal feasible solution, then we add it to the solution pool of SCIP.
If possible, then SCIP itself prunes the current node.
As mentioned and described in Section 2.4 the nonlinear solver IPOpt computes a
local optimal primal solution together with a corresponding dual solution. Primal and
dual solution together form a KKT point, i.e., they fulfill the KKT conditions (2.4.2).
These conditions have a practical interpretation for our application, as discussed
in Chapter 4. This interpretation allows to derive inequalities as described in
Chapter 5, which are valid for the topology optimization problem (3.2.1). We add
these inequalities to the cut pool of SCIP, while SCIP itself manages the handling
of these inequalities.
SCIP also allows to include primal heuristics. These are handled similarly to
the already contained heuristics in terms of frequency and order. We complement
the heuristics of SCIP by another one which is specially tailored to the topology
optimization problem (3.2.1), see Chapter 7.
Chapter 3

An MINLP Model for Gas Network

Topology Optimization

In this section we present an MINLP model for the topology optimization problem.
Therefor we first describe the technical background of gas networks in Section 3.1.
Then the MINLP model is explained in Section 3.2 followed by a complexity analysis
in Section 3.3. Two different subproblems of the MINLP are defined in Section 3.4.
They form the origin of the discussions in the ongoing parts in this thesis. Finally the
computational setup is described in Section 3.5 and different (real-world) instances
are presented which we consider for our computations in the next chapters.

3.1 Technical Background

Let us give a mathematical description of the physical and technical properties of a
gas transport network. In terms of the notation in the context of graph theory we
refer to Korte and Vygen (2007).
We use a directed graph G = (V, A) to model the network, where V denotes the
set of nodes and A ⊆ V × V the set of arcs. Apart from the arc models which are
described later we introduce the following generic variables and constraints. We have
variables for the flow through each arc a that are bounded by q a and q a :

qa ∈ [q a , q a ] ∀ a ∈ A.

A positive value corresponds to a flow in the direction of the arc, and a negative
value is a flow in the opposite direction. A gas flow through a pipe is induced if the
gas has different pressures at the end nodes of the pipe. Hence each node of the gas
transport network is associated with a pressure value. We model these pressures, but,

23
24 Chapter 3 An MINLP Model for Gas Network Topology Optimization

instead of using variables which model the pressures directly, we model the squared
pressure at each node v by

πv ∈ [π v , π v ] ∀ v ∈ V.

Here π v and π v are the specified bounds for the variable. We also say that this
variable πv models the node potential at node v, and π v , π v are the node potential
bounds. Furthermore we add the pressure variable pv when needed.
The arcs of the gas transport network (V, A) represent the various elements.
Recall that we distinguish between passive network elements, namely pipes, whose
behavior cannot be influenced by the network operator, and active network elements,
namely valves, control valves, and compressors. In the following we describe each
element individually. For a more detailed description we also refer to Fügenschuh
et al. (2015).

Pipes

The majority of the arcs in a gas transport network are pipes. A precise description
of the model of a pipe is given in Pfetsch et al. (2014), Koch et al. (2015), and Geißler
(2011). We give a brief summary. A pipe is specified by its length L, diameter D and
roughness k of the pipe wall. We assume all pipes to be straight and of cylindrical
shape and restrict to the modeling of one-dimensional flow in the pipe direction x.
Under these assumptions the mass flow q is related to gas density ρ and velocity v
via
q = A ρ v, (3.1.1)

where A = D2 π/4 denotes the constant cross-sectional area of the pipe. As pipes
in Germany are usually at least one meter below the ground it is reasonable to
assume the temperature T to be constant. In such a situation isothermal flow is an
appropriate model. Now the gas flow in such a pipe is described by the following
set of nonlinear, hyperbolic partial differential equations (see Feistauer 1993; Lurie
2008), often referred to as Euler Equations:

∂ρ 1 ∂q
+ = 0, (3.1.2a)
∂t A ∂x
1 ∂q ∂p 1 ∂(q v) |v| v
+ + + gρs + λ ρ = 0. (3.1.2b)
A ∂t ∂x A ∂x 2D

The continuity equation (3.1.2a) and the momentum equation (3.1.2b) describe
the conservation of mass and the conservation of momentum, respectively. Here,
3.1 Technical Background 25

p = p(x, t), q = q(x, t), v = v(x, t) are pressure, mass flow and velocity in the direction
of the pipe depending on time t. The constant g denotes the gravitational acceleration
(with standard value 9.806 65 m s−2 ), and s ∈ [−1, 1] denotes the constant slope of
the pipe. Furthermore ρ denotes the gas density and λ the friction factor.

Since our model is intended to solve the topology expansion problem we have
to construct it from a planner’s perspective. Expansion problems are typically
considered for the long-term network planning. In contrast to, e.g., real-time optimal
control problems for gas networks, transient effects are neglected. Here a stationary
model for the gas flow is reasonable where all time derivatives are zero. In this case
the continuity equation (3.1.2a) simplifies to

∂q
= 0, (3.1.3)
∂x

which means that the gas flow is constant within a pipe.

The momentum equation (3.1.2b) relates all forces acting on gas particles to each
other. Investigating the addends on the left-hand side of the momentum equation
from left to right, the first term represents the flow rate change over time. The
second term is the pressure gradient, followed by the so-called impact pressure. The
fourth term represents the impact of gravitational forces which are influenced by the
slope of the pipe. Finally, the last term is the most important one. It represents
the friction forces acting on the gas particles due to pipe walls. These forces are
responsible for the major part of pressure drop within pipes. As the time derivatives
are zero, the first addend can be neglected. The third term ∂x (qv)/A contributes less
than 1 % to the sum of all terms under normal operating conditions, see Wilkinson
et al. (1964). So we assume that the term can be neglected, too. In conclusion,
(3.1.2b) can be written as

∂p |v| v
+ gρs + λ ρ = 0. (3.1.4)
∂x 2D

To calculate the friction factor λ we use the formula of Nikuradse (see Nikuradse
1933; Nikuradse 1950; Mischner 2012)

  !−2
D
λ= 2 log10 + 1.138 ,
k
26 Chapter 3 An MINLP Model for Gas Network Topology Optimization

which is suitable on large transport networks where we typically have to deal with
highly turbulent flows. Moreover, for the derivation of an algebraic pressure loss
equation, we have to introduce the equation of state linking gas pressure and density

ρ0 z0 T0 p
ρ= . (3.1.5)
p0 z(p, T )T

Here T0 and p0 are the norm temperature and norm pressure. The compressibility
factor z(p, T ) characterizes the deviation of a real gas from ideal gas. We assume
that this factor can be approximated by a suitable constant along the entire pipe
which we denote zm . The compressibility factor under norm conditions is denoted by
z0 and ρ0 is the gas density under norm conditions.

Lemma 3.1.1:
The solution p(x) to (3.1.4) with initial value p(0) = pin is given by

eS̃x − 1
!
2
p(x) = p2in − Λ̃ |q| q e−S̃x (3.1.6)
S̃

with
ρ0 z0 T0 p0 zm T
S̃ := 2 g s , Λ̃ := λ .
zm T p0 ρ0 z0 T0 A2 D

Proof. In (3.1.4), we replace the gas velocity v by the mass flow q using (3.1.1) and
the gas density ρ by the pressure p using the equation of state (3.1.5). This yields

∂p ρ0 z0 T0 p |q| q p0 zm T
+g s+λ = 0,
∂x p0 zm T 2 A2 D ρ0 z0 T0 p

where we use the assumption that the gas temperature and compressibility factor
are constants T and zm , respectively. Multiplication by 2 p leads to

∂ 2
p + S̃ p2 = −Λ̃ |q| q.
∂x

By substituting y = p2 we obtain the first-order linear ordinary differential equation

(ODE)
∂
y + S̃ y = −Λ̃ |q| q, y(0) = p2in .
∂x
This ODE can be solved analytically by “variation of constants” and we arrive at

1 1
 
2
y(x) = p(x) = −Λ̃ |q| q eS̃x + p2in + Λ̃ |q| q e−S̃x ,
S̃ S̃
3.1 Technical Background 27

qa

πv − πw πv − γa πw
π v − γa π w π v − γa π w

Figure 3.1: Nonlinear relation αa qa |qa |ka = πv − γa πw of arc flow and node potential differences of a
pipe a = (v, w) for αa = 1, ka = 1, and γa = 1. Increasing the arc flow goes along with increasing the
difference of the node potentials πv and πw at the end nodes v and w.

where the last two terms in parentheses represent the integration constant obtained
from the initial value y(0) = p2in . This concludes the proof for s 6= 0. For the other
case s = 0 we proceed by taking the limit for s → 0 (equivalently S̃ → 0) in (3.1.6)
using l’Hôpital’s rule.

By evaluating the solution of (3.1.6) at x = L (with p(L) = pout ) and fixing the
notation Λ := Λ̃ L and S := S̃ L, we finally obtain a well-known relationship of inlet
and outlet pressures and the mass flow through the pipe, see, e.g., Lurie (2008):

eS − 1
!
p2out = p2in − Λ |q| q e−S (3.1.7)
S

with
p0 zm T L ρ0 z0 T0 L
Λ=λ , S = 2gs . (3.1.8)
ρ0 z0 T0 A2 D p0 zm T
Recall that the variable qa ∈ R represents the arc flow for a pipe a = (v, w) ∈ A,
where a positive value is a flow from v to w, and a negative value is a flow in the
opposite direction from w to v. Furthermore the variables πv , πw model the node
potential values. The fundamental equation we use for a pipe a = (v, w) is

αa qa |qa |ka = πv − γa πw . (3.1.9)

Here αa ∈ R≥0 , ka ∈ R≥0 and γa ∈ R≥0 \ {0} are constants that subsume all
physical properties of the pipe, the flow, and the interactions of the flow with the
pipe. According to (3.1.7) and (3.1.8) we set

eS − 1
ka := 1, αa := Λ , γa := eS .
S
28 Chapter 3 An MINLP Model for Gas Network Topology Optimization

The constant γa in particular represents the height difference between nodes v and w.
If some pipelines a1 , . . . , an form a circle, it is assumed that γa1 · . . . · γan = 1. This
reflects that there is no height difference when traversing a circuit completely. If
a = (v, w) is an arc and a0 = (w, v) is its anti-parallel counterpart, then we assume
that the constants γa are such that γa = γa−1
0 . Although each arc in principle might

have a different value for ka it is natural to assume that all pipes have the same
constant in a gas network. However, this equality is no assumption for our proceeding
work. A visualization of equation (3.1.9) is given in Figure 3.1.

Valves

A valve is installed in the network to separate or to join two independent nodes.

The spatial dimension of a valve is small in comparison to the pipes. For a discrete
decision valves allow either being open or closed. In our model the node potential
values are identified when the valve is open (πv = πw ). If the valve is closed then they
are decoupled and the flow is forced to zero (qa = 0). A binary variable xa ∈ {0, 1}
distinguishes between these states. With this variable we use the following constraints
for a valve a = (v, w):

xa = 0 ⇒ qa = 0,
xa = 1 ⇒ πv = πw ,
qa , πv , πw ∈ R,
xa ∈ {0, 1}.

Compressors

At certain locations in gas transport networks it is necessary to increase the gas

pressure. For example, if the pressure is too low after a transport distance of 100 km
to 150 km, then compressors might be used to increase it again. For the mathematical
description of such an active network element, various models exist in the literature,
see Carter et al. (1993), Carter (1996), and Wu et al. (2000) for instance. We
follow the approach of De Wolf and Smeers (2000), and make use of the following
formulation for a compressor a = (v, w):

αa qa |qa |ka ≥ πv − πw . (3.1.10)

3.1 Technical Background 29

Here αa ∈ R≥0 and ka ∈ R≥0 are constants. We introduce a slack variable

ya ∈ [y a , y a ] where y a and y a are lower and upper bounds. Then we rewrite
inequality (3.1.10) as equality

αa qa |qa |ka − βa ya = πv − πw (3.1.11)

where βa ∈ R is the weight of the slack variable ya .

In practice we need further restrictions such as a minimal and a maximal pressure
difference or a restriction of the pressure difference that depends on the flow. Therefor
we allow a linear inequality system coupling the flow qa and the pressures pv , pw such
that
Aa (qa , pv , pw ) ≤ ba . (3.1.12)
T

Here it holds Aa ∈ Rνa ×3 and ba ∈ Rνa for some value νa ∈ N. When the slack
variable ya is unbounded, then the only restrictions for the compressor are given by
constraint (3.1.12). Hence (3.1.12) specifies the operating range of the compressor.
The flow can only go in positive direction through a compressor, hence a corresponding
lower bound needs to be set by this linear inequality system, i.e., qa ≥ 0.
The previous description belongs to the active mode of a compressor. Additionally,
a compressor has other operation modes and can be in bypass and closed mode. These
modes are identical with the modes of a valve. We introduce three binary variables
xa,0 , xa,1 , xa,2 ∈ {0, 1} which model the closed, bypass and active mode respectively.
With these decision variables we use the following constraints and variable bounds
for a compressor a = (v, w):

xa,0 + xa,1 + xa,2 = 1,

p2v = πv ,
p2w = πw ,
y a ≤ ya ≤ y a ,
qa , ya , πv , πw , pv , pw ∈ R,
xa ∈ {0, 1},
xa,0 = 1 ⇒ qa = 0,
xa,1 = 1 ⇒ πv = πw (3.1.13)
xa,2 = 1 ⇒ αa qa |qa | − βa ya = πv − πw , Aa (qa , pv , pw ) ≤ ba .
ka T
30 Chapter 3 An MINLP Model for Gas Network Topology Optimization

Figure 3.2: Technical arc symbols of a valve, a compressor and a control valve.

Control Valves

It can be necessary to reduce the pressure along an arc a = (v, w) in the network in
order to protect parts of the network from high pressures due to pressure limits of
older pipelines. Technically this reduction is realized by a control valve that reduces
the gas pressure. A control valve a = (v, w) is inverse to a compressor. Hence we
model it similarly to a compressor by

αa qa |qa |ka − βa ya = πv − πw . (3.1.14)

Here αa ∈ R≥0 , ka ∈ R≥0 , and βa ∈ R are constants. The difference between a

compressor and a control valve is either the sign of βa or the bounds on ya . Note that
the flow direction through a control valve is also fixed by setting the lower bound to
zero, i.e., qa ≥ 0.
Similar to a compressor, the feasible region of a control valve is restricted by
additional constraints like (3.1.12). For instance we set a minimal and maximal
pressure difference between the end nodes v and w. Furthermore bypass and closed
mode are available. Concluding, we model a control valve a = (v, w) by (3.1.13) which
is exactly the same as the model of a compressor. For a more detailed description of
a control valve we refer to Cerbe (2008).

3.2 An MINLP Model

In the previous section we presented models for the different network elements
pipelines, valves, control valves and compressors. Let us integrate these results into a
mixed-integer nonlinear programming model for the topology optimization problem.
Recall that the problem is to compute a cost-efficient selection of network elements
that have to be added to the network in order to make the nomination feasible which
is specified by the vector d.

Multigraph

We model the topology optimization problem on a multigraph. This multigraph

is obtained as follows: Given the directed graph G = (V, A) which represents the
original gas transport network, we define an extended set of arcs AX ⊆ V × V × N≥0
3.2 An MINLP Model 31

where each arc (v, w, i) ∈ AX represents the arc a = (v, w) ∈ V × V together with
index i. This set AX contains all “original” arcs from A with the additional index
1, that is, (a, 1) ∈ AX for all a ∈ A. Each valve a ∈ A is additionally represented
by the arc (a, 0) ∈ AX to indicate the status that the valve is closed. Hence the arc
(a, 1) is associated with an open valve. Each arc a ∈ A that represents a control valve
or a compressor is additionally represented by the arcs (a, 0), (a, 2) ∈ AX to indicate
the state that the active element is closed or active, respectively. The arc (a, 1) is
associated with the state bypass.
Furthermore the extended arc set AX contains possible new network elements
(pipes, valves, compressors, or control valves), where in principle a new element can
be built between any pair of existing nodes v, w ∈ V, v 6= w. A possible new extension
between nodes v and w is represented by two arcs: (v, w, 0) to indicate in the model
below that the element is not built and (v, w, 2) to model that the new element is
built. Here we assume that a new active element is only in active mode. This is a
reasonable assumption because the bypass mode of a compressor or a control valve
equals the open mode of a valve which we refer to as its active mode. Furthermore
the closed mode of these elements means that they are not added to the network.
Several loops for an existing arc a ∈ A are represented by (a, i) for index i ∈
{2, 3, . . .} to model a given set of design parameters. We always assume that arc
(a, i) contains the capacity of (a, 1) for i ≥ 2. In the model below we select exactly
one of the arcs (a, 1), (a, 2), . . . to be allowed to transport flow.
By (V, AX ) we denote the gas transport network together with its possible
extensions. Note that (V, AX ) is a graph with multiple parallel arcs and hence a
multigraph. The ongoing model is defined on this multigraph.

Variables

Let us introduce the following variables. The flow on arc (a, i) ∈ AX , i =

6 0 is denoted
by qa,i ∈ R, where a positive value means that the flow is heading in the same
direction as the arc, and a negative value indicates the opposite direction. Note that
we do not add a flow variable for an arc (a, 0) ∈ AX because this arc corresponds to
the closed state of an active element or indicates that an extension is not added to the
network. The potential value of a node v ∈ V is modeled by πv ∈ R. The pressure
itself is modeled by pv ∈ R. The variable ya,i ∈ R specifies the slack component
in (3.1.11) and (3.1.14). For pipelines and valves this variable is fixed to zero. We
introduce a binary decision variable xa,i ∈ {0, 1} for each arc (a, i) ∈ AX , where
xa,i = 1 represents the decision that arc (a, i) is used (i.e., a necessary condition for
a nonzero flow).
32 Chapter 3 An MINLP Model for Gas Network Topology Optimization

The bounds of these variables are given as parameters. For each node v ∈ V we
have lower and upper bounds on the node potential, π v , π v ∈ R with π v ≤ π v . For
each arc (a, i) ∈ AX we have lower and upper bounds on the flow, q a,i , q a,i ∈ R with
q a,i ≤ q a,i . Furthermore bounds on the weighted slack variable y a,i , y a,i ∈ R with
y a,i ≤ y a,i are given.

Constraints

We use the following constraints for modeling the topology optimization problem.
The node flow is specified by the nomination d. To ensure a consistent flow model,
flow conservation constraints

qa,i = dv
X X
qa,i − ∀v ∈ V
+ −
(a,i)∈δ (v), (a,i)∈δ (v),
AX AX
i6=0 i6=0

state that the balance of entering and leaving flows has to match exactly the nominated
amount dv at each node v ∈ V . Furthermore each arc flow variable is coupled
with the node potential values at its end nodes. More precisely, each arc (a, i) =
(v, w, i) ∈ AX , i 6= 0, which represents either a pipeline, a valve, a control valve or a
compressor, is modeled by this basic equation:

xa,i = 1 ⇒ αa,i qa,i |qa,i |ka − βa,i ya,i − (πv − γa πw ) = 0.

It is then specified by appropriate parameters αa,i , βa,i , γa and ka as described in

the previous Section 3.1. We have an arc coefficient αa,i ∈ R≥0 , a weighting factor
βa,i ∈ R≥0 for the slack variable, a coefficient γa ∈ R≥0 \{0}, and the power ka ∈ R≥0 .
If αa,i = 0 for an arc (a, i) = (v, w, i) then we assume βa,i = 0 and γa = 1 such that
the arc is modeled by πv = πw . For γa we have the following conditions: Let (V, E)
be the undirected version of (V, A) obtained by removing the orientation of the arcs
a ∈ A0 . This way each arc a ∈ A uniquely corresponds to an edge in e ∈ E and vice
versa. Consider a circuit in (V, E). Let v1 , . . . , vn be the nodes of this circuit with
vn = v1 , e1 , . . . , en−1 the edges and a1 , . . . , an−1 the corresponding arcs. Then we
assume   
n−1 n−1

 = 0.
Y Y
γa−1 
  

 γa 


i=1 i=1
ai =(vi ,vi+1 ) ai =(vi+1 ,vi )

This condition ensures a zero height difference in total when traversing the edges of
the circuit in clockwise direction. Note that those corresponding arcs ai which are
3.2 An MINLP Model 33

not oriented in clockwise direction, i.e., ai = (vi+1 , vi ), are traversed in backward

direction. Here the value γa−1
i
is regarded.
For those arcs (a, i) ∈ AX which are representing active compressors or control
valves we assume additional data being given as a νa × 3-dimensional matrix Aa and
νa -dimensional vector ba (here νa is the number of linear constraints necessary to
describe the operating range). We add the constraint

xa,i = 1 ⇒ Aa (qa,i , pv , pw ) ≤ ba .
T

For all the other arcs we set Aa = 0 and ba = 0.

To complete the model for arcs, we ensure that the flow is fixed to zero in the
case that arc (a, i) is not used, i.e.,

xa,i = 0 ⇒ qa,i = 0.

Furthermore we add the constraint

xa,i = 1
X
∀a ∈ A
i:(a,i)∈AX

to ensure that only one state of the arc is selected.

For technical reasons only a subset of all possible discrete settings is allowed.
This subset is given by X ⊆ {0, 1}AX . The set X is described by linear inequalities
in general, i.e., n o
X = x ∈ {0, 1}AX | L x ≤ t ,

where L is a matrix of integers with |AX | columns and t is vector of integers such
that the length of t is equal to the number of rows of L.

Objective

A cost coefficient ca,i ∈ R≥0 for each arc (a, i) ∈ AX reflects the costs of using
arc (a, i). We set ca,1 := 0 for all existing arcs a ∈ A and ca,0 := 0 for every active
element a ∈ A. Furthermore we set ca,2 := 0 for all compressors and control valves
a ∈ A. The objective function then is

min
X
ca,i xa,i .
(a,i)∈AX
34 Chapter 3 An MINLP Model for Gas Network Topology Optimization

MINLP Model

We summarize the previous variables and constraints. The following nonlinear non-
convex mixed-integer program with indicator constraints is used for modeling the
topology optimization problem:

min (3.2.1a)
X
ca,i xa,i
(a,i)∈AX

s. t.
∀ (a, i) ∈ AX ,
xa,i = 1 ⇒ αa,i qa,i |qa,i |ka − βa,i ya,i − (πv − γa πw ) = 0 (3.2.1b)
a = (v, w), i 6= 0,
∀ (a, i) ∈ AX ,
xa,i = 1 ⇒ Aa (qa,i , pv , pw )T ≤ ba (3.2.1c)
a = (v, w), i ≥ 2,

xa,i = 0 ⇒ qa,i = 0 ∀ (a, i) ∈ AX , i 6= 0, (3.2.1d)

xa,i = 1 (3.2.1e)
X
∀ a ∈ A,
i:(a,i)∈AX

qa,i = dv (3.2.1f)
X X
qa,i − ∀ v ∈ V,
+ −
(a,i)∈δ (v), (a,i)∈δ (v),
AX AX
i6=0 i6=0

pv |pv | − πv = 0 ∀ v ∈ V, (3.2.1g)

L x ≤ t, (3.2.1h)

πv ≤ π v ∀ v ∈ V, (3.2.1i)

πv ≥ π v ∀ v ∈ V, (3.2.1j)

qa,i ≤ q a,i ∀ (a, i) ∈ AX , i 6= 0, (3.2.1k)

qa,i ≥ q a,i ∀ (a, i) ∈ AX , i 6= 0, (3.2.1l)

xa,i = 1 ⇒ ya,i ≤ y a,i ∀ (a, i) ∈ AX , i 6= 0, (3.2.1m)

xa,i = 1 ⇒ ya,i ≥ y a,i ∀ (a, i) ∈ AX , i 6= 0, (3.2.1n)

pv , π v ∈ R ∀ v ∈ V, (3.2.1o)

qa,i , ya,i ∈ R ∀ (a, i) ∈ AX , i 6= 0, (3.2.1p)

xa,i ∈ {0, 1} ∀ (a, i) ∈ AX . (3.2.1q)

Hereinafter we refer to this model as topology optimization problem (3.2.1). The

aforementioned model can also be used in the context of water network optimization,
3.3 Complexity Analysis 35

see Bragalli et al. (2012), Burgschweiger et al. (2009), and Gleixner et al. (2012).
There the task is to operate the network and to compute a selection of pipe diameters
such that the specified nomination is feasible in the network.

Example

In Figure 3.3 we show an example of a small network to demonstrate our notation.

In the first part 3.3a, the original network (V, A) is shown. It contains five pipes
and one control valve. In 3.3b the arc flow and node potential variables are shown.
The last part 3.3c of the figure shows the decision variables x. Simple arcs, such as
(2, 1), (5, 1), (6, 1) which correspond to the original arcs 2, 5 and 6 represent pipelines
of the network. In particular, these pipelines are not extendible via loops, i.e., by
adding parallel pipes. A valve is shown in arc pair (4, 0), (4, 1). Note that there is no
flow variable q4,0 . Multiple arcs such as (1, 1), (1, 2) and (3, 1), (3, 2), (3, 3) represent
each a pipeline (which is (1, 1) and (3, 1)), together with one or two possible loop
extensions, respectively.

3.3 Complexity Analysis

Let us characterize the complexity of the topology optimization problem (3.2.1). We
reduce the NP-hard problem 3Sat, see Korte and Vygen (2007) for details, to the
topology optimization problem (3.2.1).
3Sat is defined as follows: A collection of clauses Z = {Z1 , Z2 , . . . , Zr } in
boolean variables x1 , x2 , . . . , xs is given. Each clause Zi is a disjunction of three
literals li,1 , li,2 , li,3 where a literal li,j is either a variable xk or its negation xk for
some k ∈ {1, . . . , s}. The clause Zi is satisfied if at least one of its literals is “true”.
We denote by xi,j the boolean variable of literal li,j , i.e., if li,j = xk or li,j = xk , then
xi,j = xk . The problem is to determine whether Z is satisfiable, that is, whether
there is a truth assignment to the variables x1 , x2 , . . . , xs , which simultaneously
satisfies all the clauses in Z.
Consider an instance of 3Sat. We define a corresponding instance of the topology
optimization problem. It does not contain any extension arcs, i.e., the objective
function equals zero and every feasible solution is an optimal one. The transmission
network (V, A) of the instance is shown in Figure 3.4. It has 14r vertices, where
each clause Z1 , . . . , Zr is associated with 14 vertices. Each arc a ∈ A of the network
is either a valve, or a pipe with constants αa = βa = 0 and γa = 1, i.e., the pipe
a = (v, w) is modeled by πv = πw . The lower flow bound on each arc is zero, i.e., we
36 Chapter 3 An MINLP Model for Gas Network Topology Optimization

v5
a5 v6
v2
v1
a1
a2 a6
v3

a4
a3

v4
(a) Original network with 1 valve and 5 pipes.

q1,2 π5
q5,1 π6
π2
π1
q1,1
q2,1 q6,1
π3
q4,1

q3,3
q3,1
q3,2

π4
(b) Arc flow and node potential variables q and π.

x1,2
x5,1
x1,1
x2,1 x6,1

x4,1

x4,0 x3,3
x3,1
x3,2

(c) Discrete (switching) variables x. The variables

x2,1 and x5,1 and x6,1 are fixed to 1.

Figure 3.3: Example of a network with binary and continuous variables of the associated topology op-
timization problem (3.2.1). The dashed arcs correspond to the valve in closed mode. The dotted arcs
represent loops. Loops are available for the original arcs a1 and a3 .
3.3 Complexity Analysis 37

vi,5 ai,4
ai,13
[1, 1]
vi,2 vi,11
vi,6
ai,1 [0, 0] ai,5

vi,7 ai,6 ai,10

ai,14
ai,2 [1, 1]
vi,12
+1 −1
vi,1 vi,3 vi,8 ai,11 vi,14
ai,3 [0, 0] ai,7

vi,9 ai,8 ai,12

ai,15
[1, 1]
vi,13
vi,4 vi,10
[0, 0] ai,9

vj,5 aj,4
aj,13
[1, 1]
vj,2 vj,11
vj,6
aj,1 [0, 0] aj,5

vj,7 aj,6 aj,10

aj,14
aj,2 [1, 1]
vj,12
+1 −1
vj,1 vj,3 vj,8 aj,11 vj,14
aj,3 [0, 0] aj,7

vj,9 aj,8 aj,12

[1, 1]
vj,13
vj,4 vj,10
aj,15 aj,9
[0, 0]

Figure 3.4: Two different clauses Zi , Zj with i < j of an instance of 3Sat are associated with an instance
of the topology optimization problem. The nodes and arcs associated with clause Zi are indexed by i.
Every arc is either a pipe or a valve restricted to nonnegative arc flows. The flow on the dashed arcs is
fixed to zero. The intervals that are shown give lower and upper bounds for the node potential values.
All other node potential values are unbounded. This figure exactly matches the instance Z = {Zi , Zj }
with Zi = {x1 , x2 , x3 } and Zj = {x3 , x4 , x1 }, where x1 , . . . , x4 are the boolean variables.
38 Chapter 3 An MINLP Model for Gas Network Topology Optimization

allow nonnegative arc flow only. Let us denote the set of pipes by Ap ⊆ A and the
set of valves by Ava ⊆ A. Then the MINLP model (3.2.1) for this instance writes as

∃ q, π (3.3.1)

s. t. xa,1 = 1 ⇒ πv − πw = 0 ∀ a = (v, w) ∈ A,

xa,0 = 1 ⇒ qa = 0 ∀ a ∈ A,

qa = dv
X X
qa − ∀ v ∈ V,
+
a∈δA (v) −
a∈δA (v)

xa,0 + xa,1 = 1 ∀ a ∈ Ava ,

xa,1 = 1 ∀ a ∈ Ap ,

π v ≤ πv ≤ π v ∀ v ∈ V,

0 ≤ qa ≤ q a ∀ a ∈ A,

πv ∈ R ∀ v ∈ V,

qa ∈ R ∀ a ∈ A,

xa,0 , xa,1 ∈ {0, 1} ∀ a ∈ A.

Each clause Zi is associated with a subnetwork of (V, A). The nodes and arcs
of this subnetwork are indicated with index i. It consists of nodes vi,1 , . . . , vi,14 ,
12 valves and 3 pipes. The inflow at node vi,1 equals one (dvi,1 = 1) which is led out
at node vi,14 (dvi,14 = −1). All other nodes of the subnetwork are transshipment
nodes (dvi,2 = dvi,3 = . . . = 0). The potential value at the nodes vi,5 , . . . , vi,10 is fixed
by πvi,5 = πvi,7 = πvi,9 = 1 and πvi,6 = πvi,8 = πvi,10 = 0. For the other nodes of
the subnetwork there are no restrictions on the node potential values. The topology
of two subnetworks differs only in the arcs between node sets {vi,2 , vi,5 , vi,6 } and
{vi,3 , vi,7 , vi,8 } and {vi,4 , vi,9 , vi,10 }. Here connecting arcs exist as follows:

(vi,2 , vi,5 ) ∈ A :⇔ li,1 = xi,1 ,

(vi,2 , vi,6 ) ∈ A :⇔ li,1 = xi,1 ,
(vi,3 , vi,7 ) ∈ A :⇔ li,2 = xi,2 ,
(vi,3 , vi,8 ) ∈ A :⇔ li,2 = xi,2 ,
(vi,4 , vi,9 ) ∈ A :⇔ li,3 = xi,3 ,
3.3 Complexity Analysis 39

(vi,4 , vi,10 ) ∈ A :⇔ li,3 = xi,3 .

For each pair of clauses Zi , Zj with index i < j there is an arc between the
subnetworks associated with Zi and Zj connecting vi,10+k and vj,10+k0 if and only
if xi,k = xj,k0 for each k, k 0 ∈ {1, 2, 3}. The flow on these arcs is fixed to zero, i.e.,
upper and lower flow bounds equal zero.

Theorem 3.3.1:
The topology optimization problem (3.2.1) is NP-hard.

Proof. We prove that 3Sat polynomially transforms to the topology optimization

problem (3.3.1). We consider a feasible solution of 3Sat and define a feasible solution
for (3.3.1). Initially we set a node potential of value zero for all nodes and flow value
zero for all arcs. In the discussion below we will change some of these variables to
different values. For each clause Zi we open and close valves of the corresponding
subnetwork by the following instructions:

Case xi,1 = 1: We open valve ai,4 and close ai,5 .

Case xi,1 = 0: We open valve ai,5 and close ai,4 .

Case xi,2 = 1: We open valve ai,6 and close ai,7 .

Case xi,2 = 0: We open valve ai,7 and close ai,6 .

Case xi,3 = 1: We open valve ai,8 and close ai,9 .

Case xi,3 = 0: We open valve ai,9 and close ai,8 .

Then we set the π values as follows for each clause Zi . For those πvi,· which are fixed
by its bounds we set πvi,5 := πvi,7 := πvi,9 := 1 and πvi,6 := πvi,8 := πvi,10 := 0. The
other values πvi,· which are not fixed by its bounds are set as described below:

πvi,2 := πvi,11 := xi,1 ,

πvi,3 := πvi,12 := xi,2 , (3.3.2)
πvi,4 := πvi,13 := xi,3 .

For the remaining valves we proceed by applying one of the following cases for open
and closed status:

Case 1, li,1 = 1: We open the valves ai,1 and ai,10 and close the remaining valves of
the subnetwork. We send one unit of flow from vi,1 to vi,14 along the unique
path containing ai,1 and ai,10 . We set πvi,1 := πvi,2 and πvi,14 := πvi,11 .
40 Chapter 3 An MINLP Model for Gas Network Topology Optimization

Case 2, li,1 = 0, li,2 = 1: We open the valves ai,2 and ai,11 and close the remaining
valves of the subnetwork. We send one unit of flow from vi,1 to vi,14 along the
unique path containing ai,2 and ai,11 . We set πvi,1 := πvi,3 and πvi,14 := πvi,12 .

Case 3, li,1 = 0, li,2 = 0, li,3 = 1: We open the valves ai,3 and ai,12 and close the
remaining valves of the subnetwork. We send one unit of flow from vi,1 to
vi,14 along the unique path containing ai,3 and ai,12 . We set πvi,1 := πvi,4 and
πvi,14 := πvi,13 .

We note that at least one literal is “true” which implies that exactly one of the
previous cases applies. This in turn implies that the nomination for each subnetwork
associated with Zi is transported through the subnetwork. The arcs between two
subnetworks associated with Zi and Zj for i, j ≤ r and i 6= j ensure πvi,10+k = πvj,10+k0
if xi,k = xj,k0 for k, k 0 ∈ {1, 2, 3}. These conditions are fulfilled because of (3.3.2).
Thus we obtain a feasible solution for the topology optimization problem (3.3.1).
Now we consider a feasible solution of the topology optimization problem (3.3.1)
and derive a feasible solution for 3Sat. We proceed by assigning values to the x
variables as follows:

Case 1 (valve ai,4 or ai,5 open): If ai,4 is open, then we set xi,1 := 1. Otherwise
we set xi,1 := 0. We note that at most one of the valves ai,4 and ai,5 is open
because otherwise 1 = πv1,5 = πv1,11 = πv1,6 = 0 leads to a contradiction.

Case 2 (valve ai,6 or ai,7 open): If ai,6 is open, then we set xi,2 := 1. Otherwise
we set xi,2 := 0. We note that at most one of the valves ai,6 and ai,7 is open
because otherwise 1 = πv1,7 = πv1,12 = πv1,8 = 0 leads to a contradiction.

Case 3 (valve ai,8 or ai,9 open): If ai,8 is open, then we set xi,3 := 1. Otherwise
we set xi,3 := 0. We note that at most one of the valves ai,8 and ai,9 is open
because otherwise 1 = πv1,9 = πv1,13 = πv1,10 = 0 leads to a contradiction.

At least one of these cases applies because of the construction of the network and the
nomination. The definition above is well-defined because of the arcs connecting two
different subnetworks are associated with two different clauses. From these settings
we obtain:
(case 1 applies)
qai,1 > 0 ⇒ qai,4 + qai,5 > 0 ⇒ li,1 = 1,
(case 2 applies)
qai,2 > 0 ⇒ qai,6 + qai,7 > 0 ⇒ li,2 = 1, (3.3.3)
(case 3 applies)
qai,3 > 0 ⇒ qai,8 + qai,9 > 0 ⇒ li,3 = 1.
3.4 The Passive and Active Transmission Problem 41

We note that the cases apply because no positive flow is valid on a valve in closed
state. By construction of the nomination, we have qai,1 + qai,2 + qai,3 = 1 for each
i = 1, . . . , r. Thus we obtain from (3.3.3) that li,1 ∨ li,2 ∨ li,3 is “true”. Thus clause
Zi is fulfilled for every i = 1, . . . , r. We proceed by assigning 1 to all variables xk
which are not set by the previous definitions. This does not change the status of a
clause and thus, the instance of 3Sat is feasible.

3.4 The Passive and Active Transmission Problem

The topology optimization problem (3.2.1) has two types of subproblems. We call
them passive and active transmission problems. Both problems state the problem of
transmitting gas between entries and exits through a network where the operation
modes of valves, control valves and compressors are fixed. In Chapter 4 and 6 we
present efficient solution methods for solving them. The passive transmission problem
is obtained by fixing the variables x and y in (3.2.1). We note that, if the network
contains only valves and pipes or in the more general case y = y, this problem is
obtained after fixing all binary decisions. We define the arc set A0 so that it contains
all the arcs that are allowed to carry flow, i.e., A0 := {(a, i) ∈ AX : xa,i = 1, i > 0}.
For the ease of notation we always write a ∈ A0 as abbreviation for (a, i) ∈ A0 .
Accordingly we set γa := γã and Aa := Aã and ba := bã for a = (ã, i) ∈ A0 . Further
we define β̃a := βa ya for the current value ya for each arc a ∈ A0 . Now the passive
transmission problem writes as

∃ q, π, p (3.4.1a)

s. t. αa qa |qa |ka − β̃a − (πv − γa πw ) = 0 ∀ a = (v, w) ∈ A0 , (3.4.1b)

Aa (qa , pv , pw ) ≤ ba ∀ a = (v, w) ∈ A0 , (3.4.1c)

T

qa = dv (3.4.1d)
X X
qa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

pv |pv | − πv = 0 ∀ v ∈ V, (3.4.1e)

πv ≤ π v ∀ v ∈ V, (3.4.1f)

πv ≥ π v ∀ v ∈ V, (3.4.1g)

qa ≤ q a ∀ a ∈ A0 , (3.4.1h)

qa ≥ q a ∀ a ∈ A0 , (3.4.1i)
42 Chapter 3 An MINLP Model for Gas Network Topology Optimization

pv , πv ∈ R ∀ v ∈ V, (3.4.1j)

qa ∈ R ∀ a ∈ A0 . (3.4.1k)

This problem is called passive transmission problem because the flow on every arc is
directly related to the node potential difference at the end nodes of the arc, except
those arcs a ∈ A0 where αa = 0. Hence each arc a ∈ A0 can be regarded as a pipeline
within the context of this passive transmission problem. In turn, pipelines are passive
network elements.

The active transmission problem is obtained by fixing all binary variables x in

(3.2.1). Again we denote by A0 := {(a, i) ∈ AX : xa,i = 1, i > 0} the set of arcs that
are allowed to carry flow. Now the active transmission problem writes as follows:

∃ q, π, p, y (3.4.2a)

s. t. αa qa |qa |ka − βa ya − (πv − γa πw ) = 0 ∀ a = (v, w) ∈ A0 , (3.4.2b)

qa = dv (3.4.2c)
X X
qa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

Aa (qa , pv , pw ) ≤ ba ∀ a = (v, w) ∈ A0 , (3.4.2d)

T

pv |pv | − πv = 0 ∀ v ∈ V, (3.4.2e)

πv ≤ π v ∀ v ∈ V, (3.4.2f)

πv ≥ π v ∀ v ∈ V, (3.4.2g)

qa ≤ q a ∀ a ∈ A0 , (3.4.2h)

qa ≥ q a ∀ a ∈ A0 , (3.4.2i)

ya ≤ y a ∀ a ∈ A0 , (3.4.2j)

ya ≥ y a ∀ a ∈ A0 , (3.4.2k)

pv , πv ∈ R ∀ v ∈ V, (3.4.2l)

qa , ya ∈ R ∀ a ∈ A0 . (3.4.2m)

This problem is called active transmission problem because it might contain active
elements like compressors which are in active mode. In this case the flow is not
directly related to the node potential values at the end nodes of the arc associated
3.5 Computational Setup 43

instance nodes pipes valves other active elements

net1 20 29 0 0
net2 40 39 0 6
net3 85 80 44 5
net4 143 142 103 6
net5 396 402 261 29
net6 661 614 33 42
net7 4165 3983 308 133

Table 3.1: Information about the dimension of the networks. We use seven networks with different
topologies for our computational studies. A visualization of these networks is shown in Figure 3.5 to
Figure 3.11.

with this compressor. This is the crucial difference to the passive transmission
problem.
Taking the definition of the active transmission problem we give a different view
on the topology optimization problem: The task is to compute a cost-optimal choice
of the binary variables x such that the corresponding active transmission problem is
feasible.

3.5 Computational Setup

For our computational studies seven different networks net1 – net7 with different
nominations are considered, see Figure 3.5 to Figure 3.11. These networks contain
all types of active elements for pressure regulation, i.e., valves, control valves, and
compressors. The dimensions of the underlying graphs are summarized in Table 3.1.
The networks are obtained from literature and from our cooperation partner OGE.
Recall that we distinguish between three types of networks in this thesis as
described in Section 1.5. The networks net4 and net5 with contracted compressors
and control valves form the first type of network, namely those which contain only
pipes and valves. The networks net1 and net2 with contracted active elements are
considered with additional loop arcs in our computations and form the second type
of network. The networks net3 – net7 contain compressors and control valves and
form the third type of network.
For our computations in the following chapters we imposed a time limit of 11 h for
the topology expansion instances and a time limit of 4 h for the nomination validation
instances. In the experiments our main criteria for measuring performance are the
number of solved instances, the running time, the number of branch-and-bound
nodes needed to prove optimality and the gap. To average values over all instances
of the test set, we use a shifted geometric mean. The shifted geometric mean of
values t1 , . . . , tn with shift s is defined as n (ti + s) − s. We use a shift of s = 10
pQ
44 Chapter 3 An MINLP Model for Gas Network Topology Optimization

for time, s = 100 for nodes and s = 0 for gap in order to reduce the effect of very
easy instances in the mean values. Further, using a geometric mean prevents hard
instances at or close to the time limit from overly dominating the measures. Thus
the shifted geometric mean has the advantage that it reduces the influence of outliers
in both directions.
As hardware for our computations we use a cluster of 64bit Intel Xeon X5672
CPUs at 3.20 GHz with 12 MByte cache and 48 GB main memory, running an
OpenSuse 12.1 Linux with a gcc 4.6.2 compiler. As MINLP solver we use the mixed-
integer nonlinear branch-and-bound framework SCIP 3.0.1 as already suggested in
Chapter 2. Further we use the following software packages: Cplex 12.1 (CPLEX) as
linear programming solver, and IPOpt 3.10 (Wächter and Biegler 2006) as nonlinear
solver within SCIP. SCIP reports an optimality gap which is defined as the ratio
between the best upper bound u (i.e., the objective function value of best feasible
solution) and the best lower bound `. That is, gap = ( u−`
` ) · 100 %. The framework
(Lamatto++) is used as framework integrating the different software and handling
the input data. Hyperthreading and Turboboost, special features of the Intel CPUs,
are disabled. In all computational experiments we run only one job per node to
reduce random noise in the measured running time that might be caused by delays if
multiple processes share common resources, in particular the memory bus.
3.5 Computational Setup 45

Figure 3.5: The test network net1. The data for this network can be found in (GAMS Model Library),
and computational results for this network are given by De Wolf and Smeers (2000) and De Wolf and
Bakhouya (2008). The network represents an approximation of the backbone network of the Belgium
natural gas network. For our computational experiments we will consider a single nomination which we
upscale by increasing values 2.0, 2.1, . . . in order to model an increase of the transported gas.

Figure 3.6: The test network net2. It is an approximation of parts of the German gas network in the
Rhine-Main-Ruhr area. More precisely the length and the diameters of the pipelines are real-world data
while other parameters like roughness or compressor data are set to realistic mean values. Altered data
of network net2 with similar characteristic is publicly available at URL http://gaslib.zib.de under the
name gaslib-40. For our computational experiments we will consider a single nomination which we
upscale by increasing values 2.0, 2.1, . . . in order to model an increase of the transported gas.
46 Chapter 3 An MINLP Model for Gas Network Topology Optimization

Figure 3.7: The test network net3. For almost every arc on the outer square there exists an additional
parallel arc in series with a valve which is not visible in this picture. This test network was created
by request of the cooperation partner. The aim was to analyze the optimal solution of the topology
optimization problem for several different nominations. Especially the difference between a selection
of loops and diagonal pipelines in an optimal solution was of particular interest. We consider three
different nominations which we upscale by increasing values 2.5, 3, 4, 5 in order to model an increase of
the transported gas.

Figure 3.8: The test network net4. It is an extension of the network net2. The additional arcs were
obtained manually. They represent each a pipeline in series with a valve. The length of these pipelines is
set to the geographical distance between the end nodes. Cost associated with these pipelines reflect the
building cost. For our computational experiments we will consider a single nomination which we upscale
by increasing values 1.0, 1.1, 1.2, . . . in order to model an increase of the transported gas.
3.5 Computational Setup 47

Figure 3.9: The test network net5. This network is an approximation of the German gas network for
the high calorific gas operated by the cooperation partner. More precisely the length and the diameters
of the pipelines are real-world data while other parameters like roughness or compressor data are set to
realistic mean values. The additional arcs were obtained manually. They represent each a pipeline in
series with a valve. The length of these additional pipelines is set to the geographical distance between the
end nodes. Cost associated with these pipelines reflect the building cost. Altered data of the underlying
network, which contains no extensions, is publicly available at URL http://gaslib.zib.de under the name
gaslib-135. For our computational experiments we will consider a single nomination which we upscale
by increasing values 1.0, 1.1, 1.2, . . . in order to model an increase of the transported gas.
48 Chapter 3 An MINLP Model for Gas Network Topology Optimization

Figure 3.10: The real-world network net6. Data and nominations are provided by the cooperation partner
OGE. The network is located in the northern part of Germany.
3.5 Computational Setup 49

Figure 3.11: The real-world network net7. Data and nominations are provided by the cooperation partner
OGE. The network is located in the Rhine-Main-Ruhr area of Germany.
Chapter 4

Efficiently Solving the Passive

Transmission Problem

In this chapter we focus on the topology optimization problem (3.2.1) in the case
y = y. This case includes the first type of network we consider in this thesis. Recall
that these networks consist of pipes and valves only. We present a method for solving
the passive transmission problem (3.4.1) to global optimality efficiently. As previously
discussed this problem results from fixing the variables x and y in (3.2.1) while the
fixation of y already follows from its bounds. Therefore (3.4.1) is a continuous
nonlinear feasibility problem.
We use the solution method to speed up the solution process of the topology
optimization problem (3.2.1) as follows: Recall that (3.2.1) is solved within the branch-
and-bound framework implemented by SCIP while the nonlinearity is handled by
spatial branching. We adapt this framework and solve all nodes of the branch-
and-bound tree which correspond to passive transmission problems by the efficient
solution method which we present in this chapter. Thus there is no need to continue
with spatial branching in these nodes. A similar approach is followed by Raghunathan
(2013). He considers an MINLP arising in the context of designing water networks
and globally solves subproblems of this MINLP separately. Gentilini et al. (2013)
consider the traveling salesman problem with neighborhoods as a non-convex MINLP.
Certain subproblems turn out to be convex. These are solved to global optimality in
a separate step within a branch-and-bound approach for the overall problem.
As the passive transmission problem (3.4.1) is continuous and nonlinear we want
to apply a nonlinear solver for computing an optimal solution. A problem that can
occur is that the passive transmission problem might be infeasible, which cannot
be detected efficiently by nonlinear solvers. Our solution approach is to consider
different relaxations of the passive transmission problem, which are either feasible or
their infeasibility can be detected in a preprocessing step. All relaxations neglect the

51
52 Chapter 4 Efficiently Solving the Passive Transmission Problem

constraints (3.4.1c) and (3.4.1e). Further they successively relax all other constraints:
The first relaxation is obtained by relaxing the variable bounds (3.4.1f)–(3.4.1i),
the second one by relaxing the flow conservation constraints (3.4.1d) and the third
one by relaxing the potential-flow-coupling constraints (3.4.1b). It turns out that
under some assumptions two of these relaxations are convex and hence can be solved
efficiently by a nonlinear solver.
As highlighted earlier we solve the topology optimization problem by a special
tailored version of SCIP. It turns out that this approach enables around 29 % more
instances of the first type of network to be solved to global optimality within our
given time frame compared to solving the overall topology optimization problem by
branch-and-bound, separation, and spatial branching. On average the run time is
reduced by 72 % on those instances which are globally solvable by SCIP. In the case
that a passive transmission problem is infeasible, the dual solutions of the convex
relaxations provide an explanation for the infeasibility.
The outline of this chapter is as follows: In Section 4.1 we state the passive
transmission problem again to improve readability. In Sections 4.2 - 4.4 we present
three different relaxations of the passive transmission problem where we assume
αa > 0 for all arcs a ∈ A0 . In addition all relaxations neglect constraints (3.4.1c)
and (3.4.1e). In Section 4.5 we show how to handle the general case without these
assumptions. Finally computational results are given in Section 4.6.

4.1 Notation
Let us recall the passive transmission problem (3.4.1). It is a subproblem of the
topology optimization problem (3.2.1) obtained by fixing all binary variables x and all
slack variables y of compressors and control valves. We write the passive transmission
problem as

∃ q, π, p (4.1.1a)

s. t. αa qa |qa |ka − β̃a − (πv − γa πw ) = 0 ∀ a = (v, w) ∈ A0 , (4.1.1b)

Aa (qa , pv , pw ) ≤ ba ∀ a = (v, w) ∈ A0 , (4.1.1c)

T

qa = dv (4.1.1d)
X X
qa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

pv |pv | − πv = 0 ∀ v ∈ V, (4.1.1e)

πv ≤ π v ∀ v ∈ V, (4.1.1f)
4.2 Relaxation of Domains 53

πv ≥ π v ∀ v ∈ V, (4.1.1g)

qa ≤ q a ∀ a ∈ A0 , (4.1.1h)

qa ≥ q a ∀ a ∈ A0 , (4.1.1i)

pv , πv ∈ R ∀ v ∈ V, (4.1.1j)

qa ∈ R ∀ a ∈ A0 . (4.1.1k)

Here β̃a := βa ya for the current value ya . Recall that the arc set A0 := {(a, i) ∈ AX :
xa,i = 1, i > 0} contains all arcs such that the flow is not fixed to zero. Throughout
this chapter we assume w.l.o.g. that (V, A0 ) is a connected graph. If this is not the
case, we split the problem such that we obtain a passive transmission problem for
each connected component.

4.2 Relaxation of Domains

Let us consider the following domain relaxation of the passive transmission prob-
lem (4.1.1) with αa > 0 for each arc a ∈ A0 . We introduce a slack variable ∆v ∈ R≥0
in order to relax the potential value of node v ∈ V and another slack variable
∆a ∈ R≥0 for relaxing the flow of arc a ∈ A0 . Then the domain relaxation writes as

min ∆v + ∆a (4.2.1a)
X X

v∈V a∈A0

s. t. αa qa |qa |ka − β̃a − (πv − γa πw ) = 0 ∀ a = (v, w) ∈ A0 , (4.2.1b)

qa = dv (4.2.1c)
X X
qa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

π v − ∆v ≤ π v ∀ v ∈ V, (4.2.1d)

πv + ∆ v ≥ π v ∀ v ∈ V, (4.2.1e)

qa − ∆a ≤ q a ∀ a ∈ A0 , (4.2.1f)

qa + ∆a ≥ q a ∀ a ∈ A0 , (4.2.1g)

πv ∈ R ∀ v ∈ V, (4.2.1h)

qa ∈ R ∀ a ∈ A0 , (4.2.1i)

∆v ∈ R≥0 ∀ v ∈ V, (4.2.1j)
54 Chapter 4 Efficiently Solving the Passive Transmission Problem

∆a ∈ R≥0 ∀ a ∈ A0 . (4.2.1k)

In the following we show that this nonlinear optimization problem is feasible and
convex, see Section 4.2.1 and Section 4.2.2. Hence it can be solved very efficiently to
global optimality. Note that the convexity here is not given by the constraints but
by the feasible solution space of the relaxation. Finally in Section 4.2.3 we will give
an interpretation of the dual solution of a KKT point of (4.2.1). This dual solution
forms a network flow that is induced by dual node potentials. It fulfills constraints
that are similar to (4.2.1b) and (4.2.1c). In the case that the passive transmission
problem (4.1.1) is infeasible, this dual solution allows to filter those parts of the
network (V, A0 ) which imply the infeasibility.

4.2.1 Existence of a Solution

The existence of a primal solution for (4.2.1) was shown by Collins et al. (1978) and
Maugis (1977). In the following we review their method which basically consists
of solving the nonlinear optimization problems (4.2.2) and (4.2.3). Note that their
approach only works for constant heights, i.e., γa = 1 for all a ∈ A0 . In the subsequent
part of this section we extend their method and show how the case of inhomogeneous
heights, i.e., γa 6= 1 for some a ∈ A0 , can be treated.
Collins et al. considered the convex nonlinear optimization problem

XZ qa
min Φa (t) dt
0
qa
a∈A0

s. t. qa = −dv
X X
qa − ∀ v ∈ V, (4.2.2)
+
0 (v) 0 (v)
−
a∈δA a∈δA

qa ∈ R ∀ a ∈ A0 ,

where Φa (·) is a continuous strictly monotone function. Further qa0 is a root of

Φa (·) which implies that the objective is convex. In the context of our study we
set Φa (qa ) := αa qa |qa |ka − β̃a . Then Φa (qa ) is strictly monotone increasing because
αa > 0 by our assumption.

Lemma 4.2.1 (Collins et al. (1978) and Maugis (1977)):

The nonlinear optimization problem (4.2.2) is convex. Its optimal solution yields a
feasible solution for (4.2.1) in the case γa = 1 for each arc a ∈ A0 .

Proof. The constraints of (4.2.2) are linear and the objective is a sum of convex
functions. Hence (4.2.2) is a convex optimization problem.
4.2 Relaxation of Domains 55

Furthermore the objective and all constraints of (4.2.2) are continuously differen-
tiable. From Theorem 2.4.2 and (2.4.2a) of the KKT conditions (2.4.2) for (4.2.2)
we obtain for optimal primal values q ∗ that there exist dual values µ∗ such that

µ∗v − µ∗w = Φa (qa∗ )

for each arc a = (v, w) ∈ A0 . By setting the node potential π ∗ := µ∗ , and ∆∗v :=
max{0, πv∗ − π v , π v − πv∗ } for each node v ∈ V , and ∆∗a := max{0, qa∗ − q a , q a − qa∗ }
for each arc a ∈ A0 we obtain a primal feasible solution (q ∗ , π ∗ , ∆∗ ) of (4.2.1) with
γa = 1 (for all a ∈ A0 ).

We note that the convex optimization problem (4.2.2) is not only useful for
theoretical purpose. Raghunathan (2013) uses (4.2.2) for computing a solution of
the passive transmission problem (4.1.1) on a real-world application.
We further remark that Collins et al. (1978) provide a different proof for the
existence of a solution for (4.2.1) (with γa = 1) by considering the following nonlinear
program: Z πv −πw X Z πv
min Φa (t) dt − dv dt
X
−1

a=(v,w)∈A0 ∆a
0
v∈V 0 (4.2.3)
s. t. πv ∈ R ∀ v ∈ V.
0
Here Φ−1
a is the inverse of Φa and ∆a is a root of Φa .
−1

Lemma 4.2.2 (Collins et al. (1978) and Maugis (1977)):

The nonlinear optimization problem (4.2.3) is convex. Its optimal solution yields a
feasible solution for (4.2.1) in the case γa = 1 for each arc a ∈ A0 .

Note that problem (4.2.3) might be unbounded. In this case there exists no
optimal solution and hence we cannot ensure the feasibility of (4.2.1) by making use
of Lemma 4.2.2.

Proof. The objective is a sum of convex functions. Hence (4.2.3) is convex. For an
optimal solution π ∗ of (4.2.3) we define q ∗ by

qa∗ := Φ−1
a (πv − πw )
∗ ∗
⇔ Φa (qa∗ ) = πv∗ − πw
∗

for each arc a = (v, w) ∈ A0 while Φ−1

a is the inverse function of Φa . Furthermore the
objective of (4.2.3) is continuously differentiable. From Theorem 2.4.2 and (2.4.2a) of
56 Chapter 4 Efficiently Solving the Passive Transmission Problem

the KKT conditions (2.4.2) for (4.2.3) we obtain that the following flow conservation
constraints are fulfilled:

qa∗ − dv = 0
X X
qa∗ − ∀ v ∈ V.
+
0 (v) 0 (v)
−
a∈δA a∈δA

By setting ∆∗v := max{0, πv∗ − π v , π v − πv∗ } for each node v ∈ V and ∆∗a :=
max{0, qa∗ − q a , q a − qa∗ } for each arc a ∈ A0 we obtain a primal feasible solution
(q ∗ , π ∗ , ∆∗ ) of (4.2.1) with γa = 1 (for all a ∈ A0 ).

After this brief literary review we now turn to the general case γa 6= 1 for some
arc a ∈ A0 . We show how to obtain a feasible solution for domain relaxation (4.2.1)
by using the convex optimization problem (4.2.2).

Definition 4.2.3:
Let (V, E) be the undirected version of (V, A0 ) obtained by removing the orientation
of the arcs a ∈ A0 . This way each arc a ∈ A0 uniquely corresponds to an edge in
e ∈ E and vice versa. Let r be any node in V . For a node v ∈ V denote by Pr (v)
an undirected path from r to v. Let v1 , . . . , vn be the nodes of this path, e1 , . . . , en−1
the edges and a1 , . . . , an−1 the corresponding arcs. Recall our assumption for this
chapter that (V, A0 ) is connected. We define
  

γr,v := 
Y Y
γai   γa−1 .
  
i
i:ai =(vi ,vi+1 ) i:ai =(vi+1 ,vi )

We have γr,v > 0 as γa > 0 for every arc a ∈ A0 . The definition is actually
independent of the path Pr (v). To see this, let P 0 be a different r-v-path in (V, E).
Consider the cycle C from r to v on path P , and back from v to r on path P 0 in
reverse order. Denote the reverse path of P 0 by Q0 . Denote the nodes of this path
by ṽ1 , . . . , ṽm with ṽ1 = v and ṽm = r, the edges by ẽ1 , . . . , ẽm−1 and the arcs by
ã1 , . . . , ãm−1 . According to our assumption in Section 3.1 we have
  

1 =
Y Y
γa   γa−1 
  
i:ai =(vi ,vi+1 ) i:ai =(vi+1 ,vi )
  
Y Y
γa   γa−1  .
  

i:ãi =(ṽi+1 ,ṽi ) i:ãi =(ṽi ,ṽi+1 )
4.2 Relaxation of Domains 57

Hence γr,v is uniquely defined.

Using this value γr,v we define the function πv0 by

πv0 (π) := γr,v πv (4.2.4)

for every node v ∈ V . As a consequence of (4.2.4) we obtain lower and upper bounds
of πv0 (π) from π 0v := πv0 (π) and π 0v := πv0 (π), respectively, for each node v ∈ V .

It follows from elementary calculations that

πv0 (π) − πw
0
(π) = γr,v πv − γr,w πw = γr,v (πv − γa πw ) (4.2.5)
|{z}
=γr,v γa

holds for each arc a = (v, w) ∈ A0 .

We use the previous definitions and show how to obtain a feasible solution for
the domain relaxation (4.2.1) in the general case γa 6= 1 for some arc a ∈ A0 .

Lemma 4.2.4:
The optimization problem (4.2.1) is feasible.

Proof. We use γr,v from Definition 4.2.3 and equation (4.2.5). Now we compute a
local optimum q ∗ of the problem

XZ qa
min Φa (t) dt
0
qa
a∈A0

s. t. qa = −dv
X X
qa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

qa ∈ R ∀ a ∈ A0 ,

where Φa (qa ) := γr,v αa qa |qa |ka − γr,v β̃a and qa0 is a root of this function for each arc
a = (v, w) ∈ A0 .
By Lemma 4.2.1 this optimization problem yields a feasible solution (q ∗ , π ∗ ) for
a modified version of the domain relaxation (4.2.1) which is obtained by replacing
the constraints (4.2.1b), (4.2.1d), (4.2.1e) by

γr,v αa qa |qa |ka − γr,v β̃a − (πv − πw ) = 0 ∀ a = (v, w) ∈ A0 ,

πv − ∆v ≤ πv0 (π) ∀ v ∈ V,

πv + ∆v ≥ πv0 (π) ∀ v ∈ V,
58 Chapter 4 Efficiently Solving the Passive Transmission Problem

where we used (4.2.4).

Using (q ∗ , π ∗ ) we are going to show how to obtain a feasible solution for the
domain relaxation (4.2.1) where the previous modifications do not take place. We
define
π̂v := πv0 (π ∗ )
−1

for each node v ∈ V . Recall that γr,v > 0 because γa > 0 for each arc a ∈ A. This
implies that π 0 is well defined. In combination with (4.2.5) we obtain
−1

αa qa∗ |qa∗ |ka − β̃a = γr,v

−1
(πv∗ − πw
∗
) = π̂v − γa π̂w ∀ a = (v, w) ∈ A0 .

Thus (q ∗ , π̂) is a vector which is feasible for the constraints (4.2.1b) and (4.2.1c) of
the original domain relaxation (4.2.1). Setting ∆∗v := max{0, π̂v − π v , π v − π̂v } for
each node v ∈ V and ∆∗a := max{0, qa∗ − q a , q a − qa∗ } for each arc a ∈ A0 we obtain a
primal feasible solution (q ∗ , π̂, ∆∗ ) of (4.2.1).

4.2.2 Characterization of the Feasible Region

In the previous section we proved that the domain relaxation (4.2.1) is feasible. In
the following we are going to prove that (4.2.1) is a convex optimization problem.
First we show that the solution flow q is unique and the feasible node potentials π
form a straight line. From this we conclude that (4.2.1) is convex.
By Maugis (1977) the uniqueness of the flow vector q follows from the strict
convexity of the objectives of (4.2.2) and (4.2.3) in the case γa = 1 for each arc
a ∈ A0 . Nevertheless, we give a proof for the more general case that there exists an
arc a ∈ A0 with γa 6= 1. For this proof we make use of the following theorem and the
lemma below:

Theorem 4.2.5:
≥0 be a network flow in the (V, A ). There exists a family P of paths and a
0
Let q ∈ RA 0

family C of circuits in (V, A0 ) along with values fP > 0, P ∈ P and fC > 0, C ∈ C

such that
qa = fP0 +
X X
fC0 ∀ a ∈ A0 .
P ∈P: C∈C:
a∈A0 (P ) a∈A0 (C)

If q is a circulation then it holds P = ∅.

Proof. Let q be a network flow between two nodes s and t, i.e., an s-t-flow. In this
case the theorem follows from Korte and Vygen (2007), Theorem 8.8.
4.2 Relaxation of Domains 59

The more general case of multiple sources and sinks is obtained by adding a super
node s for the sources and another super node t for the sinks. We set

q̃(s,v) := q̃(w,t) :=
X X X X
qa − qa qa − qa
+ +
0 (v) 0 (v) 0 (w) 0 (w)
− −
a∈δA a∈δA a∈δA a∈δA

for the sources v and sinks w and q̃a := qa for all arcs a ∈ A0 . For each source v ∈ V
we add the arc (s, v) and for each sink w ∈ V we add the arc (w, t) to the graph
(V, A0 ). Then we apply Theorem 8.8 to the extended graph (V, A0 ) and the network
flow q̃ therein.

Lemma 4.2.6:
Let (q 0 , π 0 , ∆0 ) and (q 00 , π 00 , ∆00 ) be two feasible solutions of the domain relaxation
(4.2.1). W.l.o.g. it holds q 0 ≥ q 00 .

Proof. Let (q 0 , π 0 , ∆0 ) and (q 00 , π 00 , ∆00 ) be two solutions of (4.2.1). Recall that the
network (V, A0 ) is connected. If there exists an arc a with qa0 < qa00 then we reorient
this arc a = (v, w) as follows: We remove it from the graph and add the backward
(or antiparallel) arc a = (w, v) to A0 . From the constraint

αa qa |qa |ka − βa ya − (πv − γa πw ) = 0

we obtain with

α a := αa γa−1 ∈ R≥0 , β a := βa γa−1 ∈ R≥0 , γ a := γa−1 ∈ R≥0 \ {0},

y a := −y a ∈ R, y a := −y a ∈ R, q a := −q a ∈ R, q a := −q a ∈ R

the equivalent equation for the backward arc a

α a q a |q a |ka − β a y a − (πw − γ a πv ) = 0. (4.2.6)

Note that we used the relation γ a γa = 1. Hence constraint (4.2.1b) for arc a
is replaced by constraint (4.2.6) for arc a. Note that this constraint fulfills our
assumption for the constants α a , β a and γ a as described in Section 3.2 where the
constraints are explained. Note that replacing a by a goes along with the relation
q a = −qa for the arc flow variables. For a solution of (q ∗ , π ∗ , ∆∗ ) of the original
˜ ± := ∆± ∗ if we did not change the
domain relaxation we define q̃ by q̃a := qa∗ and ∆ a a
˜ + := ∆− ∗ and ∆
orientation of a and q̃a := −q ∗ , ∆ ˜ − := ∆+ ∗ if we adapted the arc
a a a a a
˜ ± := ∆± ∗ for every node v ∈ V . Then it holds that
orientation. Further we set ∆ v v
60 Chapter 4 Efficiently Solving the Passive Transmission Problem

(q ∗ , π ∗ , ∆∗ ) is feasible for the domain relaxation on the original graph (V, A0 ) if and
only if (q̃, π ∗ , ∆)˜ is feasible for the domain relaxation on the modified graph.

We adapt the direction of all arcs a ∈ A0 with qa0 < qa00 and end up with a
modified version of the passive transmission problem of the same type as the original
˜ 0 ) and
passive transmission problem. Further we have two feasible solutions (q̃ 0 , π 0 , ∆
˜ 00 ).
(q̃ 00 , π 00 , ∆

Lemma 4.2.7:
0
There exist vectors q̃ ∈ RA , π̃ ∈ RV and θ ∈ RV≥0 , such that the following holds:
(q ∗ , π ∗ , ∆∗ ) is feasible for problem (4.2.1) if and only if there exists t ∈ R with
π ∗ = π̃ + tθ and q ∗ = q̃.

Proof. It follows from Lemma 4.2.4 that (4.2.1) is feasible. Let us assume the
existence of two different solutions (q 0 , π 0 , ∆0 ) and (q 00 , π 00 , ∆00 ) of (4.2.1) with q 0 6= q 00
and π 0 =
6 π 00 . First we prove by contradiction that the primal solution flow of (4.2.1)
is unique. In a second step we analyze the difference π 0 − π 00 .

The difference q 0 − q 00 is a network flow in (V, A0 ) consisting of circulations only.

First we focus on the case q 0 ≥ q 00 . By Theorem 4.2.5 the difference q 0 − q 00 can be
split into flow along circuits C1 , . . . , Cn . Thus we obtain from q 0 ≥ q 00 , that there
exist flow values qCi > 0, i = 1, . . . , n such that

qa0 − qa00 =
X
0
qC i
∀ a ∈ A0 .
i=1,...,n:
a∈A0 (Ci )

Note that the sum can be empty for an arc a ∈ A0 . We consider a single circuit C
out of C1 , . . . , Cn . For each arc a ∈ A0 (C) we have qa0 > qa00 . As abbreviation we set
Φa (q) := αa q|q|ka − β̃a . Then Φa is a continuous strictly monotone function. We
use equation (4.2.1b) to obtain for all arcs a = (v, w) ∈ A0 (C):

qa0 > qa00 ⇔ Φa (qa0 ) > Φa (qa00 ) ⇔ πv0 − γa πw

0
> πv00 − γa πw
00
.

Let the nodes of the circuit C be ordered such that V (C) = {v1 , . . . , v` } and (as
abbreviation) v`+1 := v1 , and the arcs be ordered such that ai = (vi , vi+1 ) holds. We
consider all arcs of the circuit C and rewrite the following telescopic sums as
   
`
X i−1
Y   X` i−1
Y  
 γaj  πv0 i − γai πv0 i+1 >  γaj  πv00i − γai πv00i+1
i=1 j=1 i=1 j=1
4.2 Relaxation of Domains 61

   

πv0 1 1 − γa  > πv001 1 −

Y Y
⇔ γa  .
a∈A0 (C) a∈A0 (C)

Note that γa = 1 (see Section 3.2). Hence we have

Q
a∈A0 (C)

qa0 > qa00 , a ∈ A0 (C) ⇔ πv0 1 · 0 > πv001 · 0,

which implies that q 0 6= q 00 is infeasible. This contradicts our assumption q 0 6= q 00 . So

the solution flow q 0 = q 00 of (4.2.1) is unique.

So far we considered the case q 0 ≥ q 00 . If there exists an arc a ∈ A0 with qa0 < qa00
then we apply Lemma 4.2.6 and change the orientation of this arc. When reorienting
we obtain two flow vectors q̃ 0 and q̃ 00 which differ from q 0 and q 00 only in the signs of
the reoriented arcs. For these new flow vectors we apply the result from above and
obtain q̃ 0 = q̃ 00 . Coming back to the original orientation we change back the signs
and obtain q 0 = q 00 .

Now we prove that the feasible node potentials π of (4.2.1) form a straight line.
Therefor let (q 0 , π 0 , ∆0 ) and (q 0 , π 00 , ∆00 ) be two different feasible solutions of (4.2.1).
Let r be some node r ∈ V . Let w ∈ V be any other node in V . Consider the graph
(V, E) which is obtained from (V, A0 ) by removing the orientation. This way each arc
a ∈ A0 uniquely corresponds to an edge e ∈ E. Consider an undirected r-w-path P
in (V, E) with nodes r = v1 , . . . , vk = w, edges e1 , . . . , ek−1 and corresponding arcs
a1 , . . . , ak−1 . For an arc aj = (vj+1 , vj ) we rewrite equation (4.2.1b) as

−γa−1
j
Φaj (qaj ) = πvj − γa−1
j
πvj+1 .

For all other corresponding arcs we do not apply this reformulation. As abbreviation
we set
` `
τ` :=
Y Y
γaj γa−1
j
.
j=1 j=1
aj =(vj ,vj+1 ) aj =(vj+1 ,vj )

Then we obtain the equality

 
k−1 j−1
τk−1 = τi  Φaj (qaj )
X Y
πr0 − πw
0 
j=1: i=1
aj =(vj ,vj+1 )
 
k−1 j−1
Φaj (qaj ) = πr00 − πw
X Y
−  τi  γa−1
j
00
τk−1 .
j=1: i=1
aj =(vj+1 ,vj )
62 Chapter 4 Efficiently Solving the Passive Transmission Problem

This is equivalent to
πr0 − πr00 = (πw
0 00
− πw ) τk−1 .

We define θw := τk−1
−1
6= 0. This setting is well-defined, i.e., independent of the actual
path P from r to w as discussed in Definition 4.2.3. We set t := πr00 − πr0 . Then
the solution π 00 can be expressed as πw
00
= πw
0
+ tθw for all w ∈ V . This proves the
lemma.

Corollary 4.2.8:
The feasible solution vectors (q, π) of the domain relaxation (4.2.1) form a convex
space: The feasible flow q is unique, while the feasible node potential values π form a
straight line.

Proof. This follows from Lemma 4.2.7.

Let us briefly discuss the result of this corollary. Figure 4.1 shows the node
potential values for a test network having 34 nodes. The node potential bounds are
shown as straight lines (lower bound of 500 and upper bound of 6000). Solutions
for four different passive transmission problems are shown in different colors: Each
dot represents the node potential value at the respective node. By Lemma 4.2.7 the
feasible node potentials for the passive transmission problem are on a straight line.
This corresponds to shifting all dots in Figure 4.1, which correspond to the same
passive transmission problem, up or down at the same time. For three of the four
problems it is not possible to move all node potential values inside the bounds, hence
these solutions are infeasible. Only for the solution corresponding to the circles, all
values are inside the bounds and this solution is feasible.
Now, the combination of Lemma 4.2.4 and Corollary 4.2.8 allows us to characterize
the domain relaxation (4.2.1).

Theorem 4.2.9:
The domain relaxation (4.2.1) is a feasible and convex relaxation of the passive
transmission problem (4.1.1) in the case αa > 0 for all arcs a ∈ A0 .

Proof. A solution (q ∗ , π ∗ , p∗ ) of the passive transmission problem (4.1.1) is feasible

only if (q ∗ , π ∗ , 0) is a feasible solution of (4.2.1). Hence (4.2.1) is a relaxation of the
passive transmission problem (4.1.1).
It follows from Lemma 4.2.4 that the domain relaxation (4.2.1) is feasible. The
convexity of the feasible solution space for (q, π) follows from Corollary 4.2.8, which
states that the solution space is an affine subspace. The flow is unique, while the
feasible vectors π form a straight line. The additional constraints (4.2.1d)-(4.2.1g)
4.2 Relaxation of Domains 63

6,000

Node potential solution value

4,000

2,000

0 5 10 15 20 25 30 35

Node index

Figure 4.1: Feasible node potential values π for the domain relaxation (4.2.1) of four different passive
transmission problems (4.1.1) on a gas network with 32 nodes. The node potential values are feasi-
ble for the passive transmission problems within the marked bounds, other values are infeasible. By
Corollary 4.2.8 the node potentials π form a straight line. This corresponds to shifting all dots, which
correspond to the same passive transmission problem, up or down at the same time.

are linear and do not affect the feasibility (i.e., rendering the problem infeasible) and
the affine subspace. In total, problem (4.2.1) is a convex optimization problem over
a non-empty set of feasible solutions.

In the following Lemma we turn to the more general case where we do not assume
αa > 0 for the arcs a ∈ A0 . This implies that constraint (4.2.1b) can be written as
πv = πw for an arc (v, w) ∈ A0 .

Lemma 4.2.10:
The domain relaxation (4.2.1) is a feasible and convex optimization problem in the
general case that there exist arcs (v, w) ∈ A0 where constraint (4.2.1b) writes as
πv = πw .

Proof. The feasibility in the more general case follows from Lemma 4.2.2 because
the proof does not depend on the assumption αa > 0.
The convexity is seen as follows: Let A00 denote the set of arcs a = (v, w) ∈ A0
which are not modeled by πv = πw , i.e., A00 = {a ∈ A0 | αa > 0}. For a solution
(q ∗ , π ∗ , ∆∗ ) of the domain relaxation (4.2.1) it holds that qa∗ is unique for all arcs
a ∈ A00 . Otherwise, if there exist two solution flows q 0 , q 00 of the domain relaxation
64 Chapter 4 Efficiently Solving the Passive Transmission Problem

with qa0 0 6= qa000 for an arc a0 ∈ A00 , then (qa0 0 )a0 ∈A00 and (qa000 )a0 ∈A00 would be two
different feasible solution vectors for the domain relaxation of the preprocessed
passive transmission problem where all arcs a ∈ A0 \ A00 are contracted. This is
a contradiction to Lemma 4.2.7. It follows that the domain relaxation is a linear
program, when fixing the unique arc flow qa on all arcs a = (v, w) ∈ A00 . This proves
the convexity.

4.2.3 Interpretation of Lagrange Multipliers

Assume that the passive transmission problem (4.1.1) is infeasible and the domain re-
laxation (4.2.1) has a positive optimal objective value. It turns out (see Lemma 4.2.11)
that there exist Lagrange multipliers for the optimal solution such that the optimal
solution and the multipliers form a KKT point of (4.2.1). These multipliers have
a practical interpretation. They form a generalized network flow in (V, A0 ) which
is coupled with node potentials, similar to a primal solution (q ∗ , π ∗ ) of the passive
transmission problem. This network flow may characterize the infeasibility as dis-
cussed in Example 4.2.13, and hence allows a visualization of the conflicting parts of
the network yielding the infeasibility.

Lemma 4.2.11:
Let (q ∗ , π ∗ , ∆∗ ) be an optimal solution of the domain relaxation (4.2.1). There exist
Lagrange multipliers (µ∗ , λ∗ ) which consecutively correspond to the equality and
inequality constraints of (4.2.1), respectively, such that (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) is a KKT
point of (4.2.1). These multipliers are characterized as follows: (µ∗a )a∈A0 is a general
network flow in (V, A0 ) which is induced by dual node potentials (µ∗v )v∈V . More
precisely the multipliers (µ∗ , λ∗ ) are a feasible solution for

dΦa ∗
µa (q ) + λ+
a − λa = µv − µw
−
∀ a = (v, w) ∈ A0 ,
dqa a

γa µa = λ+
X X
−
µa − v − λv ∀ v ∈ V.
+
0 (v) 0 (v)
−
a∈δA a∈δA

Hereby the dual node flow λ+

v − λv is restricted by
−

exit innode entry

λ+
v ∈ {0} [0, 1] {1}
λ−
v ∈ {1} [0, 1] {0}
node potential πv∗ < πv πv∗ = πv π v < πv∗ πv∗ < π v πv∗ = π v πv∗ > π v
4.2 Relaxation of Domains
. 65

for each node v ∈ V , and the dual compression λ+

a − λa is constrained by
−

compression neutral regulation

λ+
a ∈ {0} [0, 1] {1}
λ−
a ∈ {1} [0, 1] {0}
flow qa∗ < q a qa∗ = q a qa∗ > q a qa∗ < q a qa∗ = q a qa∗ > q a

for each arc a ∈ A0 .

Proof. Let (q ∗ , π ∗ , ∆∗ ) be an optimal solution of (4.2.1). Recall that every locally

optimal solution is globally optimal because of the convexity of (4.2.1) by Theo-
rem 4.2.9. The objective and the constraints of (4.2.1) are continuously differentiable.
Assume there exist dual values (µ∗ , λ∗ ) such that (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) is a KKT point
of domain relaxation (4.2.1). A proof of existence follows from the next Lemma 4.2.12.
Let us write the conditions (2.4.2a) of the KKT conditions (2.4.2). We denote the La-
grange multipliers by (µ, λ) = (µv , µa , λ+ +
v , λv , λa , λa )v∈V,a∈A0 , such that µv , µa ∈ R
− −

and λ+ − +
v , λv , λa , λa ∈ R≥0 . We set Φa (qa ) := αa qa |qa |
− ka
− β̃a . Then the Lagrange
function of problem (4.2.1) has the form

L(q, π, ∆, µ, λ) = ∆v + ∆a
X X

v∈V a∈A0

+ µa Φa (qa ) − (πv − γa πw )
X

a∈A0
a=(v,w)
 

+ qa +
X X X
µv dv − qa 
 
+
0 (v) 0 (v)
v∈V −
a∈δA a∈δA

λ+
X 
+ v (πv − ∆v − π v ) + λv (π v − πv − ∆v )
−

v∈V

λ+
X 
+ a (qa − ∆a − q a ) + λa (q a − qa − ∆a )
−

a∈A0

λv ∆v − λa ∆a .
X X
−
v∈V a∈A0

From (2.4.2a) we obtain that the KKT point (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) is feasible for

∂L dΦa ∗
=0⇒ µa (q ) + λ+
a − λa = µv − µw
−
∀ a = (v, w) ∈ A0 , (4.2.7a)
∂qa dqa a
∂L
=0⇒ µa γa = λ+ (4.2.7b)
X X
−
µa − v − λv ∀ v ∈ V,
∂πv +
0 (v) 0 (v)
−
a∈δA a∈δA
66 Chapter 4 Efficiently Solving the Passive Transmission Problem

∂L
=0⇒ λ+
v + λv + λv = 1
−
∀ v ∈ V, (4.2.7c)
∂∆v
∂L
=0⇒ λ+
a + λa + λa = 1
−
∀ a ∈ A0 . (4.2.7d)
∂∆a

We conclude from (4.2.7c), (4.2.7d), and the complementarity condition (2.4.2e) that
(q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) fulfills:

v = 0,
πv > π v ⇒ λ − πv < π v ⇒ λ+
v = 0,

v ∈ [0, 1],
πv = π v ⇒ λ − πv = π v ⇒ λ+
v ∈ [0, 1], ∀v ∈ V (4.2.8a)
πv < π v ⇒ λ−
v = 1, πv > π v ⇒ λ+
v = 1,

a = 0,
qa > q a ⇒ λ− qa < q a ⇒ λ+
a = 0,

qa = q a ⇒ λ−
a ∈ [0, 1], qa = q a ⇒ λ+
a ∈ [0, 1], ∀ a ∈ A0 (4.2.8b)
qa < q a ⇒ λ−
a = 1, qa > q a ⇒ λ+
a = 1.

The tables of the Lemma follow from (4.2.8a) and (4.2.8b), respectively, and the
constraints follow from (4.2.7a) and (4.2.7b).

The interpretation of these conditions is as follows: Equality (4.2.7b) indicates,

that (µa )a∈A0 represents a network flow in (V, A0 ) where each arc a ∈ A0 has µa as
its flow variable. The in- and out-flows at sources and sinks are given by λ+
v − λv ,
−

and the relation of these values with the arc flows is given by the weighted flow
conservation (4.2.7b) (also called generalized flow conservation, see Oldham (1999)
and the references therein). As this flow conservation is weighted, the node flow
+
must not necessarily be balanced, i.e., v∈V (λv − λv ) 6= 0 might hold. The
−
P

implications (4.2.8a) ensure that a nonzero entry flow is only allowed, if πv ≥ π v .

Furthermore, a nonzero exit flow can only occur at a node fulfilling πv ≤ π v . Looking
at equation (4.2.7a), the dual value µv can be interpreted as a dual node potential at
node v. The values λ+
a , λa enforce a dual decrease or increase of the potential values
−

and so react like a dual active element (compressor or control valve) restricted by
(4.2.8b).

Lemma 4.2.12:
Let (q ∗ , π ∗ , ∆∗ ) be an optimal solution of the domain relaxation (4.2.1). There exist
Lagrange multipliers (µ∗ , λ∗ ) which consecutively correspond to the equality and
inequality constraints of (4.2.1), respectively, such that (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) is a KKT
point of (4.2.1).
4.2 Relaxation of Domains 67

Proof. We prove that (µ∗ , λ∗ ) exist such that the conditions (4.2.7) are fulfilled. It
follows from the optimality of (q ∗ , π ∗ , ∆∗ ) (especially of the ∆ variables) that there
exist λ∗v , λ∗v ± , λ∗a , λ∗a ± ∈ R≥0 such that (4.2.7c), (4.2.7d) and the complementarity
conditions (2.4.2e) are fulfilled. As the dual flow conservation (4.2.7b) must be fulfilled
by (µ∗ , λ∗ ) we observe an additional constraint for λ∗ ± . It is obtained as follows:
Summing up the dual flow conservation (4.2.7b) using γr,v from Definition 4.2.3
yields
 
−1 + ∗ ∗
(λv − λ−
v )=
X X X X
−1 
γr,v γr,v µ∗a − γa µ∗a 


+
0 (v) 0 (v)
v∈V v∈V −
a∈δA a∈δA

= µ∗a (γr,v )= µ∗a (γr,v ) = 0.

X X
−1 −1 −1 −1
− γa γr,w − γr,v
a∈A0 a∈A0

We conclude from the connectivity of the graph (V, A0 ) that λ∗v ± , v ∈ V has to be
chosen such that the balancing constraint

−1 ∗ +
(λv − λ∗v − ) = 0 (4.2.9)
X
γr,v
v∈V

is fulfilled in addition to the previously mentioned restrictions. Otherwise there does

not exist a dual arc flow µ fulfilling (4.2.7b). In order to show that (4.2.9) is an
additional feasible restriction for the definition of λ∗v ± , v ∈ V we apply Lemma 4.2.7.
Thereby the solution space of π of the domain relaxation (4.2.1) is a straight line,
i.e., π ∗ + θt is feasible for every t ∈ R. We rewrite the objective function of (4.2.1) by

f (t) := max{0, (πv∗ + θv t) − π v } + max{0, π v − (πv∗ + θv t)} .

X

v

This function f (t) is convex and optimal for t = 0 because of the local optimality of
(q ∗ , π ∗ , ∆∗ ). This implies the following conditions:

∀t0 > 0 : f (t0 ) ≥ f (0) ⇒ (4.2.10a)

X X
θv ≥ θv ,
v∈V :πv∗ ≥π v v∈V :πv∗ <π v

∀t0 < 0 : f (t0 ) ≥ f (0) ⇒ (4.2.10b)

X X
θv ≤ θv .
v∈V :πv∗ >π v v∈V :πv∗ ≤π v

Recall that λ∗v ± is fixed for those nodes v ∈ V with πv > π v or πv < π v by
(4.2.7c) and the complementarity conditions (2.4.2e). For those nodes with πv > π v
we have λ∗v + − λ∗v − = 1 and for the other ones with πv < π v we have λ∗v + − λ∗v − = −1.
Now we distinguish between three cases:
68 Chapter 4 Efficiently Solving the Passive Transmission Problem

Case 1:
θv =
X X
θv
v∈V :πv∗ >π v v∈V :πv∗ <π v

Setting λ∗v + = λ∗v − = 0 for every node v ∈ V with π v ≤ πv ≤ π v yields (4.2.9).

Recall the relation θv := γr,v
−1
, v ∈ V from the proof of Lemma 4.2.7.

Case 2:
X X
θv < θv
v∈V :πv∗ >π v v∈V :πv∗ <π v

We set λ∗v + = λ∗v − = 0 for every node v ∈ V with π v ≤ πv < π v and λ∗v + ∈ [0, 1],
λ∗v − = 0 for every node v ∈ V with πv = π v such that (4.2.9) is fulfilled. This setting
is possible due to (4.2.10b).

Case 3:
X X
θv > θv
v∈V :πv∗ >π v v∈V :πv∗ <π v

We set λ∗v + = λ∗v − = 0 for every node v ∈ V with π v < πv ≤ π v and λ∗v + = 0,
λ∗v − ∈ [0, 1] for every node v ∈ V with πv = π v such that (4.2.9) is fulfilled. This
setting is possible due to (4.2.10b).

We note that exactly one of the above cases applies which means that the previous
discussion yields a feasible definition for λ∗v ± , v ∈ V . Recall that the values of
λ∗v , v ∈ V have to be set according to (4.2.7c).
Now we compute a local optimum (µ∗a )a∈A0 of the problem

XZ µa
min −1
γr,v Φ̃a (t) dt
a∈A0 µ0a

s. t. µa = −λ∗v + + λ∗v −
X X
γa µa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

µa ∈ R ∀ a ∈ A0 ,

∗+
where Φ̃a (µ) := µa dqa (qa ) + λa
dΦa ∗
− λ∗a − and µ0a is a root of this function for each arc
a = (v, w) ∈ A0 . The nonlinear optimization problem is feasible due to the previous
discussion. Similar as in the proof of Lemma 4.2.1 we conclude that there exists
(µ0v )v∈V such that
µ0v − γa µ0w = γr,v
−1
Φ̃a (µ∗a )
4.2 Relaxation of Domains 69

holds for every arc a = (v, w) ∈ A0 . We define µ∗v := γr,v µ0v and make use of (4.2.5)
to obtain
µ∗v − µ∗w = Φ̃a (µ∗a )

for every arc a = (v, w) ∈ A0 . This way we obtain a solution (µ∗ , λ∗ ) which is feasible
for (4.2.7a) and (4.2.7b). Hence (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) is a KKT point of (4.2.1).

Example 4.2.13:
Figure 4.2 shows a visualization of primal and dual flow of a KKT point of the
domain relaxation (4.2.1) in a network of practical dimension (net6). The color of a
node corresponds to the primal / dual node potential. The arc width represents the
primal / dual flow value (the thicker the more flow), while its color depicts the mean
value of the node potentials at both end nodes. The figures show an infeasible primal
flow, where the node potentials at some entries are above their respective upper limit
and some exit node potentials are below their respective lower limit. (Note that there
are other primal entries and exits, too.) Those nodes exceeding their respective upper
limit are dual entries, those which exceed their respective lower limit are dual exits.
For this example the arc set A00 which contains all arcs having a nonzero dual
flow, i.e., A00 = {a ∈ A0 | µa 6= 0}, characterizes the infeasibility of the nomination d
˜ be a solution of the domain relaxation (4.2.1).
in the following sense: Let (q̃, π̃, ∆)
We observe from Figure 4.2 that there exists a subset A000 ⊆ A00 such that from

αa q̃a |q̃a |ka − β̃a − (π̃v − γa π̃w ) = 0

for all arcs a = (v, w) ∈ A000 the existence of nodes s, t ∈ V such that π̃s − π̃t > π s −π t
follows.
70 Chapter 4 Efficiently Solving the Passive Transmission Problem

max min
Figure 4.2: Flow and node potential corresponding to the primal (upper picture) and dual (lower picture)
parts of a KKT point of domain relaxation (4.2.1) for network net6. The line width represents the flow
value (the thicker the more flow), while its color depicts the mean value of the node pressures at both end
nodes. The nodes are depicted by squares and the node colors represent the node pressures. All active
elements which are not closed are in bypass mode and the corresponding arcs are contracted. It follows
from Corollary 4.2.8 that the primal solution flow is unique.
4.3 Relaxation of Flow Conservation Constraints 71

4.3 Relaxation of Flow Conservation Constraints

In this section we consider the flow conservation relaxation of the passive transmission
problem with αa > 0 for each arc a ∈ A0 . This relaxation is obtained by relaxing the
flow conservation constraints (4.1.1d). It is as follows:

∆+ +
X  X 
min v + ∆ −
v + ∆ a + ∆ −
a (4.3.1a)
v∈V a∈A0

∀ a ∈ A0 ,
s. t. αa qa |qa |ka − β̃a − (πv − γa πw ) = 0 (4.3.1b)
a = (v, w),

(qa − (∆+
a − ∆a ))
X
−

+
a∈δA0 (v)

(qa − (∆+ +
a − ∆a )) − (∆v − ∆v ) = dv (4.3.1c)
X
− −
− ∀ v ∈ V,
0 (v)
−
a∈δA

πv ≤ π v ∀ v ∈ V, (4.3.1d)

πv ≥ π v ∀ v ∈ V, (4.3.1e)

qa ≤ q a ∀ a ∈ A0 , (4.3.1f)

qa ≥ q a ∀ a ∈ A0 , (4.3.1g)

∆−
v (π v − πv ) = 0 ∀ v ∈ V, (4.3.1h)

∆+
v (πv − π v ) = 0 ∀ v ∈ V, (4.3.1i)

a (q a − qa ) = 0
∆− ∀ a ∈ A0 , (4.3.1j)

∆+
a (qa − q a ) = 0 ∀ a ∈ A0 , (4.3.1k)

πv ∈ R ∀ v ∈ V, (4.3.1l)

qa ∈ R ∀ a ∈ A0 , (4.3.1m)

∆+
v , ∆v ∈ R≥0 ∀ v ∈ V,
−
(4.3.1n)

∆+
a , ∆a ∈ R≥0 ∀ a ∈ A .
− 0
(4.3.1o)

On an arc a ∈ A0 a positive slack value ∆+

a > 0 or ∆a > 0 is feasible only if the flow
−

variable qa reaches its bounds q a or q a , respectively. Accordingly, a positive slack

value ∆+
v > 0 or ∆v > 0 at a node v ∈ V is feasible only if the potential value πv
−

attains a boundary value π v or π v , respectively.

72 Chapter 4 Efficiently Solving the Passive Transmission Problem

We will show how to solve this nonlinear optimization problem (4.3.1) efficiently.
It can be infeasible, but this infeasibility can be detected in a preprocessing step as
described in Section 4.3.1. Otherwise, if this preprocessing does not detect infeasibility,
then the problem is feasible and convex, see Section 4.3.2 and Section 4.3.3. Hence
it can be solved very efficiently to global optimality. Note that the convexity here
is not given by the constraints but by the feasible solution space of the relaxation.
Finally in Section 4.3.4 we will give an interpretation of the dual solution of a KKT
point of (4.3.1). Similar to the discussion in Section 4.2.3 this dual solution forms a
general network flow that is induced by a dual node potential.

4.3.1 Preprocessing

Problem (4.3.1) can be infeasible. This happens, if the flow bounds (4.3.1f) and
(4.3.1g) enforce such a high amount of flow on an arc, that the potential loss (as
deduced by equation (4.3.1b)) is in conflict with the bounds on the node potentials
on both end nodes of the arc which are given by (4.3.1d) and (4.3.1e). The conflicting
constraints in this case are

αa qa |qa |ka − β̃a = πv − γa πw ,

q a ≤ qa ≤ q a ,

π v ≤ πv ≤ π v ,

π w ≤ πw ≤ π w ,

for arc a = (v, w) ∈ A0 . This conflict situation can easily be detected by solving the
linear program

∃π
s. t. Φa (q a ) ≤ πv − γa πw ≤ Φa (q a ) ∀ a = (v, w) ∈ A0 ,
(4.3.2)
π v ≤ πv ≤ π v ∀ v ∈ V,

πv ∈ R ∀ v ∈ V,

with Φa (q) := αa q|q|ka − β̃a . If this LP turns out to be infeasible, then the passive
transmission problem is infeasible, because the constraints that induce the infea-
sibility are part of the passive transmission problem. Otherwise it follows from
Lemma 4.3.3 (stated and proven in the following section) that the flow conservation
relaxation (4.3.1) is feasible. In the following discussion we assume that we applied
4.3 Relaxation of Flow Conservation Constraints 73

this preprocessing technique which means to solve LP (4.3.2) and did not detect
infeasibility.

4.3.2 Existence of a Solution

We consider the preprocessed problem (4.3.1) where the preprocessing is applied as

described in Section 4.3.1. Recall that (4.2.2) and (4.2.3), which were first presented
by Collins et al. (1978) and Maugis (1977), allowed to compute a feasible solution
for the domain relaxation (4.2.1). We extend these convex optimization problems in
order to compute a feasible solution for the flow conservation relaxation (4.3.1). This
approach only works for constant heights, i.e., γa = 1 for all a ∈ A0 . In the subsequent
part of this section we show how to treat the general case of inhomogeneous heights,
i.e., γa 6= 1 for some a ∈ A0 .
First we describe the extension of (4.2.2). We introduce slack variables and add
further terms to the objective function. Again we set Φa (qa ) := αa qa |qa |ka − β̃a .
Then this extension is of the following form:

XZ qa
min Φa (t) dt
0
qa
a∈A0
(4.3.3a)
π v ∆+ Φa (q a ) ∆+
X  X 
+ π v ∆−
v − v + Φa (q a ) ∆−
a − a
v∈V a∈A0

qa − (∆+
 
s. t. a − ∆a )
X
−

0 (v)
−
a∈δA

qa − (∆+ + (∆+
 
∆ ) v − ∆v ) = −dv (4.3.3b)
X
− −
− a − a ∀ v ∈ V,
+
a∈δA 0 (v)

∆±
v ≥0 ∀ v ∈ V, (4.3.3c)

∆±
a ≥0 ∀ a ∈ A0 , (4.3.3d)

∆±
v ∈R ∀ v ∈ V, (4.3.3e)

q a , ∆±
a ∈R ∀ a ∈ A0 . (4.3.3f)

Here qa0 is the root of Φa (·). In the next lemma we characterize this nonlinear
optimization problem by analyzing the KKT conditions for this constraint system.

Lemma 4.3.1:
The nonlinear optimization problem (4.3.3) is convex. Every optimal solution for
74 Chapter 4 Efficiently Solving the Passive Transmission Problem

(4.3.3) can be transformed into a feasible solution for (4.3.1), if γa = 1 for all arcs
a ∈ A0 .

Proof. The optimization problem (4.3.3) is convex, as Φa is a monotone increasing

function which implies that the objective consists of a sum of convex functions.
Furthermore the constraints are of linear type.
Let (q ∗ , ∆∗ ) be a local minimum of (4.3.3). Because of the convexity every local
minimum is global. The objective and the constraints of (4.3.3) are continuously
differentiable. Hence, by Theorem 2.4.2, there exist dual values (π ∗ , λ∗ ), which
consecutively correspond to the equality and inequality constraints of (4.3.3), such
that (q ∗ , ∆∗ , π ∗ , λ∗ ) is a KKT point of (4.3.3). The Lagrange function with the
Lagrange multipliers (π, λ) = (πv , λ+ +
v , λv , λa , λa )v∈V,a∈A0 with πv ∈ R, v ∈ V and
− −

λ+ − +
v , λv ∈ R≥0 , v ∈ V and λa , λa ∈ R≥0 , a ∈ A is as follows:
− 0

L(q, ∆, π, λ)
X Z qa
+ +
X  X 
= Φa (t) dt + π v ∆−
v − π v ∆ v + Φ a (q a ) ∆−
a − Φ a (q a
) ∆a
0
qa
a∈A0 v∈V a∈A0
 

+ (qa − (∆+
a − ∆a )) + (qa − (∆+
a − ∆a ))
X X X
− − 
πv dv −

+
0 (v) 0 (v)
v∈V −
a∈δA a∈δA
 
+ 
(∆−
v − ∆v )
X X
− πv 
v∈V v∈V

λ+ +
λ+ +
X  X 
v ∆v + λ v ∆v a ∆a + λa ∆a .
− − − −
− −
v∈V a∈A0

We obtain from (2.4.2a) of the KKT conditions (2.4.2) that (q ∗ , ∆∗ , π ∗ , λ∗ ) is feasible

for

∂L
=0⇒ Φa (qa ) = πv − πw ∀ a = (v, w) ∈ A0 , (4.3.4a)
∂qa
∂L
=0⇒ v = πv
π v − λ− ∀ v ∈ V, (4.3.4b)
∂∆−v

∂L
=0⇒ π v + λ+
v = πv ∀ v ∈ V, (4.3.4c)
∂∆+v

∂L
= 0 ⇒ Φa (q a ) − λ−
a = πv − πw ∀ a = (v, w) ∈ A0 , (4.3.4d)
∂∆−a

∂L
=0⇒ Φa (q a ) + λ+
a = πv − πw ∀ a = (v, w) ∈ A0 . (4.3.4e)
∂∆+a
4.3 Relaxation of Flow Conservation Constraints 75

It follows from (4.3.4a) that the vector (q ∗ , π ∗ ) satisfies constraint (4.3.1b). Using
(4.3.4b)–(4.3.4c) in combination with the complementarity conditions (2.4.2e) which
write here as

λ+ +
v ∆v = 0 and λv ∆v = 0,
− −
(4.3.5a)
and
λ+ +
a ∆a = 0 and λa ∆a = 0,
− −
(4.3.5b)

we observe that the following constraints are fulfilled by (∆∗ , π ∗ ):

v (π v − πv ) = 0,
∆−
∆+
v (πv − π v ) = 0.

Using the strict monotonicity of Φa and (4.3.4a), (4.3.4d), (4.3.4e), (4.3.5b), we

obtain that (q ∗ , ∆∗ ) fulfills

∆−
a (Φa (q a ) − Φa (qa )) = 0 ⇒ ∆−
a (q a − qa ) = 0,

∆+
a (Φa (qa ) − Φa (q a )) = 0 ⇒ ∆+
a (qa − q a ) = 0.

So constraints (4.3.1h)–(4.3.1k) are fulfilled by (q ∗ , ∆∗ , π ∗ ). Note that λ ≥ 0. There-

fore (4.3.1d)–(4.3.1g) are also fulfilled because of (4.3.4a)–(4.3.4e). Furthermore
the flow conservation constraints (4.3.1c) are fulfilled due to constraints (4.3.3b).
Altogether, (q ∗ , π ∗ , ∆∗ ) is a feasible solution for (4.3.1), if γa = 1, a ∈ A0 .

Note that problem (4.3.3) might be unbounded. In this case there exists no
optimal solution and that is why we cannot ensure the feasibility of (4.3.1) by
Lemma 4.3.1. Thus we consider another optimization problem which extends (4.2.3)
as follows:
Z πv −πw XZ πv
min Φ−1
a (t) dt dv dt
X
− (4.3.6a)
a=(v,w)∈A0 ∆0a v∈V 0

s. t. πv ≤ π v ∀ v ∈ V, (4.3.6b)

πv ≥ π v ∀ v ∈ V, (4.3.6c)

πv − πw ≤ Φa (q a ) ∀ a = (v, w) ∈ A0 , (4.3.6d)

πv − πw ≥ Φa (q a ) ∀ a = (v, w) ∈ A0 , (4.3.6e)

πv ∈ R ∀ v ∈ V. (4.3.6f)
76 Chapter 4 Efficiently Solving the Passive Transmission Problem

Here ∆0a is the root of the function Φ−1

a (·), the inverse of Φa (·).

Lemma 4.3.2:
The nonlinear optimization problem (4.3.6) is convex and bounded. Its optimal
solution yields a feasible solution for (4.3.1), if γa = 1 for all arcs a ∈ A0 .

Note that problem (4.3.6) is bounded and hence allows to compute a feasible
solution for the flow conservation relaxation (4.3.1) which is not guaranteed for the
previous problem (4.3.3). However, we consider (4.3.3) and (4.3.6) for theoretical
purpose only.

Proof. We note that the constraints of (4.3.6) are of linear type. The objective
function is convex, because of the definition of ∆0a . Hence (4.3.6) is convex.
The problem (4.3.6) might be infeasible because of the linear constraints, but this
situation is excluded because of the preprocessing described in Section 4.3.1. The
linear constraints ensure that the optimization problem is bounded. Hence there
exists a local optimum which we denote by π ∗ . This is a global minimum because of
the convexity.
The objective and the constraints of (4.3.6) are continuously differentiable. Hence,
by Theorem 2.4.2, there exist dual values ∆∗ consecutively corresponding to the
inequality constraints of (4.3.6) such that (π ∗ , ∆∗ ) is a KKT point of (4.3.6). We
denote the Lagrange function with Lagrange multipliers ∆+
v , ∆v ∈ R≥0 for each
−

node v ∈ V and ∆+
a , ∆a ∈ R≥0 for each arc a ∈ A as follows:
− 0

Z πv −πw XZ πv
L(π, ∆) = Φ−1
a (t) dt − dv dt
X

a=(v,w)∈A0 ∆0a v∈V 0

+
+ (∆−
v (πv − π v ) + ∆v (π v − πv ))
X

v∈V

+ ∆−
a (πv − πw − Φa (q a ))
X

a=(v,w)∈A0

+ ∆+
a (Φa (q a ) − (πv − πw )).
X

a=(v,w)∈A0

We obtain from (2.4.2a) of the KKT conditions (2.4.2) that (π ∗ , ∆∗ ) is feasible for

∂L +
=0⇒ (Φ−1
a (πv − πw ) − (∆a − ∆a ))
X
−
∂πv +
a=(v,w)∈δA 0 (v)
4.3 Relaxation of Flow Conservation Constraints 77

+
(Φ−1
a (πv − πw ) − (∆a − ∆a ))
X
−
−
0 (v)
−
a=(v,w)∈δA

−(∆+
v − ∆v ) = dv
−
∀ v ∈ V.

Setting q ∗ by

qa∗ := Φ−1
a (πv − πw )
∗ ∗
⇔ Φa (qa∗ ) = πv∗ − πw
∗
∀ a = (v, w) ∈ A0 (4.3.7)

we derive that (q ∗ , π ∗ , ∆∗ ) fulfills (4.3.1b) and the flow conservation constraints

(4.3.1c).
Constraints (4.3.1d) and (4.3.1e) are fulfilled by (q ∗ , π ∗ , ∆∗ ) because of constraints
(4.3.6b) and (4.3.6c). Constraints (4.3.1f) and (4.3.1g) are fulfilled by (q ∗ , π ∗ , ∆∗ )
because of constraints (4.3.7), (4.3.6d) and (4.3.6e) and the strictly monotonicity
(hence bijectivity) of Φa and Φ−1
a . From the complementary slackness conditions we
observe that the following constraints are fulfilled by (q ∗ , π ∗ , ∆∗ ):

∆−
v (π v − πv ) = 0, ∆+
v (πv − π v ) = 0 ∀ v ∈ V,
∆−
a (q a − qa ) = 0, ∆+
a (qa − q a ) = 0 ∀ a ∈ A0 .

This gives (4.3.1h)–(4.3.1k). Therefore, if γa = 1, a ∈ A0 , then (q ∗ , π ∗ , ∆∗ ) is a

solution for (4.3.1).

Making use of (4.3.6) we are able to prove the following characterization of the
feasibility of the flow conservation relaxation (4.3.1) in the general case that there
exists an arc a ∈ A0 such that γa 6= 1.

Lemma 4.3.3:
If the preprocessing described in Section 4.3.1 does not detect infeasibility, then the
flow conservation relaxation (4.3.1) is feasible.

Proof. If the preprocessing described in Section 4.3.1 detects infeasibility, then the
passive transmission problem is infeasible as discussed in that section. Otherwise, we
proceed as follows. We use γr,v from Definition 4.2.3, the function πv0 (π) from (4.2.4)
and equation (4.2.5). Now we compute a local optimum π ∗ for the following problem
Z πv −πw XZ πv
min Φ0−1
a (t) dt − dv dt
X

a=(v,w)∈A0 ∆0a v∈V 0

s. t. πv ≤ πv0 (π) ∀ v ∈ V,
78 Chapter 4 Efficiently Solving the Passive Transmission Problem

πv ≥ πv0 (π) ∀ v ∈ V,

πv − πw ≤ Φ0a (q a ) ∀ a = (v, w) ∈ A0 ,

πv − πw ≥ Φ0a (q a ) ∀ a = (v, w) ∈ A0 ,

πv ∈ R ∀ v ∈ V,

where Φ0−1
a (·) is the inverse of Φa (qa ) := γr,v αa qa |qa |
0 ka
− γr,v β̃a and ∆0a is a root of
this inverse function for each arc a = (v, w) ∈ A0 .
This problem is feasible because of the assumption that we applied the pre-
processing described in Section 4.3.1. By Lemma 4.3.2 this optimization problem
yields a feasible solution (q ∗ , π ∗ , ∆∗ ) for a modified version of the flow conservation
relaxation (4.3.1) which is obtained by replacing the constraints (4.3.1b), (4.3.1d),
(4.3.1e), (4.3.1h), (4.3.1i) by

γr,v αa qa |qa |ka − γr,v β̃a − (πv − πw ) = 0 ∀ a = (v, w) ∈ A0 ,

πv ≤ πv0 (π) ∀ v ∈ V,

πv ≥ πv0 (π) ∀ v ∈ V,

∆−
v (πv (π) − πv ) = 0
0
∀ v ∈ V,

∆+
v (πv − πv (π)) = 0
0
∀ v ∈ V.

Using (q ∗ , π ∗ , ∆∗ ) we are going to show how to obtain a feasible solution for the
flow conservation relaxation (4.3.1) which is not modified. We define

π̂v := πv0 (π ∗ ).
−1

for each node v ∈ V . We obtain π ≤ π̂ ≤ π. Furthermore we obtain ∆−

v (π v − π̂v ) = 0
and ∆+
v (π̂v − π v ) = 0. In combination with (4.2.5) we obtain

αa qa∗ |qa∗ |ka − β̃a = γr,v

−1
(πv∗ − πw
∗
) = π̂v − γa π̂w ∀ a = (v, w) ∈ A0 .

Thus (q ∗ , π̂, ∆∗ ) is a feasible solution for the flow conservation relaxation (4.3.1).

4.3.3 Characterization of the Feasible Region

The next lemmata characterize the feasible region of the flow conservation relax-
ation (4.3.1). They state that all local optimal solutions differ only in the π values,
which lie together on a straight line segment. This then allows to prove that the flow
4.3 Relaxation of Flow Conservation Constraints 79

conservation relaxation is either infeasible, which can be detected by the preprocessing

described in Section 4.3.1, or a feasible convex optimization problem.

We start with proving a simple result which is needed in the following part of
this section.

Lemma 4.3.4:
Let (q 0 , π 0 ), (q 00 , π 00 ) ∈ RA × RV with q 0 ≤ q 00 be two vectors fulfilling
0

αa qa |qa |ka − β̃a = πv − πw ∀ a = (v, w) ∈ A0

with αa > 0 for all arcs a ∈ A0 . Let P be a v-w-path in (V, A0 ). The following
implications hold for the node potential values in v and w:

πv0 ≥ πv00 ⇒ 0
πw 00
≥ πw ,
0 00
πw ≤ πw ⇒ πv0 ≤ πv00 ,

and

πv0 > πv00 ⇒ 0

πw 00
> πw ,
0 00
πw < πw ⇒ πv0 < πv00 .

If there exists an arc a in this path with qa0 < qa00 then it holds:

πv0 ≥ πv00 ⇒ 0
πw 00
> πw ,
0 00
πw ≤ πw ⇒ πv0 < πv00 .

Proof. Let P be a v-w-path in (V, A0 ). Let the nodes of P be given by v1 , . . . , vn+1

where v1 = v and vn+1 = w and connecting arcs by a1 , . . . , an . We obtain from
qa0 i ≤ qa00i :
πv0 i − γai πv0 i+1 = αai qa0 i |qa0 i |kai − β̃ai
(4.3.8)
≤ αai qa00i |qa00i |kai − β̃ai = πv00i − γai πv00i+1

From this we obtain, again using qa0 i ≤ qa00i for every i = 1, . . . , n:

πv0 1 ≥ πv001 ⇒ πv0 2 ≥ πv002 ⇒ . . . ⇒ πv0 n+1 ≥ πv00n+1 ,

πv0 n+1 ≤ πv00n+1 ⇒ πv0 n ≤ πv00n ⇒ . . . ⇒ πv0 1 ≤ πv001 .
80 Chapter 4 Efficiently Solving the Passive Transmission Problem

If there exists an arc ai0 with qa0 i0 < qa00i0 , then inequality (4.3.8) is strict. We obtain:

πv0 1 ≥ πv001 ⇒ πv0 2 ≥ πv002 ⇒ . . . ⇒πv0 i0 +1 > πv00i0 +1 ⇒ . . . ⇒ πv0 n+1 > πv00n+1 ,
πv0 n+1 ≤ πv00n+1 ⇒ πv0 n ≤ πv00n ⇒ . . . ⇒ πv0 i0 < πv00i0 ⇒ ... ⇒ πv0 1 < πv001 .

Lemma 4.3.5:
There exists a vector q̃ ∈ RA such that the following holds: (q ∗ , π ∗ , ∆∗ ) is a feasible
0

solution of (4.3.1) if and only if q ∗ = q̃.

Proof. Assume that there exist two solutions (q 0 , π 0 , ∆0 ) and (q 00 , π 00 , ∆00 ) of problem
(4.3.1). Note that there exists at least one by Lemma 4.3.3. We are going to prove
that q 0 = q 00 . We proceed as follows:
+
1. At first we show that there exist ∆000 , ∆000 ∈ RA such that the vector
− 0

(q 00 , π 00 , (∆0v , ∆000
a )v∈V,a∈A0 ) is feasible for (4.3.1).
± ±

2. Afterwards we conclude q 0 = q 00 .

In order to prove the first item we have to show the following because of (4.3.1h)
+ +
and (4.3.1i): If there exists a node v ∈ V with one of the values ∆0 v , ∆0 v , ∆00v , ∆00v
− −

not equal to zero, then the node potentials πv0 and πv00 are equal, i.e., πv0 = πv00 .
+ +
Let v ∗ ∈ V be a node such that one of the values ∆0 v∗ , ∆0 v∗ , ∆00v∗ , ∆00v∗ is not
− −

+ +
equal to zero. In the case that ∆0 v∗ − ∆0 v∗ = ∆00v∗ − ∆00v∗ 6= 0 we conclude from
− −

+ +
(4.3.1h) and (4.3.1i) that πv0 ∗ = πv00∗ . In the case ∆0 v∗ − ∆0 v∗ = ∆00v∗ − ∆00v∗ =0
− −

+ +
and some value ∆0 v∗ , ∆0 v∗ , ∆00v∗ , ∆00v∗ 6= 0 we conclude that π v∗ = π v∗ and hence
− −

+ +
πv0 ∗ = πv00∗ . Next we consider the remaining case ∆0 v∗ − ∆0 v∗ 6= ∆00v∗ − ∆00v∗ .
− −

We define q̂a for all arcs a ∈ A0 by

+ +
q̂a := (qa00 − (∆00a − ∆00a )) − (qa0 − (∆0 a − ∆0 a )).
− −

+ +
W.l.o.g. we assume that ∆0 v∗ − ∆0 v∗ < ∆00v∗ − ∆00v∗ (otherwise, exchange the two
− −

solutions). Hence v acts as a source for q̂.

∗

W.l.o.g. we assume q̂ ≥ 0. If this is not the case then we change the orientation of
every arc a ∈ A0 with q̂a < 0 in the same way as described in the proof of Lemma 4.2.6.
+ +
This implies qa00 − (∆00a − ∆00a ) ≥ qa0 − (∆0 a − ∆0 a ) for every arc a ∈ A0 . This in
− −

combination with (4.3.1j) and (4.3.1k) implies

qa00 ≥ qa0 ∀ a ∈ A0 . (4.3.9)

4.3 Relaxation of Flow Conservation Constraints 81

Using Theorem 4.2.5 we split the flow q̂ into a flows along paths P1 , . . . , Pm
and flows along cycles C1 , . . . , Cn . We denote the path flow values by q̂Pi > 0,
i = 1, . . . , m and the cycle flow values by q̂Ci > 0, i = 1, . . . , n. Then

q̂a = q̂Pi +
X X
q̂Ci ∀ a ∈ A0 .
i=1,...,m: i=1,...,n:
a∈A0 (Pi ) a∈A0 (Ci )

Let q̂P` be a path flow that starts at source node v ∗ and ends at some other sink
node w∗ . Then we have

+ + + +
∆0 v∗ − ∆0 v∗ < ∆00v∗ − ∆00v∗ and ∆0 w∗ − ∆0 w∗ > ∆00w∗ − ∆00w∗ . (4.3.10)
− − − −

We analyze the node potential differences in the two end nodes of the path P` , that
is, πv00∗ − πw
00
∗ versus πv ∗ − πw ∗ , and show that πv ∗ = πv ∗ and πw ∗ = πw ∗ . Therefor
0 0 0 00 0 00
+ +
we distinguish three cases. Note that the signs of ∆0 v∗ − ∆0 v∗ and ∆00 w∗ − ∆00 w∗ are
− −

unknown.

+
Case 1, ∆0 v∗ − ∆0 v∗ > 0. From constraint (4.3.1i) for v ∗ it follows that πv0 ∗ = π v∗
−

and (4.3.10) implies πv00∗ = π v∗ . We distinguish three subcases.

+
Case 1.1, ∆0 w∗ − ∆0 w∗ > 0. Then again it follows that πw ∗ = π w∗ . From
− 0

Lemma 4.3.4 we obtain using πv00∗ = πv0 ∗ and (4.3.9) that 00

πw ∗ ≤ 0
πw ∗ holds.
From π w∗ ≤ πw
00
∗ ≤ πw ∗ = π w ∗ we obtain πw ∗ = π w ∗ . Hence πv ∗ = πv ∗ and
0 00 0 00

∗ = πw ∗ . From (4.3.9) we conclude qa = qa for every arc a of path P` because

0 00 00 0
πw
otherwise qa00 > qa0 for an arc of path P` in combination with πv00∗ = πv0 ∗ means
∗ < πw ∗ by Lemma 4.3.4.
00 0
πw
+ +
Case 1.2, ∆0 w∗ − ∆0 w∗ = 0. By (4.3.10) we have ∆00w∗ − ∆00w∗ < 0. According
− −

to constraint (4.3.1h) we obtain that πw

00
∗ = π w ∗ . Now the node potentials π
00

are at the lower boundary on the source and at the upper boundary on the sink
side (πv00∗ = π v∗ , πw
00
∗ = π w ∗ ). From Lemma 4.3.4 we obtain using πv ∗ = πv ∗
0 00

and (4.3.9) that πw

0
∗ ≥ πw ∗ holds. From π w ∗ ≥ πw ∗ ≥ πw ∗ = π w ∗ we obtain
00 0 00

∗ = πw ∗ . Hence πv ∗ = πv ∗ and πw ∗ = πw ∗ . From (4.3.9) we conclude qa = qa

00 0 0 00 0 00 00 0
πw
for every arc a of path P` because otherwise qa00 > qa0 for an arc of path P` in
combination with πv00∗ = πv0 ∗ means πw
00
∗ < πw ∗ by Lemma 4.3.4.
0

+
Case 1.3, ∆0 w∗ − ∆0 w∗ < 0. Then from constraint (4.3.1h) we have πw ∗ = π w∗ ,
− 00

and we are in the same situation as in Case 1.2.

+ +
Case 2, ∆0 v∗ − ∆0 v∗ = 0. From (4.3.1i) in combination with 0 = ∆0 v∗ − ∆0 v∗ <
− −

+
∆00v∗ ∆00v∗ by (4.3.10) we get that πv00∗ = π v∗ .
−
−
82 Chapter 4 Efficiently Solving the Passive Transmission Problem

+
Case 2.1, ∆0 w∗ − ∆0 w∗ > 0. Then again it follows that πw ∗ = π w∗ . From
− 0

Lemma 4.3.4 we obtain using π v∗ = πv00∗ ≤ πv0 ∗ and (4.3.9) that 00

0
≤ πw
πw ∗∗

holds. From π w∗ ≤ 00
πw ∗ ≤ πw∗ = π w∗ we obtain πw∗ = πw∗ . Again we obtain
0 00 0

from Lemma 4.3.4 using πw 0

∗ ≤ πw ∗ and (4.3.9) that πv ∗ ≤ πv ∗ holds. Hence
00 0 00

πv0 ∗ = πv00∗ and πw

0
∗ = πw ∗ . From (4.3.9) we conclude qa = qa for every arc a of
00 00 0

path P` because otherwise qa00 > qa0 for an arc of path P` in combination with
πv00∗ = πv0 ∗ means πw00
∗ < πw ∗ by Lemma 4.3.4.
0

+ +
Case 2.2, ∆0 w∗ − ∆0 w∗ = 0. By (4.3.10) we have ∆00w∗ − ∆00w∗ < 0. According
− −

to constraint (4.3.1h) we obtain that πw

00
∗ = π w ∗ . Now the node potentials π
00
are
at the lower boundary on the source and at the upper boundary on the sink side
(πv00∗ = π v∗ , πw
00
∗ = π w ∗ ). From Lemma 4.3.4 we obtain using πv ∗ ≥ πv ∗ = π v ∗
0 00

and (4.3.9) that πw

0
∗ ≥ πw ∗ holds. From π w ∗ ≥ πw ∗ ≥ πw ∗ = π w ∗ we conclude
00 0 00

∗ = πw ∗ . Again we obtain from Lemma 4.3.4 using πw ∗ ≤ πw ∗ and (4.3.9) that

00 0 0 00
πw
πv0 ∗ ≤ πv00∗ holds. Hence πv0 ∗ = πv00∗ and πw
0
∗ = πw ∗ . From (4.3.9) we conclude
00

qa00 = qa0 for every arc a of path P` because otherwise qa00 > qa0 for an arc of path
P` in combination with πv00∗ = πv0 ∗ means πw
00
∗ < πw ∗ by Lemma 4.3.4.
0

+
Case 2.3, ∆0 w∗ − ∆0 w∗ < 0. Then from constraint (4.3.1h) we have πw ∗ = π w∗ ,
− 00

and we are in the same situation as in Case 2.2.

+
Case 3, ∆0 v∗ − ∆0 v∗ < 0. Then it follows from (4.3.1h) that πv0 ∗ = π v∗ .
−

+
Case 3.1, ∆0 w∗ − ∆0 w∗ > 0. Then it follows from (4.3.1i) that πw ∗ = π w ∗ . Now
− 0

πv∗ = π v and πw∗ = π w∗ holds. From Lemma 4.3.4 we obtain using πv00∗ ≤
0 ∗
0

π v∗ = πv0 ∗ and (4.3.9) that πw

0
∗ ≥ πw ∗ holds. From π w ∗ = πw ∗ ≥ πw ∗ ≥ π w ∗
00 0 00

we conclude πw
00
∗ = πw ∗ . Again we obtain from Lemma 4.3.4 using πw ∗ ≤ πw ∗
0 0 00

and (4.3.9) that πv0 ∗ ≤ πv00∗ holds. Hence πv0 ∗ = πv00∗ and πw
0
∗ = πw ∗ . Again
00

from (4.3.9) we conclude qa00 = qa0 for every arc a of path P` because otherwise
qa00 > qa0 for an arc of path P` in combination with πv00∗ = πv0 ∗ means πw
00 0
∗ < πw ∗

by Lemma 4.3.4.
+ + +
Case 3.2, ∆0 w∗ − ∆0 w∗ = 0. We derive 0 = ∆0w∗ − ∆0w∗ > ∆00w∗ − ∆00w∗
− − −

from (4.3.10). In combination with (4.3.1i) we obtain πw

00
∗ = π w ∗ . Furthermore

we have πw∗ ≤ π w∗ = πw∗ . From Lemma 4.3.4 and (4.3.9) we conclude πv00∗ ≥ πv0 ∗ .
0 00

From π v∗ ≥ πv00∗ ≥ πv0 ∗ = π v∗ we obtain πv0 ∗ = πv00∗ . Again from Lemma 4.3.4
and (4.3.9) we conclude πw
0
∗ ≥ πw ∗ which then in combination with the previous
00

observation πw
0
∗ ≤ πw ∗ implies πw ∗ = πw ∗ . We derive qa = qa for every arc a of
00 0 00 00 0

path P` because otherwise qa00 > qa0 for an arc of path P` in combination with
πv00∗ = πv0 ∗ means πw
00
∗ < πw ∗ by Lemma 4.3.4.
0
4.3 Relaxation of Flow Conservation Constraints 83

+
Case 3.3, ∆0 w∗ − ∆0 w∗ < 0. Then we get from (4.3.1h) that πw ∗ = π w ∗ . So we
− 00

are in the same situation as in Case 3.2, hence the same conclusions remain valid.

From these cases we conclude that πv0 ∗ = πv00∗ for any node v ∗ ∈ V such that at least
+ +
one value ∆0 v∗ , ∆0 v∗ , ∆00v∗ , ∆00v∗ is not equal to zero. This implies that (πv00 , ∆0 v )v∈V
− − ±

is feasible for constraints (4.3.1h) and (4.3.1i).

In order to complete the proof of the first item we turn to the definition of
+ +
(∆000
a )a∈A0 . The flow q̂ is a network flow with node flow (∆00v − ∆00v ) − (∆0 v − ∆0 v )
± − −

for every node v ∈ V . Hence we obtain from the previous 3 × 3 cases that qa00 = qa0 for
every arc a ∈ A0 (Pi ), i = 1, . . . , m. This implies for every arc a ∈ A0 (Pi ), i = 1, . . . , m:

+ +
(∆00a − ∆00a ) − (∆0 a − ∆0 a ) = q̂a = q̂Pi + q̂Ci > 0. (4.3.11)
− −
X X
i=1,...,m: i=1,...,n:
a∈A0 (Pi ) a∈A0 (Ci )

We define
+ +
(∆000 − ∆000
a ) := (∆a − ∆a ) − (4.3.12)
− −
X
00 00
a q̂Pi
i=1,...,m:
a∈A0 (Pi )

+
for every arc a ∈ A0 where at least one value ∆000 or ∆000 equals zero. Now we
−
a a
consider an arc a ∈ A (Pi ), i = 1, . . . , m. From ∆000 > 0 it follows by q̂Pi > 0,
0 −
a
+
i = 1, . . . , m and (4.3.12) that ∆00a > 0. This implies qa00 = q a . From ∆000 > 0 it
−
a
+
follows from (4.3.11) that ∆0a > 0 holds. This implies qa00 = qa0 = q a . For all other
arcs a ∈ A0 which are not considered before it holds that they are not part of any
+ +
path Pi , i = 1 . . . , m. Hence ∆000 − ∆000 = ∆00a − ∆00a holds for them and (4.3.1j)
− −
a a
and (4.3.1k) are fulfilled. We conclude that (qa00 , ∆000
a )a∈A0 is feasible for constraints
±

(4.3.1j) and (4.3.1k).

It follows from (4.3.11) and (4.3.12) that the flow conservation constraint (4.3.1c)
is fulfilled by (q 00 , (∆0v , ∆000
a )v∈V,a∈A0 ). Hence a feasible vector for (4.3.1) is given
± ±

by (q 00 , π 00 , (∆0v , ∆000
a )v∈V,a∈A0 ). This proves the first item.
± ±

+ +
Turning to the second item we define q̂a0 := (qa00 −(∆000
a −∆a ))−(qa −(∆ a −∆ a ))
000 0 0 0 − −

for each arc a ∈ A0 . Then q̂ 0 consists of circulations only because of

+ +
q̂a0 = dv + (∆0v − ∆0v ) − (dv + (∆0v − ∆0v )) = 0 ∀ v ∈ V.
− −
X X
q̂a0 −
+
0 (v) 0 (v)
−
a∈δA a∈δA

But a circulation can only take place in the (∆±

a )a∈A0 variables because of the
following reason: By (4.3.9) it holds q 00 ≥ q 0 . If there exists an arc a ∈ A0 with
qa00 > qa0 , then, using αa > 0, a ∈ A0 and (4.3.1b), we derive the contradiction 0 > 0
similar as done in the end of the proof of Lemma 4.2.7. Thus we conclude q 0 = q 00 ,
which proves the Lemma.
84 Chapter 4 Efficiently Solving the Passive Transmission Problem

Theorem 4.3.6:
The flow conservation relaxation (4.3.1) is a relaxation of the passive transmission
problem (4.1.1). The relaxation is either infeasible, which can be detected by the
preprocessing technique described in Section 4.3.1, or a feasible convex optimization
problem.

Proof. Every feasible solution (q ∗ , π ∗ , p∗ ) of the passive transmission problem (4.1.1)

yields a feasible solution (q ∗ , π ∗ , 0) for (4.3.1). Hence (4.3.1) is a relaxation of the
passive transmission problem (4.1.1).
If infeasibility is not detected during the preprocessing described in Section 4.3.1,
then (4.3.1) is feasible by Lemma 4.3.3. In this case we consider two feasible solutions
(q 0 , π 0 , ∆0 ) and (q 00 , π 00 , ∆00 ) of (4.3.1). From Lemma 4.3.5 we know that q 0 and q 00
are equal, i.e., q 0 = q 00 . By the same argumentation as in the last part of the proof of
Lemma 4.2.7 we obtain the existence of θ ∈ RV≥0 such that there exists t ∈ R such
that π 00 = π 0 + tθ. The definition of θ is independent of the two solutions. Now we
distinguish two cases:

1. If there exist two nodes v and w with πv0 = π v and πw

0
= π w then θ ≥ 0 implies
either t = 0 or θ = 0 and hence π 0 = π 00 . This means that the two solutions
(q 0 , π 0 , ∆0 ) and (q 00 , π 00 , ∆00 ) differ only in the variables ∆ = (∆±
a , ∆v )a∈A0 ,v∈V .
±

Fixing the node potential variables π = π 0 and the flow variables q = q 0 in

(4.3.1) results in a linear program. This means that (q 0 , π 0 , ∆0 + (1 − )∆00 ) is
a feasible solution for (4.3.1) for every  ∈ [0, 1]. We conclude that the feasible
solution space of (4.3.1) is convex in this case.

2. It holds either πv0 > π v for every node v ∈ V or πv0 < π v for every node
+
v ∈ V . This means either ∆0v = 0 for all nodes v ∈ V or ∆0v = 0 for all
−

nodes v ∈ V . Summing (4.3.1c) over all nodes v ∈ V yields the condition

0+
v∈V (∆v − ∆v ) = 0. We conclude ∆v = 0 for all nodes v ∈ V . From
0− 0±
P

θ ≥ 0 we obtain that there does not exist a solution (q 0 , π 000 , ∆000 ) of (4.3.1) with
πv000 = π v and πw
000
= π w for two nodes v, w ∈ V . Hence the previous discussion
also applies for π 00 and ∆00v , v ∈ V and we obtain ∆00v = 0, v ∈ V . Fixing the
± ±

node slack variables ∆v ±

= 0 and the flow variables q = q 0 in (4.3.1) results in
a linear program. This means that (q 0 , 1 π 0 + (1 − 1 )π 00 , 2 ∆0 + (1 − 2 )∆00 )
is a feasible solution for (4.3.1) for every 1 , 2 ∈ [0, 1]. We conclude that the
feasible solution space of (4.3.1) is convex in this case.

Exactly one of the previous cases applies. This proves that the feasible solution space
of (4.3.1) is convex and hence (4.3.1) is a convex optimization problem.
4.3 Relaxation of Flow Conservation Constraints 85

4.3.4 Interpretation of Lagrange Multipliers

Assume that the passive transmission problem (4.1.1) is infeasible and the flow
conservation relaxation (4.3.1) has a positive optimal objective value. It turns out
(see Lemma 4.3.7) that there exist Lagrange multipliers for the optimal solution such
that the optimal solution and the multipliers form a KKT point of (4.3.1). These
multipliers have a practical interpretation. They form a generalized network flow
which is coupled with node potentials, similar to a primal solution (q ∗ , π ∗ ) of the
passive transmission problem. This result is comparable to the interpretation in
Section 4.2.3.

Lemma 4.3.7:
Let (q ∗ , π ∗ , ∆∗ ) be an optimal solution of the flow conservation relaxation (4.3.1).
Let (µ∗ , λ∗ ) be Lagrange multipliers which consecutively correspond to the equality
and inequality constraints of (4.3.1), respectively, such that (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) is a
KKT point of (4.3.1). These multipliers are characterized as follows: (µ∗a )a∈A0 is
a general network flow in (V, A0 ) which is induced by dual node potentials (µ∗v )v∈V .
More precisely the multipliers (µ∗ , λ∗ ) are a feasible solution for

dΦa ∗
µa (q ) + λ+ + −∗ − +∗
a − λa = µv − µw + µa ∆a − µa ∆a
−
∀ a = (v, w) ∈ A0 ,
dqa a

γa µa = λ+ + −∗ − +∗
v − λv + µv ∆v − µv ∆v
X X
−
µa − ∀ v ∈ V.
+
0 (v) 0 (v)
−
a∈δA a∈δA

Hereby the dual node potential µv is restricted by

µv ∈ {1} [−1, 1] {−1}

∆− >0 ∆− =0= ∆+ ∆+
v >0
∗ ∗ ∗ ∗
node slack v v v

for each node v ∈ V , and the dual node potential difference µv − µw is constrained by

µv − µw ∈ {1} [−1, 1] {−1}

∆− >0 ∆− =0= ∆+ ∆+
a >0
∗ ∗ ∗ ∗
arc slack a a a

for each arc a = (v, w) ∈ A0 .

Proof. Let (q ∗ , π ∗ , ∆∗ ) be an optimal solution of (4.3.1). Recall that a local

optimal solution is globally optimal because of the convexity of (4.3.1) by The-
orem 4.3.6. Let (µ∗ , λ∗ ) be dual values such that (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) is a KKT
point of flow conservation relaxation (4.3.1). Let us write the conditions (2.4.2a)
of the KKT conditions (2.4.2). We denote the Lagrange multipliers by (µ, λ) =
86 Chapter 4 Efficiently Solving the Passive Transmission Problem

(µv , µ+ + + + + +
v , µv , µa , µa , µa , λv , λv , λ̃v , λ̃v , λa , λa , λ̃a , λ̃a )v∈V,a∈A0 , such that for the do-
− − − − − −

mains µv , µ+ + + + + +
v , µv , µa , µa , µa ∈ R and λv , λv , λ̃v , λ̃v , λa , λa , λ̃a , λ̃a ∈ R≥0 holds.
− − − − − −

Furthermore we define the function Φa (qa ) := αa qa |qa |ka − β̃a . Then the Lagrange
function of problem (4.3.1) has the form

L(q, π, ∆, µ, λ)

∆+ +
X  X 
= v + ∆ −
v + ∆a + ∆ −
a
v∈V a∈A0

+ µa Φa (qa ) − (πv − γa πw )
X

a=(v,w)∈A0
 

+ (qa − ∆+
a + ∆a ) + (qa − ∆+
a + ∆ a )
X X X
− − 
µv dv −

+
0 (v) 0 (v)
v∈V −
a∈δA a∈δA

+ µv (∆+
v − ∆v )
X
−

v∈V

λ+
X 
+ v (πv − π v ) + λv (π v − πv )
−

v∈V

λ+
X 
+ a (q a − q a ) + λ−
a (q a
− qa )
a∈A0

µ+ +
X 
+ v (π v − πv ) ∆v + µv (πv − π v ) ∆v
− −

v∈V

µ+ +
X 
+ a (q a − qa ) ∆a + µa (qa − q a ) ∆a
− −

a∈A0

λ̃+ +
X 
v ∆v + λ̃v ∆v
− −
−
v∈V

λ̃+ +
X 
− a ∆ a + λ̃a ∆
− −
a .
a∈A0

From (2.4.2a) we obtain that the KKT point (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) is feasible for

∂L
=0⇒ −µ+ − +
a ∆ a + µa ∆ a
−
∂qa
dΦa ∀ a ∈ A0 ,
+µa (qa ) + λ+
a − λa = µv − µw
−
(4.3.13a)
dqa a = (v, w),
∂L
=0⇒ µ+ − +
v ∆ v − µv ∆ v
−
∂πv
+ γ a µa = λ + (4.3.13b)
X X −
µa − v − λv ∀ v ∈ V,
a∈δ +0 (v) a∈δ −0 (v)
A A

∂L +
=0⇒ µv + µ−
v (πv − π v ) − λ̃v = −1 ∀ v ∈ V, (4.3.13c)
∂∆+v
4.3 Relaxation of Flow Conservation Constraints 87

∂L
=0⇒ −µv + µ+
v (π v − πv ) − λ̃v = −1
−
∀ v ∈ V, (4.3.13d)
∂∆−v

∂L + ∀ a ∈ A0 ,
a (qa − q a ) − λ̃a = −1
= 0 ⇒ µ v − µw + µ− (4.3.13e)
∂∆+a a = (v, w),
∂L ∀ a ∈ A0 ,
= 0 ⇒ µ w − µv + µ+
a (q a − qa ) − λ̃a = −1
−
(4.3.13f)
∂∆−a a = (v, w).

We conclude from (4.3.13c)-(4.3.13f), and the complementarity condition (2.4.2e)

that (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) fulfills:

∆−
v > 0 ⇒ µv = 1, ∆+
v > 0 ⇒ µv = −1, ∀ v ∈ V, (4.3.14a)

+ ∀ a ∈ A0 ,
∆−
a > 0 ⇒ µv − µw = 1, ∆a > 0 ⇒ µv − µw = −1, (4.3.14b)
a = (v, w).

The equations (4.3.13a) and (4.3.13b) yield the constraints of the Lemma, while
the tables follow from (4.3.14a) and (4.3.14b). From the complementarity con-
straints (4.3.1h) and (4.3.1i) we obtain

πv < π v ⇒ λ + + −
v = 0, µv ∆v = 0,
∀ v ∈ V, (4.3.15a)
− +
v = 0, µv ∆v = 0,
πv > π v ⇒ λ −

qa < q a ⇒ λ+ + −
a = 0, µa ∆a = 0,
∀ a ∈ A0 . (4.3.15b)
− +
a = 0, µa ∆a = 0,
qa > q a ⇒ λ−

We compare (4.3.13) with (4.2.7), the corresponding results for the domain
relaxation (4.2.1). Most of the interpretation of (4.2.7) remains also valid for (4.3.13).
In the following we will focus on the differences. Conditions (4.3.13a) correspond to
the potential flow coupling (4.2.7a), and conditions (4.3.13b) correspond to the flow
conservation (4.2.7b). The derived conditions (4.2.8a) state that in certain cases for
the node potential values we have to fix the dual node flows. Complementary, the
derived conditions (4.3.14a) state that in certain cases for the node flow slack values
we have to fix the dual node potentials. Similarly, conditions (4.2.8b) state that in
certain cases for the arc flow values we have to fix the dual variables λ+
a and λa .
−

Complementary, the derived conditions (4.3.14b) state that in certain cases for the
arc flow slack values we have to fix the dual node potential difference at the end
nodes of the arc.
88 Chapter 4 Efficiently Solving the Passive Transmission Problem

4.4 Relaxation of Potential-Flow-Coupling Constraints

A relaxation of the passive transmission problem (4.1.1) which is obtained by relaxing
the potential-flow-coupling constraint (4.1.1b) writes as

min (∆+
a + ∆a ) (4.4.1a)
X
−

a∈A0

∀ a ∈ A0 ,
s. t. αa qa |qa |ka − (πv − γa πw ) − (∆+
a − ∆a ) = β̃a
−
(4.4.1b)
a = (v, w),

qa = dv (4.4.1c)
X X
qa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

πv − π v ≤ 0 ∀ v ∈ V, (4.4.1d)

π v − πv ≤ 0 ∀ v ∈ V, (4.4.1e)

qa − q a ≤ 0 ∀ a ∈ A0 , (4.4.1f)

q a − qa ≤ 0 ∀ a ∈ A0 , (4.4.1g)

πv ∈ R ∀ v ∈ V, (4.4.1h)

qa ∈ R ∀ a ∈ A0 , (4.4.1i)

∆+
a , ∆a ∈ R≥0
−
∀ a ∈ A0 . (4.4.1j)

Lemma 4.4.1:
The optimization problem (4.4.1) is a relaxation of the passive transmission prob-
lem (4.1.1).

Proof. A solution (q ∗ , π ∗ , p∗ ) is feasible for the passive transmission problem (4.1.1)

only if (q ∗ , π ∗ , 0) is feasible for the nonlinear optimization problem (4.4.1). Hence
(4.4.1) is a relaxation of (4.1.1).

In the following we will show that this relaxation (4.4.1) is a non-convex opti-
mization problem having different KKT points with different objective values.

4.4.1 Conditions of the KKT System

In order to show that there exist multiple KKT points of (4.4.1) with different objec-
tive values we proceed as follows: The objective and all constraints of (4.4.1) are con-
tinuously differentiable. Now let us consider a KKT point (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ). Further
let us write the conditions (2.4.2a) of the KKT system (2.4.2) for (4.4.1). We denote
4.4 Relaxation of Potential-Flow-Coupling Constraints 89

the Lagrange multipliers by (µ, λ) = (µv , µa , λ+ + +

v , λv , λa , λa , λ̃a , λ̃a )v∈V,a∈A0 , such
− − −

that µv , µa ∈ R and λ+ − + − +
v , λv , λa , λa , λ̃a , λ̃a ∈ R≥0 . We set Φa (qa ) := αa qa |qa | − β̃a .
− ka

Then the Lagrange function of problem (4.4.1) has the form

∆+
X 
L(q, π, ∆, µ, λ) = a + ∆a
−

a∈A0

µa Φa (qa ) − (πv − γa πw ) − (∆+

 
+ a − ∆a )
X
−

a=(v,w)∈A0
 

+ qa +
X X X
µv dv − qa 
 
+
0 (v) 0 (v)
v∈V −
a∈δA a∈δA

λ+
X 
+ v (πv − π v ) + λv (π v − πv )
−

v∈V

λ+
X 
+ a (q a − q a ) + λ −
a (q a
− q a )
a∈A0

λ̃+ +
a ∆a + λ̃a ∆a .
X
− −
−
a∈A0

Now (2.4.2a) which are fulfilled by the KKT point (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) write as

∂L dΦa ∗ ∀ a ∈ A0 ,
=0⇒ µa (q ) + λ+
a − λa = µv − µw
−
(4.4.2a)
∂qa dqa a a = (v, w),
∂L
=0⇒ γa µa = λ+ (4.4.2b)
X X
−
µa − v − λv ∀ v ∈ V,
∂πv +
0 (v) 0 (v)
−
a∈δA a∈δA

∂L
=0⇒ µa + λ̃+
a =1 ∀ a ∈ A0 , (4.4.2c)
∂∆+a

∂L
=0⇒ a = −1
µa − λ̃− ∀ a ∈ A0 . (4.4.2d)
∂∆−a

In combination with the complementarity conditions (2.4.2e) the subsequent condi-

tions follow for (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ):

qa < q a ⇒ λ+
a = 0, a =0
qa > q a ⇒ λ − ∀ a ∈ A0 , (4.4.3a)

πv < π v ⇒ λ+
v = 0, π v = π v ⇒ λ+
v ≥0 ∀ v ∈ V, (4.4.3b)

v = 0,
πv > π v ⇒ λ− πv = π v ⇒ λ −
v ≤0 ∀ v ∈ V, (4.4.3c)

∆+
a > 0 ⇒ µa = 1, ∆−
a > 0 ⇒ µa = −1 ∀ a ∈ A0 . (4.4.3d)
90 Chapter 4 Efficiently Solving the Passive Transmission Problem

α1 = 1.0, γ1 = 1.0

dv = 1, πv ∈ [0, 10] dw = −1, πw ∈ [20, 30]

α2 = 1.5, γ2 = 1.0

Figure 4.3: Example of an instance (network and nomination) demonstrating the nonconvexity of the
relaxation (4.4.1) of the passive transmission problem (4.1.1). Two different KKT points of (4.4.1) having
different objective values are presented in Example 4.4.2.

We compare (4.4.3) with (4.3.14) and (4.2.8). Recall that in the domain relaxation,
(4.2.7), we had an enforcement of the variables λ±v , λa . In the flow conservation
±

relaxation, (4.3.13), we derived an enforcement of the variables µv . Now in the

potential-flow-coupling relaxation, (4.4.2), only an enforcement of the variables µa
to nonzero values remains.

4.4.2 Different KKT Points

The following example shows that the feasible domain of (4.4.1) is non-convex in
general, as there exist two different KKT points with different optimal objective
values for a test instance being a planar graph. The convex combination of the primal
parts is not feasible. Hence the feasible solution space is non-convex and a nonlinear
solver, like IPOpt, which computes KKT points, cannot guarantee to compute the
global optimal solution of (4.4.1).

Example 4.4.2:
Consider the network shown in Figure 4.3. It consists of two nodes v, w and two
parallel arcs a1 , a2 from v to w. We set Φa (qa ) := αa qa |qa | and assume arc constants
α1 = 1.0, β1 = 0, γ1 = 1.0 and α2 = 1.5, β2 = 0, γ2 = 1.0, and node potential bounds
π v = 0, π v = 10, π w = 20, π w = 30. Node v is an entry with flow +1, node w is an
exit with flow −1. The following two solutions both fulfill the KKT system (4.4.2):

• Let q1 = 3, q2 = −2, πv = 10, πw = 20, ∆+ +

1 = 19, ∆2 = 4, ∆1 = 0, ∆2 = 0.
− −

The objective function value is 23. The dual values are µ1 = 1, µ2 = 1, µv = 6,

µw = 0.

• Let q1 = 0.6, q2 = 0.4, πv = 10, πw = 20, ∆+ +

1 = 10.36, ∆2 = 10.24, ∆1 = 0,
−

∆−
2 = 0. The objective function value is 20.6. The dual values are µ1 = 1,
µ2 = 1, µv = 1.2, µw = 0.

Thus we found two different KKT points. It is easy to see that a convex combination
of both primal feasible solutions is not feasible for the relaxation (4.4.1) because of
4.5 Solving the Passive Transmission Problem 91

constraint (4.4.1b). Hence the feasible solution space of (4.4.1) is not convex. As
both solutions have different objective values this shows that even a nonlinear solver,
like IPOpt, which computes KKT points, cannot guarantee to compute the global
optimal solution of (4.4.1).

4.5 Solving the Passive Transmission Problem

The convex relaxations (4.2.1) and (4.3.1) of the passive transmission problem (4.1.1)
described in the previous Sections 4.2 and 4.3 neglect constraints (4.1.1c) and (4.1.1e).
Hence they can be used to compute a solution for the passive transmission problem
efficiently only if Aa = 0, ba = 0 for each arc a ∈ A0 . Furthermore it is required that
there exists no arc a = (v, w) ∈ A0 that is modeled by the constraint πv = πw , which
implies αa > 0 for all arcs a ∈ A0 , see Section 3.2.
In this section we turn to the general case and describe how the relaxations
can be used to compute a solution for the passive transmission problem without
these restrictions. First we consider the case that the passive transmission problem
contains an arc (v, w) which is modeled by πv = πw . Our strategy for solving this
problem splits up into three steps as follows:

Step 1: First we apply a preprocessing technique: For each arc a = (v, w) ∈ A0

where constraint (4.1.1b) writes as πv = πw , we contract arc a and identify the end
nodes v and w. This identification goes along with setting node potential bounds
[π v , π v ] ∩ [π w , π w ] for the node which represents the contraction of v and w. If this
intersection is empty, then the passive transmission problem (4.1.1) is infeasible
because constraint (4.1.1b) is in contradiction with the node potential bounds at v
and w. If we do not detect infeasibility here, then we continue with the next step.

Step 2: From the contraction we obtain the node set V 0 and the arc set A00 . Now
we solve the corresponding preprocessed passive transmission problem by making
use of either the domain relaxation (4.2.1) presented in Section 4.2 or the flow
conservation relaxation (4.3.1) described in Section 4.3. A solution of one of these
relaxations with optimal objective zero yields a feasible solution for the preprocessed
passive transmission problem. Both problems are convex by Theorem 4.2.9 and
Theorem 4.3.6 and so can be solved efficiently. The relaxation also means that an
optimal solution with positive objective value implies that the preprocessed passive
transmission problem is infeasible.
If the preprocessed problem turns out to be infeasible, then again, the passive
transmission problem (4.1.1) is infeasible. This is proven as follows: For every
92 Chapter 4 Efficiently Solving the Passive Transmission Problem

solution (q 0 , π 0 ) of the passive transmission problem it holds that (qa0 )a∈A00 is a feasible
flow for the preprocessed passive transmission problem. Hence the preprocessed
passive transmission problem cannot be infeasible if the non-preprocessed passive
transmission problem is feasible. If we do not detect infeasibility here, then, similar
to step 1, we continue with the next step.

Step 3: For an optimal solution (q ∗ , π ∗ ) of the preprocessed passive transmission

problem, we solve the non-preprocessed passive transmission problem and fix the flow
on all arcs a ∈ A00 to qa∗ before. We define Φa (qa ) := αa qa |qa |ka − β̃a . The arising
problem is a linear program and writes as follows:

∃ q, π (4.5.1)

s. t. πv − γa πw = Φa (qa∗ ) ∀ a = (v, w) ∈ A00 ,

πv − πw = 0 ∀ a = (v, w) ∈ A0 \ A00 ,
qa = qa∗ ∀ a = (v, w) ∈ A00 ,
qa = dv
X X
qa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

π v ≤ πv ≤ π v ∀ v ∈ V,
q a ≤ qa ≤ q a ∀ a ∈ A0 ,
πv ∈ R ∀ v ∈ V,
qa ∈ R ∀ a ∈ A0 .

For every solution flow q 0 of the passive transmission problem it holds that qa0 = qa∗ for
every arc a ∈ A00 . This can be seen as follows: A solution q 0 of the passive transmission
problem yields a flow vector (qa0 )a∈A00 which is feasible for the preprocessed passive
transmission problem and hence feasible for its domain relaxation. By Lemma 4.2.7
this flow vector is unique. As q ∗ is also a feasible flow vector for the domain relaxation
we obtain qa0 = qa∗ for every arc a ∈ A00 . We conclude that LP (4.5.1) is feasible if and
only if the passive transmission problem is feasible. Hence (4.5.1) is either feasible
and the optimal solution (q̃, π̃) yields a feasible solution (q̃, π̃, p̃) with p̃ := sgn(π) |π|
p

for the passive transmission problem, or otherwise it is infeasible, which implies that
the passive transmission problem is infeasible, too.

The discussion above describes how to solve the passive transmission problem,
if it contains arcs a = (v, w) ∈ A0 that are modeled by πv = πw . Now we turn to
4.5 Solving the Passive Transmission Problem 93

the more general case Aa , ba 6= 0 for an arc a ∈ A0 . In order to solve the passive
transmission problem in this case, we proceed as described in the following two steps:

Step 1: We ignore the constraints Aa (qa , pv , pw ) ≤ ba for each arc a ∈ A0 and

T

solve the arising passive transmission problem as described above. If we detect

infeasibility, then the passive transmission problem (4.1.1) is infeasible because
neglecting constraints (4.1.1c) means to consider a relaxation of (4.1.1). Otherwise,
if we obtain a feasible solution (q ∗ , π ∗ , p∗ ), then it is not guaranteed that (q ∗ , π ∗ , p∗ )
is a feasible solution for the passive transmission problem (4.1.1). In this case we
proceed with the next step.

Step 2: From the previous discussion in Step 3 we know that the arc flow qa
has a unique solution for every arc a = (v, w) ∈ A0 which is not modeled by
πv = πw . Consequently, for a feasible solution (q ∗ , π ∗ , p∗ ) of the passive transmission
problem without constraints (4.1.1c) and a feasible solution (q 0 , π 0 , p0 ) of the passive
transmission problem (including constraints (4.1.1c)) it holds qa0 = qa∗ for every arc
a = (v, w) ∈ A0 not modeled by πv = πw .
Let us now analyze the restrictions induced by constraints (4.1.1c) on the solution
space of the passive transmission problem without constraints (4.1.1c). Recall that
(4.1.1c) writes as 0 ≤ 0 for all arcs a = (v, w) ∈ A0 which are modeled by πv = πw
as Aa = 0 and ba = 0, see Section 3.2. For all other arcs a ∈ A0 the previous
argumentation implies that the feasible flow values are given by qa∗ . Let us now
consider an arc a = (v, w) ∈ A0 and analyze the constraints

Φ(qa∗ ) = πv − γa πw , Aa (qa∗ , pv , pw ) ≤ 0,
T
(4.5.2)
pv |pv | = πv , pw |pw | = πw .

Below we are going to reformulate this as

(i) (i)
Φ(qa∗ ) = πv − γa πw , (πv , πw ) ∈ (4.5.3)
[
Ia,v × Ia,w .
i

(i) (i)
where Ia,v , i = 1, . . . and Ia,w , i = 1, . . . are disjoint intervals, respectively. Then, after
fixing the flows of all arcs a ∈ A00 to qa∗ and neglecting the pressure variables pv , v ∈ V ,
we solve the equivalent formulation of the passive transmission problem (4.1.1) which
writes as
∃ q, π (4.5.4)

s. t. πv − γa πw = Φa (qa∗ ) ∀ a = (v, w) ∈ A00 ,

πv − πw = 0 ∀ a = (v, w) ∈ A0 \ A00 ,
94 Chapter 4 Efficiently Solving the Passive Transmission Problem

qa = qa∗ ∀ a = (v, w) ∈ A00 ,

qa = dv
X X
qa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

(i) (i)
(πv , πw ) ∈ Ia,v ∀ a = (v, w) ∈ A0 ,
_
× Ia,w
i

π v ≤ πv ≤ π v ∀ v ∈ V,
q a ≤ qa ≤ q a ∀ a ∈ A0 ,
πv ∈ R ∀ v ∈ V,
qa ∈ R ∀ a ∈ A0 .

This problem is a disjunctive programming problem and its optimal solution (q̃, π̃)
yields a feasible solution (q̃, π̃, p̃) with p̃ := sgn(π) |π| for the passive transmission
p

problem (4.1.1) if and only if the passive transmission problem is feasible.

To complete this step, we show how (4.5.2) can be reformulated as (4.5.3). First we
rewrite (4.5.2) and reformulate the equalities to obtain equivalent formulas

Aa (qa∗ , pv , pw ) ≤ 0,
T

r  (4.5.5)
pw = sgn(pv |pv | − Φa (qa∗ )) pv |pv | − Φa (qa∗ ),
 

describing the feasible pressures pv , pw of (4.5.2). We observe that the constraints

Aa (qa∗ , pv , pw ) ≤ 0 form a polyhedron Pa∗ . According to (4.5.5) the feasible solution
T

space of pv and pw is then given as an intersection of this polyhedron Pa∗ with a

curve defined by
r 
g : pv 7→ sgn(pv |pv | − Φa (qa∗ )) pv |pv | − Φa (qa∗ ). (4.5.6)
 

This intersection, as illustrated in Figure 4.4, leads to intervals I˜a,v and I˜a,w such
(i) (i)

that

(pv , pw ) ∈ Pa∗ : g(pv ) = pw (pv , pw ) ∈ I˜a,v

(i)
× I˜a,w
(i)
: g(pv ) = pw
[
⇔
i

The computation of these intervals can be performed as follows: First we compute all
vertices u1 , . . . , un of the polyhedron Pa∗ . We assume an order such that ui and ui+1
are neighbored vertices for all i = 1, . . . , n − 1. Considering two neighbored vertices
we check whether their connecting line segment intersects the function (4.5.6). If
this is the case, then we store the intersection point. As a result we get a set of
4.6 Integration and Computational Results 95

pw

g(pv )

p00w

Pa∗

p0w

pv
p0v p00v

Figure 4.4: The figure is used to analyze the impact of the constraints (4.1.1c) on a feasible solution of
the active transmission problem without (4.1.1c). For this solution especially the flow value is available.
Assume that we fix this value and concentrate on the constraints (4.1.1b), (4.1.1c) and (4.1.1e). For a
single arc they write as (4.5.2). The figure shows the intersection of the polyhedron Pa∗ describing the
feasible operation range for qa = qa ∗
with the curve g(pv ) determined by the equation Φ(qa ∗
) = πv − πw =
pv |pv | − pw |pw |. In this case the intersection intervals yield Aa (qa ∗
, pv , pw )T ≤ 0 : g(pv ) = pw ⇔
(pv , pw ) ∈ [p0v , p00
v ] × [pw , pw ] : g(pv ) = pw .
0 00

intersection points that we sort in non-decreasing order. Two points consecutively

limit feasible and infeasible regions alternately. From a pair of these nodes limiting a
feasible region we derive the intervals I˜a,v × I˜a,w . From these intervals which limit
(i) (i)

(i) (i)
the pressure variables p we derive intervals Ia,v × Ia,w limiting the node potentials π
by taking the relation p|p| = π into account.

4.6 Integration and Computational Results

In this chapter we focused on the topology optimization problem (3.2.1) arising from
the first type of network that we consider in this thesis. Recall that these networks
consist of pipes and valves only. In this case it holds y = y for our model (3.2.1).
The outline of the solution framework that we apply is shown in Figure 4.5. We solve
(3.2.1) by SCIP as described in Section 2.2. For each node of the branching tree we
check whether the solution of the LP relaxation yields integral values for the integral
variables. If this is the case, then we consider the corresponding passive transmission
problem (4.1.1). We solve this nonlinear optimization problem to global optimality
and further classify the current MILP feasible node. If the passive transmission
problem is feasible, then we obtain a global optimal solution for the current node
of the branching tree and SCIP itself prunes the node. Otherwise, if the passive
transmission problem is infeasible, and all integral variables are fixed by branching,
96 Chapter 4 Efficiently Solving the Passive Transmission Problem

Branch-and-bound, separation, and spatial branching

node of branching tree

feasible node with feasible node infeasible node

MILP solution with fixed x, y

infeasible NLP feasible NLP infeasible NLP

LP
N

feasible node
infeasible node
globally solved

Figure 4.5: Solution framework presented in Section 4.6. The topology optimization problem (3.2.1) is
solved with SCIP essentially by branch-and-bound, separation, and spatial branching, see Section 2.2.
We adapt this framework and solve globally the passive transmission problem (4.1.1) as discussed in
Section 4.5, classify the current node of the branching tree and prune it if possible.

then we prune the current node of the branch-and-bound tree manually. If it is not
possible to solve the passive transmission problem globally due to numerical troubles
then we continue with branching.
We implemented the algorithms described in the previous section in C, i.e.,
domain relaxation (4.2.1), flow conservation relaxation (4.3.1), and the disjunctive
problem (4.5.4). The computational setup is described in Section 3.5. We compare
four strategies for solving the passive transmission problem (4.1.1).

1. The first strategy is to solve the topology optimization problem by SCIP

without any adaptations on the solver settings. All branching decisions are up
to the solver, and the topology optimization problem (3.2.1) is basically solved
by branch-and-bound, separation and spatial branching.

2. The second strategy is to solve the topology optimization problem by SCIP

and enforce a certain branching priority rule, so that SCIP first branches on
binary variables x. Only after all these variables are fixed, it is allowed to
perform spatial branching on continuous variables.

3. The third strategy implements the domain relaxation from Section 4.2 for
solving the passive transmission problem (4.1.1). We consider the passive
transmission problem and apply presolve and then solve the LP relaxation.
If these solution methods detect infeasibility, then the passive transmission
problem is infeasible. Otherwise we proceed as described in Section 4.5 and use
4.6 Integration and Computational Results 97

the nonlinear solver IPOpt for solving the convex domain relaxation (4.2.1).
Additionally we set branching priorities according to the second strategy.

4. The fourth strategy uses the relaxation of the flow conservation constraints
from Section 4.3 for solving the passive transmission problem (4.1.1). Again, we
first presolve the passive transmission problem and then solve the LP relaxation.
If infeasibility is detected, then the passive transmission problem is infeasible.
Otherwise we proceed as described in Section 4.5 and use the nonlinear solver
IPOpt for solving the convex flow conservation relaxation (4.3.1). Additionally
we set branching priorities according to the second strategy.

We did not implement the flow-coupling-constraint relaxation (4.4.1) described in

Section 4.4, because the relaxation is non-convex, and thus a local solver cannot
guarantee to find a global optimal solution. This is necessary for pruning nodes of
the branch-and-bound tree. Further we note that presolving and solving the LP
relaxation is done efficiently by SCIP and hence included in strategy 3 and 4.

Computational Results
We consider those test instances described in Section 3.5 which belong to the first
type of network. These networks are net4 and net5 while we contract all arcs which
represent compressors and control valves. This means that we identify the end nodes
of these arcs. This way the networks consist of pipes and valves only. Note that
the instances become harder at first sight because the deleted active elements do
not enforce a fixed relation between the flow and the node potentials at their end
nodes compared to pipes. But on the other hand a possible flow bound enforcing a
positive arc flow through a compressor is removed by this contraction. In this sence
the instances are relaxed and easier to solve.
We have 52 nominations in total. For every pipe a of these networks it holds
αa > 0 and Aa = 0, ba = 0. Hence solving the passive transmission problem in the
third and fourth strategy means solving the convex domain relaxation (4.2.1) and
flow conservation relaxation (4.3.1), respectively. It is not necessary to solve (4.5.4)
in advance as described in Section 4.5. Furthermore these networks do not contain
any compressors and control valves (as they are contracted). So fixing all binary
variables of the topology optimization problem (3.2.1) yields the passive transmission
problem (4.1.1) as assumed in this chapter.
For the computations we imposed a time limit of 39 600 s and used the com-
putational setup described in Section 3.5. The results are available in Table A.4
and Table A.5. A summary is shown in Table 4.1 and Table 4.2. Here we use the
98 Chapter 4 Efficiently Solving the Passive Transmission Problem

strategy 1 2 3 4 all
solved instances 24 30 45 46 46

Table 4.1: Summary of the Tables A.4 and A.5 showing the globally solved instances out of 52 nominations
in total. The third strategy globally solves all instances which are solved to global optimality by the second
and the first strategy.

(A,B) = (2,3) (A,B) = (4,3)

solved(30) incomp.(1) solved(45) incomp.(4)
time [s] nodes gap [%] time [s] nodes gap [%]
strategy A 25.9 1,038 15 66.2 827 141
strategy B 7.2 147 15 33.1 746 122
shifted geom. mean −72 % −86 % 0% −50 % −10 % −13 %

Table 4.2: Run time, number of branch-and-bound nodes and gap comparison for the strategies 2 and
3 and additionally 4 and 3 (aggregated results). The columns “solved” contain mean values for those
instances globally solved by both strategies A and B. The columns “incomplete” show mean values for
those instances having a primal feasible solution available but were not globally solved by both strategies
A and B. The underlying data are available in Tables A.4 and A.5.

geometric mean of run time, number of branch-and-bound nodes and gap as described
in Section 3.5.
Table 4.1 shows a clear order of the four strategies: Strategy 1 solves less instances
than strategy 2. We conclude that branching priorities as set by the second strategy
are a first step to improve the solving performance of SCIP. Approximately 57 %
of the instances (30 out of 52) are solved to global optimality by strategy 2. These
instances are also globally solved by strategy 3. Additionally around 29 % more
instances of the test set (15 out of 52) are solved to global optimality compared to
the second strategy. The fourth strategy solves one more instance than the third
strategy.
Table 4.2 shows that the third strategy is two times faster than the fourth
one while the the number of nodes is only reduced by 10 %. Figure 4.6 yields an
explanation. One can see that the run time for the domain relaxation is one order
of magnitude lower than for the flow conservation relaxation. We depict only those
instances that were solved by IPOpt, but not those that were detected as infeasible
during the presolve. In average the gap is reduced by 13 % following strategy 3
in comparison to strategy 4 on those instances which remain with a finite positive
gap value following both strategies. As both strategies solve globally nearly the
same number of instances we conclude that the domain relaxation (4.2.1) described
in Section 4.2 for solving the passive transmission problem (4.1.1) yields the most
4.6 Integration and Computational Results 99

net5
net4
100

Run time domain relaxation (sec)

10−1

10−2

10−2 10−1 100

Run time flow conservation relaxation (sec)

Figure 4.6: Run time comparison for the domain relaxation (4.2.1) and the flow conservation relax-
ation (4.3.1) on instances of the networks net4 and net5. The two crosses above the straight line belong
to passive transmission problems where the domain relaxation had numerical troubles.

efficient results comparing strategy 3 and 4. Hence we decided to use strategy 3 in

our practical application.
In order to demonstrate the benefit of strategy 3 we finally compare it with the
second strategy. Recall from the previous analysis that the third strategy solves
approximately 29 % more instances to global optimality. Table 4.2 shows that
strategy 3 saves 72 % of run time and 86 % of nodes in comparison to strategy 2.
The savings in terms of number of nodes are higher than the run time reduction due
to the following reason: the relaxations are set up and the nonlinear solver IPOpt
has to be called. In summary we conclude that more instances are globally solved
and less run time is needed.
Figure 4.7 summarizes the run time results of the four strategies in a performance
plot. The four graphs show the share of instances (in per cent) that could be solved
within a time limit of 39 600 s. Evaluating Figure 4.7 shows a result coherent to
our previous observations: the graphs for SCIP (strategy 1 and 2) are below the
graph for the flow conservation relaxation, which is below the graph for the domain
relaxation. Figure 4.8 shows a consistent result. It shows a scatter plot comparing
the run times of the second and the third strategy (the best strategy following spatial
branching vs. the best strategy using the convex relaxations of this chapter). The
run time of many instances is reduced when using the third strategy.
100 Chapter 4 Efficiently Solving the Passive Transmission Problem

100
strategy 3
strategy 4
strategy 2
Number of solved instances (%)
strategy 1
75

50

25

0
10−1 100 101 102 103 104
Run time (sec)

Figure 4.7: Performance plot for different nominations on the networks net4 and net5 (aggregated) and
a time limit of 39 600 s. The different strategies are described in Section 4.6. Strategy 1 and 2 mainly
consist of SCIP. Strategies 3 and 4 also correspond to SCIP together with our elaborated solution methods
presented in this chapter. The underlying data are available in Tables A.4 and A.5.

105

104
Run time strategy 3 (sec)

103

102

101

100

10−1

10−2

10−2 10−1 100 101 102 103 104 105

Run time strategy 2 (sec)
Figure 4.8: Run time comparison for different nominations on the networks net4 and net5. Each cross
(×) corresponds to a single instance of the test set. Note that multiple crosses are drawn in the upper
right corner of the plots that cannot be differed. They represent those instances that ran into the time
limit of 39 600 s for both strategies. The underlying data are available in Tables A.4 and A.5.
4.6 Integration and Computational Results 101

We refer to Humpola et al. (2015b) where we carried out a computational study

for different nominations on the real-world network net6 by applying the results
of this chapter. This network contains compressors and control valves. Therefor
we discretized the feasible solution space of these active elements so that ya,i ∈ Z
holds for every arc (a, i) ∈ AX . Recall that y are continuous variables in our
model. So fixing binary and discrete variables x and y in the topology optimization
problem (3.2.1) yields the passive transmission problem. Additionally a method is
presented which computes the coefficients αa for each compressor and control valve
a ∈ A according to this discretization. The computational results show large running
times that increase further with the number of discretizations. A large number of
discretizations is in turn necessary to reduce the approximation error for compressors
and control valves. We conclude that the algorithms of this chapter are mainly useful
for networks containing only pipelines and valves.

Summary

We presented different solution methods for the passive transmission problem. They
allow to speed up the solution process of the topology optimization problem in the
case y = y. This case includes the first type of network of our test instances, namely
those which contain only pipes and valves. From the computational results we
conclude that a strategy which is based on solving the domain relaxation yields the
most convincing results. This strategy shows faster computation times in comparison
to SCIP. On average the run time is reduced by 72 % in comparison to SCIP with
branching priorities set. Approximately 29 % more instances of the test set are also
solved to global optimality compared to the solver SCIP. Therefore in our practical
applications we use the domain relaxation method for networks that contain only
pipes and valves.
Let us now try to understand the reason for the performance of SCIP. Therefor
we considered a specific part of the solution process for a test instance. We analyzed
the computational effort for handling the nonlinear equations of our model that are
associated with pipes. As a result we collected all cutting planes that are generated
during the branch-and-bound process for each pipe individually. The result for one
single pipe is visualized in Figure 4.9.
Many inequalities are required for the approximation of the convex part of the
function qa 7→ αa qa |qa |ka . We conclude that the use of nonlinear solvers on specific
nodes of branch-and-bound reduces the computational effort in handling nonlinear
equations compared to spatial branching.
102 Chapter 4 Efficiently Solving the Passive Transmission Problem

3,000

2,000

1,000
αa qa |qa |

−1,000

−2,000

−3,000
-10000 0 10000
qa

Figure 4.9: Handling the nonlinearity during the branch-and-bound process. For handling the nonlinear
function qa 7→ αa qa |qa | of a single pipe a with αa = 1/37 700, cutting planes were generated for all
marked points.
Chapter 5

An Improved Benders Cut for the

Topology Optimization Problem

In Chapter 4 we presented two convex relaxations (4.2.1) and (4.3.1) of the passive
transmission problem (4.1.1). Both allowed us to solve the passive transmission
problem efficiently and thereby reduce the use of time consuming spatial branching
when solving the topology optimization problem (3.2.1). We focused on the case
y = y which includes networks containing only pipes and valves. These networks
are the first type of network which we consider in this thesis. We now focus on the
second type of gas network which additionally contains loops or rather parallel pipes.
Again we restrict to the case y = y.
In this chapter we show that a KKT point of the previously mentioned relaxations
can be used to generate a new linear inequality for (3.2.1). By fixing all integral
variables in (3.2.1) we obtain the passive transmission problem. Whenever one of the
relaxations (4.2.1) and (4.3.1) of the passive transmission problem is solved during
the branch-and-bound process and when we conclude that the passive transmission
problem is infeasible, then we dynamically add such an inequality to the problem
formulation. This act allows to solve approximately 13 % more instances of the second
type of network to global optimality within our given time limit. For those instances
which are already solvable by SCIP the run time is reduced by approximately 33 %.
The linear inequality that we are going to present contains only binary variables.
When fixing the discrete variables to those values leading to the passive transmission
problem from which the inequality is derived, then it reflects the infeasibility of
the corresponding passive transmission problem, i.e., it is violated if and only if
the passive transmission problem is infeasible. As the inequality does not contain
any continuous variable we are reminded of generalized Benders decomposition (see
Geoffrion 1972). Here the master problem would consist of the integral variables
and the constraints containing only integral variables and we would write it as

103
104 Chapter 5 An Improved Benders Cut

min{cx | x ∈ X , x ∈ {0, 1}AX }. Each subproblem would be a feasibility problem

equal to the passive transmission problem (4.1.1). Nevertheless, we do not follow
this decomposition approach, but restart the overall branch-and-bound process after
a predefined number of inequalities is generated.
The inequality presented in this chapter is obtained by a Benders argument from
the Lagrange function of the domain relaxation (4.2.1) augmented by a specially
tailored pc-regularization. This regularization is necessary to derive a globally valid
cut. A certain choice of the Lagrange multipliers finally yields the described properties.
Hence we consider our inequality as an improved Benders cut.
The outline of this chapter is as follows: In Section 5.1 we present a linear
inequality for the passive transmission problem (4.1.1) which represents its feasibility.
In Section 5.2 we derive an extended formulation of this inequality which is then
valid for the topology optimization problem (3.2.1). Computational results are given
in Section 5.3. They show the benefit of our inequalities when added to the topology
optimization problem (3.2.1).

5.1 Valid Inequalities for the Passive Transmission

Problem
Let us consider the passive transmission problem which is obtained from the topology
optimization problem (3.2.1) by fixing all integral and continuous variables x and y.
Recall that the variables y are fixed for the networks considered in this chapter. In
Section 5.1.1 we first formulate a valid nonlinear inequality for this problem. This
inequality bases on the definition of a dual transmission flow (see Definition 5.1.1). In
Section 5.1.2 we explain how to obtain a linear inequality, which is constant on both
sides and valid for the passive transmission problem. The choice of the parameters
of the linear inequality is discussed in Section 5.1.3. In Section 5.1.4 we explain the
relation between the linear inequality that we derive in this section and the Lagrange
function of the domain relaxation (4.2.1).
Throughout this section let A0 contain all arcs such that the flow is not fixed to zero,
i.e., A0 := {(a, i) ∈ AX | xa,i = 1, i > 0}. We review the domain relaxation (4.2.1) to
improve readability:

min ∆v + ∆a (5.1.1a)
X X

v∈V a∈A0

s. t. αa qa |qa |ka − β̃a − (πv − γa πw ) = 0 ∀ a = (v, w) ∈ A0 , (5.1.1b)

5.1 Valid Inequalities for the Passive Transmission Problem 105

qa = dv (5.1.1c)
X X
qa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

π v − ∆v ≤ π v ∀ v ∈ V, (5.1.1d)

πv + ∆ v ≥ π v ∀ v ∈ V, (5.1.1e)

qa − ∆a ≤ q a ∀ a ∈ A0 , (5.1.1f)

qa + ∆a ≥ q a ∀ a ∈ A0 , (5.1.1g)

πv ∈ R ∀ v ∈ V, (5.1.1h)

qa ∈ R ∀ a ∈ A0 , (5.1.1i)

∆v ∈ R≥0 ∀ v ∈ V, (5.1.1j)

∆a ∈ R≥0 ∀ a ∈ A0 . (5.1.1k)

Let us turn to the definition of a dual transmission flow. Therefor we write Φa (qa ) :=
αa qa |qa |ka − β̃a and recall the Lagrange function of this problem as

L(q, π, ∆, µ, λ) = ∆v + ∆a (5.1.2)
X X

v∈V a∈A0

+ µa Φa (qa ) − (πv − γa πw )
X

a=(v,w)∈A0
 

+ qa +
X X X
µv dv − qa 
 
+
0 (v) 0 (v)
v∈V −
a∈δA a∈δA

λ+
X 
+ v (πv − ∆v − π v ) + λv (π v − πv − ∆v )
−

v∈V

λ+
X 
+ a (qa − ∆a − q a ) + λa (q a − qa − ∆a )
−

a∈A0

λv ∆v − λa ∆a .
X X
−
v∈V a∈A0

In Section 4.2.3 we showed that the dual values of a KKT point (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ )

of domain relaxation (4.2.1), which is indexed in this chapter by (5.1.1), form a
network flow (µ∗a )a∈A0 , which is induced by dual node potentials (µ∗v )v∈V . We recall
the constraints (4.2.7) that are fulfilled by the dual variables (µ∗ , λ∗ ):

dΦa ∗
µa (q ) + λ+
a − λa = µv − µw
−
∀ a = (v, w) ∈ A0 , (5.1.3a)
dqa a
106 Chapter 5 An Improved Benders Cut

µa γa = λ+ (5.1.3b)
X X
−
µa − v − λv ∀ v ∈ V.
+
0 (v) 0 (v)
−
a∈δA a∈δA

Relation (5.1.3b) is the basis for the inequalities that we consider in this chapter.
Therefor we give the following definition:

Definition 5.1.1:
Every vector (µ, λ) = (µv , µa , λ+ +
v , λv , λa , λa )v∈V,a∈A , such that µv , µa ∈ R and
− −

λ+ +
v , λv , λa , λa ∈ R≥0 , which fulfills the constraints
− −

γa µa = λ+ (5.1.4)
X X
−
µa − v − λv ∀ v ∈ V,
+
a∈δA (v) −
a∈δA (v)

is called dual transmission flow. We regard the vector (µa )a∈A of this dual
transmission flow as a generalized flow in the original network (V, A).

In the remainder of this section we do not differ between an arc a ∈ A and the
corresponding arc a0 = (a, i) ∈ A0 . For a dual transmission flow (µ, λ) this implies
that we speak of µa0 which is defined as µa0 := µa for a ∈ A and a0 = (a, i) ∈ A0 .

5.1.1 A Nonlinear Inequality

In the next two lemmata we derive two different inequalities for the passive transmis-
sion problem (4.1.1). They both only contain the flow variables q. The right-hand
side of both inequalities is constant. The left-hand side of the first inequality is a
non-convex function in q while the left-hand side of the second inequality is quasi-
convex in q. We will consider a linear combination of both inequalities which then
allows to project out the flow variables q as described in Section 5.1.2.
The first inequality is derived from any dual transmission flow.

Lemma 5.1.2:
For any dual transmission flow (µ, λ) with µa = 0 for all arcs a ∈
/ A0 , the inequality
in q

λ+
X 
µa Φa (qa ) ≤
X
−
v π v − λv π v
a∈A0 v∈V

is valid for the passive transmission problem (4.1.1).

5.1 Valid Inequalities for the Passive Transmission Problem 107

Proof. We prove the following estimation which is valid for any solution (q ∗ , π ∗ ) of
the passive transmission problem (4.1.1):

µa Φa (qa ) = µa (πv − γa πw ) (5.1.5a)

X X

a∈A0 a=(v,w)∈A0
 

= (5.1.5b)
X  X X
πv  µa − γa µa 

+
0 (v) 0 (v)
v∈V −
a∈δA a∈δA
 

= (5.1.5c)
X  X X
πv  µa − γa µa 

+
v∈V a∈δA (v) −
a∈δA (v)

= πv (λ+
v − λv ) (5.1.5d)
X
−

v∈V

(λ+
v π v − λv π v ). (5.1.5e)
X
−
≤
v∈V

To obtain (5.1.5a) we multiply equation (4.1.1b) by µa and sum over all arcs a ∈ A0 .
Then we rewrite the right-hand side by changing the order of summation and obtain
(5.1.5b). Note that all arcs a ∈ A\A0 have µa = 0, hence they can be added and we
obtain (5.1.5c). We use equation (5.1.4) in Definition 5.1.1 of a dual transmission flow
and obtain (5.1.5d). Finally we estimate the right-hand side and obtain (5.1.5e).

Remark 5.1.3:
We briefly explain the practical meaning of the inequality of Lemma 5.1.2. Therefor
let (µ, λ) be a dual transmission flow with µa = 0 for all arcs a ∈
/ A0 . Furthermore
we assume µ ≥ 0 and γa = 1, a ∈ A for this explanation. We split the network
flow µ into sets of flow along paths P1 , . . . , Pm and flow along circuits C1 , . . . , Cn ,
see Theorem 4.2.5. This way we obtain flow values µPi > 0, i = 1, . . . , m and
µCi > 0, i = 1, . . . , n such that

µa = µPi +
X X
µCi ∀ a ∈ A0
i=1,...,m: i=1,...,n:
a∈A0 (Pi ) a∈A0 (Ci )

holds. We use this equation and denote the start and end node of Pi by s(Pi ) and
t(Pi ) to write the inequality of Lemma 5.1.2 in a different way. For any primal
solution (q ∗ , π ∗ ) of the passive transmission problem it holds

m
X  
µPi πs(Pi ) − πt(Pi )
i=1
108 Chapter 5 An Improved Benders Cut

=πs(P ) −πt(P ) =0
m z X i}| i
{ X n z X }| {
= Φa (qa ) + Φa (qa )
X
µPi µCi
i=1 a∈A0 (Pi ) i=1 a∈A0 (Ci )

= Φa (qa ) µPi + Φa (qa )

X X X X
µCi
a∈A0 i:a∈A0 (Pi ) a∈A0 i:a∈A0 (Ci )

= Φa (qa ) µa
X

a∈A0
n
λ+
X  X  
≤ π
v v − λ−
π
v v = µPi π s(Pi ) − π t(Pi ) .
v∈V i=1

In this argumentation we used the pressure conservation along circuits, i.e., the sum
of potential differences along a circuit equals zero. We conclude that Lemma 5.1.2
presents a valid inequality for the passive transmission problem that requires the
weighted sum of the potential losses πs(P
∗
i)
∗
− πt(Pi)
along the paths Pi to be bounded
by the weighted sum of the available potential losses π s(Pi ) − π t(Pi ) along these paths.

Now we turn to the second inequality. For the next lemma we make use of γr,v
and the function π 0 from Definition 4.2.3, and recall the reformulation (4.2.5):

πv0 (π) − πw
0
(π) = γr,v (πv − γa πw ).

For abbreviations we defined π 0v := πv0 (π) and π 0v := πv0 (π) for each node v ∈ V .
Furthermore we set
Φ0a (qa ) := γr,v Φa (qa ) (5.1.6)

for every arc a = (v, w) ∈ A0 .

Lemma 5.1.4:
Let q ∗ ∈ RA , ∆±
v ∈ R≥0 , v ∈ V , and ∆a ∈ R≥0 , a ∈ A be vectors such that the flow
0
± 0

conservation

(qa∗ − (∆+
a − ∆a )) − (qa∗ − (∆+ +
a − ∆a )) − (∆v − ∆v ) = dv
X X
− − −

+
0 (v) 0 (v)
−
a∈δA a∈δA

is fulfilled. Then the inequality in q

γr,v (qa − qa∗ ) Φa (qa )

X

a=(v,w)∈A0

(∆− − ∆+
v πv ) + (∆− + 0
a Φa (q a ) − ∆a Φa (q a ))
X X
0 0 0
≤ v πv
v∈V a∈A0
5.1 Valid Inequalities for the Passive Transmission Problem 109

is valid for the passive transmission problem (4.1.1)

Proof. Let (q 0 , π 0 ) be a feasible solution of the passive transmission problem (4.1.1).

By (4.1.1d) the flow vector q 0 fulfills the flow conservation

qa = dv
X X
qa −
+
0 (v) 0 (v)
−
a∈δA a∈δA

for all nodes v ∈ V . From this we derive that q 0 is feasible for

(qa − qa∗ ) − (qa − qa∗ ) =

X X
+
0 (v) 0 (v)
−
a∈δA a∈δA

+ + +
(∆−
v − ∆v ) + (∆−
a − ∆a ) − (∆−
a − ∆a )
X X
+
0 (v) 0 (v)
−
a∈δA a∈δA

for all nodes v ∈ V . We multiply each side by πv0 (π), take the sum over all nodes
v ∈ V and obtain:

(qa − qa∗ )(πv0 (π) − πw (π))

X
0

a=(v,w)∈A0
+ +
= πv0 (π)(∆−
v − ∆v ) + (∆−
a − ∆a )(πv (π) − πw (π))
X X
0 0

v∈V a=(v,w)∈A0
+ +
v − πv (π)∆v ) +
(πv0 (π)∆− (∆−
a − ∆a ) Φa (qa )
X X
0 0
≤
v∈V a∈A0

(π 0v ∆− π 0v ∆+
v) + + 0
a Φa (q a ) − ∆a Φa (q a )).
(∆−
X X
0
≤ v −
v∈V a∈A0

The estimations are obtained by taking the lower and upper bounds on πv0 (π) and qa
into account. We use this estimation to obtain from (4.1.1b) and (4.2.5)

γr,v (qa − qa∗ ) Φa (qa )

X

a=(v,w)∈A0

= (qa − qa∗ )γr,v (πv − γa πw )

X

a=(v,w)∈A0

= (qa − qa∗ )(πv0 (π) − πw (π))

X
0

a=(v,w)∈A0
+ + 0
(π 0v ∆−
v − π v ∆v ) + a Φa (q a ) − ∆a Φa (q a )).
(∆−
X X
0 0
≤
v∈V a∈A0

Hence every primal solution of the passive transmission problem fulfills the inequality
of the Lemma.
110 Chapter 5 An Improved Benders Cut

Remark 5.1.5:
We briefly explain the practical meaning of the inequality of Lemma 5.1.4. Therefor
let q ∗ fulfill the flow conservation constraint. Furthermore we assume γa = 1, a ∈ A
for this explanation. For any feasible solution (q̃, π̃) for the passive transmission
problem we obtain that q̃ − q ∗ forms a circulation. In this case the inequality of
Lemma 5.1.4 writes as

(q̃a − qa∗ ) Φa (qa ) ≤ 0

X

a=(v,w)∈A0

Let us assume that q̃ ≥ q ∗ . We split this network flow into sets of flow along circuits
C1 , . . . , Cn , see Theorem 4.2.5. We obtain flow values µCi > 0, i = 1, . . . , n such that

q̃a − qa∗ =
X
µCi ∀ a ∈ A0
i=1,...,n:
a∈A0 (Ci )

holds. We use this equation to obtain

(q̃a − qa∗ ) Φa (q̃a ) = Φa (q̃a )

X X X
µCi
a=(v,w)∈A0 a∈A0 i:a∈A0 (Ci )
n
= Φa (q̃a ).
X X
µCi
i=1 a∈A0 (Ci )

We conclude that Lemma 5.1.4 presents a valid inequality for the passive transmission
problem which requires the weighted sum of the potential losses along the circuits Ci
to be bounded by zero. Obviously a true statement because the loss of potential along
a circuit equals zero.

Now we consider a linear combination of both inequalities from the previous two
Lemmata 5.1.2 and 5.1.4:

Corollary 5.1.6:
Let (µ, λ) be a dual transmission flow with µa = 0 for all arcs a ∈
/ A0 . Furthermore
let q ∗ ∈ RA , ∆±
v ∈ R≥0 , v ∈ V , and ∆a ∈ R≥0 , a ∈ A be vectors such that the flow
0
± 0

conservation

(qa∗ − (∆+
a − ∆a )) − (qa∗ − (∆+ +
a − ∆a )) − (∆v − ∆v ) = dv
X X
− − −

+
0 (v) 0 (v)
−
a∈δA a∈δA
5.1 Valid Inequalities for the Passive Transmission Problem 111

is fulfilled. Then for any ζ ∈ [0, 1] the inequality in q

(ζ γr,v (qa − qa∗ ) + (1 − ζ)µa ) Φa (qa )

X

a∈A0
+ 0 + 0
(∆−
v π v − ∆v π v ) + ζ a Φa (q a ) − ∆a Φa (q a ))
(∆−
X X
0 0
≤ζ (5.1.7)
v∈V a∈A0

+(1 − ζ) (λ+
v π v − λv π v )
X
−

v∈V

is valid for the passive transmission problem (4.1.1).

Proof. The inequality is a linear combination of the inequalities from Lemma 5.1.2
and Lemma 5.1.4.

It is easy to see that the inequality from Lemma 5.1.4 is used to circumvent
unboundedness of the left-hand-side of (5.1.7). Hence we regard this second inequality
as a regularization. It allows us to use a linear underestimator for the left-hand-side
of inequality (5.1.7) in the ongoing part of this chapter. As this regularization
expresses the conservation of potential along circuits by Remark 5.1.5 we call it the
pc-regularization.

5.1.2 A Linear Inequality

So far we derived a nonlinear inequality stated in Corollary 5.1.6 which is valid
for the passive transmission problem (4.1.1). In principle, it would be possible
to add this inequality to the topology optimization problem (3.2.1) at different
nodes of the branch-and-bound tree. But this act would increase the number of
nonlinearities of the model. Hence, in the following, we describe how to derive a linear
inequality. Therefor we will use a certain linear underestimator for the left-hand side
of inequality (5.1.7). This left-hand side is a sum of functions in qa over the arcs
a ∈ A0 . We consider each of these functions

fζ,q∗ ,µ (qa ) := (ζ γr,v (qa − qa∗ ) + (1 − ζ)µa ) Φa (qa )

separately and give a linear underestimator. For the underestimation we use the
function

`ζ,π,µ,λ (qa ) := ζ (πv0 (π) − πw

0
(π)) qa + (1 − ζ)(µv − µw − λ+
a + λa ) qa
−

for an arc a = (v, w) ∈ A0 . This function is used to linearize inequality (5.1.7) as

follows:
112 Chapter 5 An Improved Benders Cut

Lemma 5.1.7:
Let (µ, λ) be dual transmission flow with µa = 0 for all arcs a ∈
0
/ A0 . Let q ∗ ∈ RA ,
π ∗ ∈ R V , ∆±
v ∈ R≥0 , v ∈ V , and ∆a ∈ R≥0 , a ∈ A be vectors such that the flow
± 0

conservation

(qa∗ − (∆+
a − ∆a )) − (qa∗ − (∆+ +
a − ∆a )) − (∆v − ∆v ) = dv
X X
− − −

+
0 (v) 0 (v)
−
a∈δA a∈δA

is fulfilled for each node v ∈ V . Furthermore let ζ ∈ [0, 1]. Then for constants
τa := inf{fζ,q∗ ,µ (qa ) − `ζ,π∗ ,µ,λ (qa ) | q a ≤ qa ≤ q a } for each arc a ∈ A0 the inequality
X
τa ≤
a∈A0
 
+ 0 + 0
(∆−
v π v − ∆v π v ) + a Φa (q a ) − ∆a Φa (q a )) −
(∆− dv πv0 (π ∗ )
X 0
X 0
X
ζ
(5.1.8)
v∈V a∈A0 v∈V
 
X 
+(1 − ζ)  λ+
v π v − λv π v + λ+
−
X −

X
a q a − λa q a − d v µv 
v∈V a∈A0 v∈V

is valid for the passive transmission problem (4.1.1).

Remark 5.1.8:
We note that the unfixed variables in the passive transmission problem are flow vari-
ables q and node potentials π. These variables are not contained in inequality (5.1.8).
Thus inequality (5.1.8) is constant on both sides, left- and right-hand side, and hence
linear. If the inequality is violated, then the passive transmission problem is infeasible.

Using the underestimator `(qa ) to reformulate (5.1.7) we obtain inequality (5.1.8)

which depends on the dual transmission flow (µ, λ) and other parameters q ∗ , π ∗ , ∆,
and ζ. The choice of these parameters such that (5.1.8) represents the infeasibility of
the passive transmission problem (4.1.1) will be discussed in Section 5.1.3.

Proof. [Lemma 5.1.7 ] Let τa := inf{fζ,q∗ ,µ (qa ) − `ζ,π∗ ,µ,λ (qa ) | q a ≤ qa ≤ q a } for
each arc a ∈ A0 . We consider inequality (5.1.7) of Corollary 5.1.6. By the definition
of τa we obtain the underestimator for each summand on the left-hand side as

fζ,q∗ ,µ (qa ) ≥ τa + `ζ,π∗ ,µ,λ (qa )

≥ τa + ζ (πv0 (π ∗ ) − πw
0
(π ∗ )) qa + (1 − ζ)(µv − µw − λ+
a + λa ) qa .
−
5.1 Valid Inequalities for the Passive Transmission Problem 113

We rewrite the underestimator. Each flow vector q 0 which is feasible for (4.1.1) fulfills
the flow conservation constraint

qa = dv .
X X
qa −
+
0 (v) 0 (v)
−
a∈δA a∈δA

Multiplying this equation with πv0 (π ∗ ) and summing over the nodes v ∈ V we obtain
that q 0 is feasible for

(πv0 (π ∗ ) − πw (π ∗ )) qa = ζ dv πv0 (π ∗ ) (5.1.9)

X X
0
ζ
a=(v,w)∈A0 v∈V

and multiplying the flow conservation constraint with µv and summing over the
nodes v ∈ V we derive

(1 − ζ) (µv − µw ) qa = (1 − ζ) (5.1.10)
X X
dv µv .
a=(v,w)∈A0 v∈V

Using the reformulations (5.1.9), (5.1.10) we obtain that q 0 is feasible for

fζ,q∗ ,µ (qa )
X

a∈A0
(5.1.11)
τa + ζ dv πv0 (π ∗ ) + (1 − ζ) dv µv − (1 − ζ) (λ+
a − λa )qa .
X X X X
−
≥
a∈A0 v∈V v∈V a∈A0

We use the lower and upper bounds on qa to obtain

(1 − ζ) (λ+
a − λa )qa ≤ (1 − ζ) (λ+
a q a − λa q a ).
X X
− −

a∈A0 a∈A0

Now (5.1.11) writes as

X
τa ≤
a∈A0

fζ,q∗ ,µ (qa ) + (1 − ζ) (λ+

a q a − λa q a )
X X
−

a∈A0 a∈A0

dv πv0 (π ∗ ) − (1 − ζ)
X X
−ζ dv µv .
v∈V v∈V

Inequality (5.1.7) of Corollary 5.1.6 yields an over estimator for fζ,q∗ ,µ (qa ).
P
a∈A0
Applying this over estimator proves the lemma.
114 Chapter 5 An Improved Benders Cut

5.1.3 Feasibility Characterization by a Linear Inequality

So far we proved that inequality (5.1.8) is valid for the passive transmission prob-
lem (4.1.1). The upcoming question is how to choose the dual transmission flow
(µ, λ) and the parameters q ∗ , π ∗ , ∆, ζ in order to obtain an inequality that represents
the infeasibility of the passive transmission problem. We want to obtain an inequality
that is violated if the passive transmission problem is infeasible. Therefor, as stated
in Definition 5.1.11, we derive a dual transmission flow from a KKT point of the
domain relaxation (4.2.1) or the flow conservation relaxation (4.3.1) and discuss a
suitable choice of the parameters q ∗ , π ∗ , ∆, ζ.
The first lemma describes the choice of the parameter ζ for inequality (5.1.8) for
the case that the dual transmission flow (µ, λ) and the parameters q ∗ and π ∗ are
given. The results are important for the subsequent lemmata of this section.

Lemma 5.1.9:
Let (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) be a KKT point of the domain relaxation (5.1.1) or the flow
conservation relaxation (4.3.1). Let ζ ∈]0, 1] such that the following conditions hold
for every arc a = (v, w) ∈ A0 :

1. if µa qa∗ > 0, then (1 − ζ)|µ∗a | < ζ γr,v |qa∗ |,

2. if µa qa∗ < 0, then (1 − ζ) |µ∗v − µ∗w − λ∗a + + λ∗a − | < ζ γr,v |πv∗ − γa πw
∗
− β̃a |,

3. if µa qa∗ = 0, then (1 − ζ) µ∗a = 0.

Then the minimum of the function

fζ,q∗ ,µ∗ (qa ) − `ζ,π∗ ,µ∗ ,λ∗ (qa ) (5.1.12)

is attained at qa∗ for every arc a ∈ A0 .

Remark 5.1.10:
It is not guaranteed that a value ζ ∈]0, 1] exists which fulfills the conditions of
Lemma 5.1.9.

Proof. [Lemma 5.1.9 ] Let (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) be a KKT point of the domain relax-

ation (5.1.1) or the flow conservation relaxation (4.3.1). Let ζ ∈]0, 1] such that condi-
5.1 Valid Inequalities for the Passive Transmission Problem 115

tions 1, 2 and 3 are fulfilled for every arc a ∈ A0 . We consider an arc a = (v, w) ∈ A0
and write the derivative of function (5.1.12) as

∗ ,µ∗ (qa ) − `ζ,π ∗ ,µ∗ ,λ∗ (qa )

0 0
fζ,q
 
dΦa dΦa
= ζ γr,v (qa − qa∗ ) (qa ) + Φa (qa ) + (1 − ζ)µ∗a (qa )
dqa dqa
− ζ (πv0 (π ∗ ) − πw
0
(π ∗ )) − (1 − ζ)(µ∗v − µ∗w − λ∗a + + λ∗a − ).

We set qa = qa∗ and obtain from (5.1.1b), (4.3.1b), Lemma 4.2.11 and Lemma 4.3.7
and the relation (4.2.5)

∗ ,µ∗ (qa ) − `ζ,π ∗ ,µ∗ ,λ∗ (qa )

0 ∗ 0 ∗
fζ,q
= ζ (γr,v Φa (qa∗ ) − (πv0 (π ∗ ) − πw
0
(π ∗ )))
 
∗ dΦa ∗ ∗+
+ (1 − ζ) µa (q ) − (µv − µw − λa + λa )
∗ ∗ ∗−
dqa a
= ζ γr,v (Φa (qa∗ ) − πv∗ + γa πw
∗
)
 
∗ dΦa ∗ ∗+
+ (1 − ζ) µa (q ) − (µv − µw − λa + λa )
∗ ∗ ∗−
dqa a
= 0 + 0.

This implies that function (5.1.12) has an extreme point for qa = qa∗ .

We still have to prove that the point qa = qa∗ is a global minimum of (5.1.12)
for the choice of ζ. We write function (5.1.12) as g(qa ) − h(qa ) where g and h are
defined as
=:b>0 =:c =:d
z }| { z }| {
g(qa ) : = (ζ γr,v qa + (1 − γr,v qa∗ )αa − ((1 − − ζγr,v qa∗ )β̃a
z }| {
ζ)µ∗a −ζ ka
qa |qa | ζ)µ∗a
= (bqa + c) αa qa |qa |ka − d

and

+ β̃a ) qa + (1 − ζ)(µ∗v − µ∗w − λ+

∗ ∗
h(qa ) := ζγr,v (πv∗ − γa πw
∗
a + λa ) qa .
−

In the case αa = 0 it follows that g(qa ) is constant and (5.1.1b) and (5.1.3a) imply
h(qa ) ≡ 0. This implies that qa∗ is a global minimum of (5.1.12). To prove this for
the case αa > 0 we briefly show that the choice of ζ means that g(qa ) and h(qa ) have
the form as indicated in Figure 5.1 depending on the value of qa∗ . This implies that
g(qa ) − h(qa ) is a convex function. In combination with our previous analysis we
obtain that g(qa ) − h(qa ) has a global optimum at qa = qa∗ .
116 Chapter 5 An Improved Benders Cut

g(qa )
g(qa )
g(qa ) h(qa )
qa
qa qa
h(qa ) h(qa )
(a) Case qa
∗
> 0. (b) Case qa
∗
= 0. (c) Case qa
∗
< 0.

Figure 5.1: Visualization of the functions g(qa ) and h(qa ) defined in the proof of Lemma 5.1.9 for different
values of qa
∗
and αa > 0.

In the following we characterize g(qa ) and show c < 0 if qa∗ > 0, c > 0 if qa∗ < 0
and c = 0 if qa∗ = 0. This then proves that g(qa ) has the form as shown in Figure 5.1.
We distinguish three cases:

Case qa∗ > 0: It holds c = (1 − ζ)µ∗a − ζ γr,v qa∗ < 0:

• If µ∗a ≤ 0 then it holds (1 − ζ)µ∗a ≤ 0, −ζ γr,v qa∗ < 0. Note that ζ 6= 0.
• If µ∗a > 0 then it follows from assumption 1 that (1 − ζ)µ∗a = (1 − ζ)|µ∗a | <
|ζ γr,v qa∗ | = ζ γr,v qa∗ .

Case qa∗ = 0: It holds c = (1 − ζ)µ∗a − ζ γr,v qa∗ = 0 because (1 − ζ)µ∗a = 0 holds by

assumption 3.

Case qa∗ < 0: It holds c = (1 − ζ)µ∗a − ζ γr,v qa∗ > 0:

• If µ∗a ≥ 0 then it holds (1 − ζ)µ∗a ≥ 0, −ζ γr,v qa∗ > 0. Note that ζ 6= 0.
• If µ∗a < 0 then it follows from assumption 1 that −(1 − ζ)µ∗a = (1 − ζ)|µ∗a | <
|ζ γr,v qa∗ | = −ζ γr,v qa∗ and hence (1 − ζ)µ∗a > ζ γr,v qa∗ .

We now turn to the analysis of h(qa ).

Case qa∗ > 0: It holds ζγr,v (πv∗ − γa πw − β̃a ) + (1 − ζ)(µ∗v − µ∗w − λ+

a + λa ) > 0:
∗ − ∗ ∗

From (5.1.1b) it follows πv∗ − γa πw

∗
− β̃a > 0.
• If µ∗a ≥ 0 then it holds µ∗v − µ∗w − λ+
a + λa ≥ 0 by (5.1.3a).
− ∗ ∗

• If µ∗a < 0 then it follows from (5.1.3a) and assumption 2 that (1 − ζ)|µ∗v − µ∗w −
λ+
a + λa | < ζγr,v |πv − γa πw − β̃a | = ζγr,v (πv − γa πw − β̃a ).
∗ − ∗ ∗ ∗ ∗ ∗

Case qa∗ = 0: It holds ζγr,v (πv∗ −γa πw − β̃a )+(1−ζ)(µ∗v −µ∗w −λ+
a +λa ) = 0: From
∗ − ∗ ∗

(5.1.1b) it follows πv∗ −γa πw −β̃a = 0. From (5.1.3a) we obtain µ∗v −µ∗w −λ+
a +λa = 0.
∗ − ∗ ∗

Hence g(qa ) ≡ 0.

Case qa∗ < 0: It holds ζγr,v (πv∗ − γa πw − β̃a ) + (1 − ζ)(µ∗v − µ∗w − λ+

a + λa ) < 0:
∗ − ∗ ∗

From (5.1.1b) it follows πv∗ − γa πw

∗
− β̃a < 0.
5.1 Valid Inequalities for the Passive Transmission Problem 117

• If µ∗a ≤ 0 then it holds µ∗v − µ∗w − λ+

a + λa ≤ 0 by (5.1.3a).
− ∗ ∗

• If µ∗a > 0 then it follows from (5.1.3a) and assumption 2 that −(1 − ζ)(µ∗v −
µ∗w − λ+ +
a + λa ) = −(1 − ζ)|µv − µw − λa + λa | > ζγr,v (πv − γa πw − β̃a ).
∗ − ∗ ∗ ∗ − ∗ ∗ ∗ ∗

These cases show that h(qa ) is a linear function with positive slope if qa∗ > 0, negative
slope if qa∗ < 0 and slope zero if qa∗ = 0. Hence h(qa ) has the form as shown in
Figure 5.1.

The dual transmission flow which is necessary to write inequality (5.1.8) is

obtained from a KKT point of the domain relaxation (5.1.1) or the flow conservation
relaxation (4.3.1) as follows:

Definition 5.1.11:
A dual transmission flow (µ, λ) is derived from a KKT point (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) of
domain relaxation (5.1.1) as follows: The dual values fulfill the dual flow conservation

γa µ∗a = λ+
∗ ∗
X X
µ∗a − −
v − λv
+
0 (v) 0 (v)
−
a∈δA a∈δA

for all nodes v ∈ V by Lemma 4.2.11. We extend this network flow to all arcs
a ∈ A by setting zero flow values to those arcs that are contained in A but not in
A0 . Hence we obtain a dual transmission flow (µ, λ) in (V, A) by setting µv := µ∗v
v := λv for each node v ∈ V and µa := µ∗a , λ± a := λa
±∗ ±∗
and λ± if a ∈ A0 and
µa := 0, λ±
a := 0 otherwise for each arc a ∈ A \ A .
0

A dual transmission flow (µ, λ) is derived from a KKT point (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) of

the flow conservation relaxation (4.3.1) as follows: The dual values fulfill the dual
flow conservation

γa µ∗a = (λ+ + +
∗ ∗ ∗ ∗ ∗ ∗
v − λv ) + (µv ∆v − µv ∆v )
X X
µ∗a − − − −

+
0 (v) 0 (v)
−
a∈δA a∈δA

for each node v ∈ V . Setting µv := µ∗v and λ+ + +

v := λv + µv ∆v
∗ ∗ ∗
−
and further
+
v := λv + µv ∆v for each node v ∈ V and µa := µ∗a , λ±
a := λa
∗ ∗ ∗ ∗
λ− − − ±
if a ∈ A0 and
µa := 0, λ±
a := 0 otherwise for each arc a ∈ A \ A yields a dual transmission flow
0

(µ, λ) in (V, A).

In the next lemma we derive a dual transmission flow from a KKT point of the
domain relaxation (5.1.1). We use the primal solution of this KKT point and describe
a suitable choice of ζ such that inequality (5.1.8) is violated if and only if the passive
118 Chapter 5 An Improved Benders Cut

transmission problem (4.1.1) is infeasible. For the choice of ζ we refer to the previous
Lemma 5.1.9.

Lemma 5.1.12:
Let (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) be a KKT point of the domain relaxation (5.1.1) and let (µ, λ)
be a dual transmission flow derived from this KKT point. Let ζ ∈]0, 1[ such that
the conditions of Lemma 5.1.9 are fulfilled. Then we obtain inequality (5.1.8) from
Lemma 5.1.7 with dual transmission flow (µ, λ), and parameters q ∗ , π ∗ , ζ and ∆ = 0
which is violated if and only if the passive transmission problem (4.1.1) is infeasible.
The violation (i.e., the absolute difference of the left-hand and the right-hand side of
inequality (5.1.8)) is greater than or equal to (1 − ζ) times the objective value of the
KKT point.

Proof. First we write inequality (5.1.8) of Lemma 5.1.7. It is valid for the passive
transmission problem and we only need to concentrate on the violation of this
inequality. Therefor we simply rewrite the inequality in the remainder of this proof.
Note that we do not need to refer to Theorem 4.2.9 stating the convexity of the
domain relaxation under some assumptions.

Since qa∗ realizes the nomination, we derive from

(qa∗ − (∆+
a − ∆a )) − (qa∗ − (∆+ +
a − ∆a )) − (∆v − ∆v ) = dv
X X
− − −

+
0 (v) 0 (v)
−
a∈δA a∈δA

for each node v ∈ V , that the slack variables are zero, i.e., ∆+
v = ∆v = 0 and
−
+
∆a = ∆a = 0. Then inequality (5.1.8) reduces to
−

dv πv0 (π ∗ )
X X
τa ≤ −ζ
a∈A0 v∈V

(5.1.13)
 
X 
+(1 − ζ)  λ+
v π v − λv π v + λ+
−
X −

X
a q a − λa q a − dv µv  .
v∈V a∈A0 v∈V

We recall the definition of τa = inf{fζ,q∗ ,µ (qa ) − `ζ,π∗ ,µ,λ (qa ) | q a ≤ qa ≤ q a } from

Lemma 5.1.7. By Lemma 5.1.9 the minimum of the function

fζ,q∗ ,µ (qa ) − `ζ,π∗ ,µ,λ (qa )

5.1 Valid Inequalities for the Passive Transmission Problem 119

is attained at qa∗ , which is not necessarily contained in [q a , q a ]. Hence the left-hand-

side of (5.1.13) is estimated as follows:
 
(1 − ζ)µa Φa (qa∗ ) − ζ γr,v (πv∗ − γa πw ) qa∗
X X
∗
τa ≥
a∈A0 a=(v,w)∈A0

(1 − ζ)(µv − µw − λ+
a + λa ) qa
X
− ∗
−
(5.1.14)
a=(v,w)∈A0

= (1 − ζ) µa Φa (qa∗ ) − ζ dv πv0 (π ∗ )
X X

a∈A0 v∈V

− (1 − ζ) dv µv + (1 − ζ) (λ+
a − λa )qa .
X X
− ∗

v∈V a∈A0

Combining (5.1.13) with (5.1.14) results in

(1 − ζ) µa Φa (qa∗ ) + (1 − ζ) (λ+
a − λa )qa
X X
− ∗

a∈A0 a∈A0
(5.1.15)
λ+ λ+
X  X 
≤ (1 − ζ) v π v − λv π v + (1 − ζ)
− −
a q a − λa q a .
v∈V a∈A0

In the remainder of this proof we analyze this inequality (5.1.15) and its violation.
First we prove

λ+ (λ+
X  X
µa Φa (qa∗ ) − v v = v + λv )∆v , (5.1.16)
X
− − ∗
v π v − λ π
a∈A0 v∈V v∈V

(λ+ λ+ (λ+
X 
a − λa )qa − a a = a + λa )∆a . (5.1.17)
X X
− ∗ − − ∗
q
a a − λ q
a∈A0 a∈A0 a∈A0

Equality (5.1.17) follows from Lemma 4.2.11 because the vector (q ∗ , π ∗ , ∆∗ , µ∗ ,

λ∗ ) is a KKT point of the domain relaxation (5.1.1). By this lemma it holds
λ+ +
a = λa > 0 only if qa∗ ≥ q a and λ−
a = λa > 0 only if qa∗ ≤ q a . Furthermore
∗ − ∗

∆∗a = max{0, qa∗ − q a , q a − qa∗ }, a ∈ A0 because other values imply that ∆∗ is not
optimal and hence that (q ∗ , π ∗ , ∆∗ ) is not the primal part of a KKT point.

To show equality (5.1.16) we proceed as follows. From the proof of Lemma 5.1.2
we obtain the equality

µa Φa (qa∗ ) = (λ+
v − λv )πv .
X X
− ∗

a∈A0 v∈V

By Lemma 4.2.11 it holds that λ+ +

v = λv > 0 only if πv∗ ≥ π v and λ−
v = λv > 0
∗ − ∗

only if πv∗ ≤ π v . Furthermore ∆∗v = max{0, πv∗ − π v , π v − πv∗ }, v ∈ V because other

120 Chapter 5 An Improved Benders Cut

values imply that ∆∗ is not optimal and hence that (q ∗ , π ∗ , ∆∗ ) is not a primal part
of a KKT point. This implies

(λ+ λ+ (λ+
X  X
v − λv )πv − v π v − λv π v = v + λv )∆v .
X
− ∗ − − ∗

v∈V v∈V v∈V

It follows from the reformulations (5.1.16) and (5.1.17) that (5.1.15) equivalently
rewrites to
X 
(1 − ζ) (λ+
v + λv )∆v + (λ+
a + λa )∆a ≤ 0. (5.1.18)
X
− ∗ − ∗

v∈V a∈A0

Now it follows from (4.2.8) that ∆∗v > 0 (which means πv∗ < π v or πv∗ > π v ) implies
λ+ − ∗ +
v + λv = 1 for each node v ∈ V and ∆a > 0 implies λa + λa = 1 for each arc
−

a ∈ A0 . Thus we rewrite (5.1.18) equivalently as

X 
(1 − ζ) ∆v + ∆a ≤ 0.
X
∗ ∗

v∈V a∈A0

This inequality is violated if and only if the primal solution (q ∗ , π ∗ , ∆∗ ) has positive
slack.

Next we turn to the flow conservation relaxation (4.3.1). Similar to the previous
lemma we derive a dual transmission flow from a KKT point of the relaxation (4.3.1).
We describe a suitable choice of the parameters q ∗ , π ∗ , ∆, ζ such that we obtain an
inequality by Lemma 5.1.7 which is violated if and only if the passive transmission
problem (4.1.1) is infeasible. Again we refer to Lemma 5.1.9 for the choice of ζ.

Lemma 5.1.13:
Let (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) be a KKT point of the flow conservation relaxation (4.3.1) and
let (µ, λ) be a dual transmission flow derived from this KKT point. Let ζ ∈]0, 1[ such
that the conditions of Lemma 5.1.9 are fulfilled. Then we obtain inequality (5.1.8)
from Lemma 5.1.7 with dual transmission flow (µ, λ), and parameters q ∗ , π ∗ , ∆∗ , ζ
which is violated if and only if the passive transmission problem (4.1.1) is infeasible.
The violation is greater than or equal to (1 − ζ) times the optimal objective value of
the relaxation (4.3.1).

Proof. We rewrite inequality (5.1.8) of Lemma 5.1.7. It is valid for the passive
transmission problem and we only need to concentrate on the violation of this
inequality. Therefor we simply rewrite the inequality in the remainder of this proof.
5.1 Valid Inequalities for the Passive Transmission Problem 121

Note that we do not need to refer to Theorem 4.3.6 stating the convexity of the flow
conservation relaxation under some assumptions.

We recall the definition of

τa = inf{fζ,q∗ ,µ (qa ) − `ζ,π∗ ,µ,λ (qa ) | q a ≤ qa ≤ q a }

from Lemma 5.1.7. By Lemma 5.1.9 the minimum of the function

fζ,q∗ ,µ (qa ) − `ζ,π∗ ,µ,λ (qa )

is attained at qa∗ , which is not necessarily contained in [q a , q a ]. Hence, similar as in

the previous proof of Lemma 5.1.12, a∈A0 τa rewrites as
P


(1 − ζ)µa Φa (qa∗ )
X X
τa ≥
a∈A0 a=(v,w)∈A0

− ζ (πv0 (π ∗ ) − πw
0
(π ∗ ))qa∗
− (1 − ζ)(µv − µw − λ+

a + λa )qa .
− ∗

After rearranging we obtain

µa Φa (qa∗ ) + (1 − ζ) (λ+
a − λa )qa
X X X
− ∗
τa ≥(1 − ζ)
a∈A0 a∈A0 a∈A0
 
+ζ πv0 (π ∗ ) +
X X X
dv − qa∗ qa∗
+
0 (v) 0 (v)
v∈V −
a∈δA a∈δA
  (5.1.19)
+ (1 − ζ) qa∗ +
X X X
µv dv − qa∗
+
0 (v) 0 (v)
v∈V −
a∈δA a∈δA

dv πv0 (π ∗ ) − (1 − ζ)
X X
−ζ dv µv .
v∈V v∈V

We rewrite the first three lines of this equality separately as follows:

1. From (4.3.15a) and (4.3.15b) we obtain

µa Φa (qa∗ ) = (λ+
v π v − λv π v ),
X X
−

a∈A0 v∈V

(λ+ a )qa = (λ+

a q a − λa q a ).
X X
a − λ− ∗ −

a∈A0 a∈A0
122 Chapter 5 An Improved Benders Cut

2. Setting ∆±
v := ∆v for all nodes v ∈ V and ∆±
a := ∆a for all arcs a ∈ A the
± ∗ ± 0 ∗

vector q ∗ fulfills the flow conservation

(qa∗ − (∆+
a − ∆a )) − (qa∗ − (∆+ +
a − ∆a )) − (∆v − ∆v ) = dv
X X
− − −

+
0 (v) 0 (v)
−
a∈δA a∈δA

for all nodes v ∈ V . We multiply each side by πv0 (π ∗ ), take the sum over all
nodes v ∈ V and obtain
 
πv0 (π ∗ ) + =
X X X
dv − qa∗ qa∗
+
0 (v) 0 (v)
v∈V −
a∈δA a∈δA

+ 0 +
(∆−
v − ∆v )πv (π ) + (∆−
a − ∆a )(πv (π ) − πw (π ))
X X
∗ 0 ∗ 0 ∗

v∈V a=(v,w)∈A0

From the value of ∆ from (4.3.1h), (4.3.1i), (4.3.1j) and (4.3.1k) we derive the
equalities

+ 0 + 0
(∆−
v − ∆v )πv (π ) = (∆−
v πv (π) − ∆v πv (π))
X X
∗ 0

v∈V v∈V

(∆− ∆+
a )(πv (π ) (π ∗ )) = (∆− + 0
a Φa (q a ) − ∆a Φa (q a ))
X X
0 ∗ 0 0
a − − πw
a=(v,w)∈A0 a∈A0

3. With the same definition of ∆ from the previous item and similar reasoning we
obtain (by multiplication by µv instead of πv0 (π ∗ ))
 
+
X X X
µv dv − qa∗ qa∗
+
0 (v) 0 (v)
v∈V −
a∈δA a∈δA

+ +
= (∆−
v − ∆v )µv + (∆−
a − ∆a )(µv − µw ).
X X

v∈V a=(v,w)∈A0

Using all these reformulations together we obtain from (5.1.19)

X
τa ≥
a∈A0
X  X 
(1 − ζ) λ+
v π v − λv π v + (1 − ζ)
−
λ+ −
a q a − λa q a
v∈V a∈A0
X 
+ 0 + 0
+ζ (∆−
v πv (π) − ∆v πv (π)) + (∆−
a Φa (q a ) − ∆a Φa (q a )) − dv πv0 (π ∗ )
0
X 0
X

v∈V a∈A0 v∈V

X 
+(1 − ζ) (∆− ∆+
v )µv + (∆− ∆+
a )(µv − µw ) −
X X
v − a − d v µv .
v∈V a=(v,w)∈A0 v∈V
5.1 Valid Inequalities for the Passive Transmission Problem 123

We take this inequality and substitute the left-hand side of inequality (5.1.8) which
then reduces to
X 
+ +
(1 − ζ) (∆−
v − ∆v )µv + (∆−
a − ∆a )(µv − µw ) ≤0 (5.1.20)
X

v∈V a=(v,w)∈A0

+
Now it follows from Lemma 4.3.7 that ∆−
v > 0 implies µv = µv = 1 and ∆v > 0
∗

implies µv = µ∗v = −1 for each node v ∈ V . Furthermore by Lemma 4.3.7 it holds that
+
∆−
a > 0 implies µv − µw = µv − µw = 1 and ∆a > 0 implies µv − µw = µv − µw = −1
∗ ∗ ∗ ∗

for each arc a ∈ A0 . Thus we rewrite (5.1.20) equivalently as

X 
+ +
(1 − ζ) (∆− + ∆ ) + (∆ + ∆ ) ≤0
X
−
v v a a
v∈V a∈A0

This inequality is violated if and only if the primal solution (q ∗ , π ∗ , ∆∗ ) has positive
slack.

The next corollary summarizes Lemma 5.1.12 and Lemma 5.1.13. It presents a
choice of the parameters for inequality (5.1.8) such that it is violated if and only if
the passive transmission problem (4.1.1) is infeasible. Thus the corollary enhances
Lemma 5.1.7.

Corollary 5.1.14:
Let (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) be a KKT point of relaxation (5.1.1) or (4.3.1), and let (µ, λ)
be a dual transmission flow derived from this KKT point. Furthermore let ∆ = 0
if the KKT point belongs to the domain relaxation and ∆ = ∆∗ otherwise. Let
ζ ∈]0, 1[ such that the conditions of Lemma 5.1.9 are fulfilled. Then for constants
τa := inf{fζ,q∗ ,µ (qa ) − `ζ,π∗ ,µ,λ (qa ) | q a ≤ qa ≤ q a } for each arc a ∈ A0 the inequality

(5.1.21)
X
τa ≤
a∈A0
 
+ 0 + 0
(∆−
v π v − ∆v π v ) + (∆−
a Φa (q a ) − ∆a Φa (q a )) − dv πv0 (π ∗ )
X X X
0 0
ζ
v∈V a∈A0 v∈V
 

λ+ λ+
X  X 
+(1 − ζ)  v π v − λv π v +
X
− −
a q a − λa q a − dv µv 
v∈V a∈A0 v∈V

is violated if and only if the passive transmission problem (4.1.1) is infeasible. The
violation is greater than or equal to (1 − ζ) times the optimal objective value of the
relaxations (4.2.1) or (4.3.1) respectively.
124 Chapter 5 An Improved Benders Cut

Proof. This follows from Lemma 5.1.7, Lemma 5.1.12 and Lemma 5.1.13.

We note that it is not guaranteed that ζ as assumed in Corollary 5.1.14 exists. If

it does not exist, then inequality (5.1.21) cannot be set up.

5.1.4 A Linear Inequality derived from the Lagrange Function of

the Domain Relaxation
We show that inequality (5.1.21) of the previous Corollary 5.1.14 with ζ = 0 is
motivated by the Lagrange function (5.1.2) of the domain relaxation (5.1.1). However,
a restriction of the flow directions of the passive transmission problem is necessary
for the validity of the inequality.

Lemma 5.1.15:
Let L(q, π, ∆, µ, λ) be the Lagrange function (5.1.2) of the nonlinear domain relax-
ation (5.1.1). Let (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) be a KKT point of this relaxation. Assume
that the flow directions of the passive transmission problem (4.1.1) are restricted by
qa µ∗a ≥ 0 for each arc a ∈ A0 , i.e., q a ≥ 0 if µ∗a ≥ 0 and q a ≤ 0 if µ∗a ≤ 0.
If primal and dual arc flows have the same direction, i.e., qa∗ µ∗a ≥ 0 holds for each
arc a ∈ A0 , and q a ≤ qa ≤ q a , a ∈ A0 then the inequality

inf L(q, π, ∆, µ∗ , λ∗ ) ≤ 0
q,π,∆

is equal to inequality (5.1.21) of Corollary 5.1.14 with ζ = 0 and the KKT point
(q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ).

Proof. Let (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) be a KKT point of the domain relaxation (5.1.1). We

write the Lagrange function (5.1.2):

L(q, π, ∆, µ∗ , λ∗ ) = ∆v + ∆a
X X

v∈V a∈A0

+ µ∗a Φa (qa ) − (πv − γa πw )

X

a∈A0
a=(v,w)
 

+ qa +
X X X
µ∗v dv − qa 
 
+
0 (v) 0 (v)
v∈V −
a∈δA a∈δA

λ+
X ∗

−∗
+ v (πv − ∆v − π v ) + λv (π v − πv − ∆v )
v∈V

λ+
X ∗

−∗
+ a (qa − ∆a − q a ) + λa (q a − qa − ∆a )
a∈A0
5.1 Valid Inequalities for the Passive Transmission Problem 125

λ∗v ∆v − λ∗a ∆a .
X X
−
v∈V a∈A0

By Lemma 4.2.11 the dual solution of the KKT point forms a general network flow
in (V, A0 ), i.e.,

µ∗a γa = λ+
∗ ∗
X X
µ∗a − −
v − λv ∀ v ∈ V.
+
0 (v) 0 (v)
−
a∈δA a∈δA

This implies
(λ+
∗ ∗
µ∗a (πv − γa πw ) = v πv − λv πv ).
X X
−

a∈A0 v∈V
a=(v,w)

This means that the Lagrange function L(q, π, ∆, µ∗ , λ∗ ) is independent of the values
of π. Recall the conditions (4.2.7c) and (4.2.7d) from the proof of Lemma 4.2.11
which state that the dual solution (µ∗ , λ∗ ) also fulfills

λ+
∗ ∗
v + λv + λv = 1 ∀ v ∈ V,
− ∗

λ+
∗ ∗
a + λa + λa = 1 ∀ a ∈ A .
− ∗ 0

This implies

λ+
∗ ∗
v ∆v + λv ∆v + λv ∆v = ∆v
− ∗
∀ v ∈ V,

λ+
∗ ∗
a ∆a + λa ∆a + λa ∆a = ∆a
− ∗
∀ a ∈ A0 .

Again we conclude that the Lagrange function L(q, π, ∆, µ∗ , λ∗ ) is independent of

the values of ∆. From these observations we conclude that L(q, π, ∆, µ∗ , λ∗ ) writes
as follows:

L(q, π, ∆, µ∗ , λ∗ ) = µ∗a Φa (qa )

X

a∈A0
 

+ qa +
X X X
µ∗v dv − qa 
 
+
0 (v) 0 (v)
v∈V −
a∈δA a∈δA

λ+
X ∗ ∗

−
− v π v − λv π v
v∈V

λ+
X ∗

−∗
+ a (qa − q a ) + λa (q a − qa ) .
a∈A0
126 Chapter 5 An Improved Benders Cut

We rearrange this as

µ∗a Φa (qa ) − (µ∗v − µ∗w − λ+

X ∗ ∗

L(q, π, ∆, µ∗ , λ∗ ) = a + λa ) qa
−

a∈A0

+
X
µ∗v dv
v∈V

λ+
X ∗ ∗

v π v + λv π v
−
−
v∈V

λ+
X ∗

−∗
− a q a − λa q a
.
a∈A0

We derive a dual transmission flow (µ, λ) from the KKT point (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ )

and define τa for each arc a ∈ A0 by

τa := inf µa Φa (qa ) − (µv − µw − λ+

n o
a + λa ) qa | q a ≤ qa ≤ q a , µa qa ≥ 0 .
−

Then it follows from (5.1.3a) of the KKT conditions and µ∗a qa∗ ≥ 0 and q a ≤ qa ≤ q a
that
τa = µ∗a Φa (qa∗ ) − (µ∗v − µ∗w − λ+
∗ ∗
a + λa ) qa
− ∗

holds. Then we have:

inf L(q, π, ∆, µ∗ , λ∗ ) ≤ 0
q≤q≤q,π≤π≤π,∆≥0

⇔ (5.1.22)
X  X 
λ+ πv + πv + λ+
X X
τa ≤ v λ−
v a qa − λ−
a qa − µv dv .
a∈A0 v∈V a∈A0 v∈V

This inequality equals inequality (5.1.21) of Corollary 5.1.14 for ζ = 0.

5.2 A Valid Inequality for the Topology Optimization

Problem

Corollary 5.1.14 states the main result of the previous Section 5.1. The corollary
states that inequality (5.1.21) is violated if and only if the passive transmission
problem is infeasible. This inequality contains no variables and is valid for the passive
transmission problem (4.1.1). In this section we extend this inequality such that it
is valid for the topology optimization problem (3.2.1), and not only for the passive
transmission problem. Recall that the passive transmission problem is obtained by
5.2 A Valid Inequality for the Topology Optimization Problem 127

fixing all binary variables x as we focus on y = y in this chapter. Hence x and y are
the variables of our extended inequality.

We recall the definition of the functions fζ,q∗ ,µ and `ζ,π∗ ,µ,λ and extend them to
the more general case of a flow value qa,i for an arc (a, i) ∈ AX , i > 0 of the extended
graph:

f˜ζ,q∗ ,µ (qa,i , ya,i ) := (ζ γr,v (qa,i − qa∗ ) + (1 − ζ)µa ) (αa,i qa,i |qa,i |ka − βa,i ya,i )
`˜ζ,π∗ ,µ,λ (qa,i ) := ζ (πv0 (π ∗ ) − πw
0
(π ∗ )) qa,i + (1 − ζ)(µv − µw − λ+
a + λa ) qa,i
−

For abbreviations we set, similarly as in (5.1.6),

Φa,i (q, y) := αa,i q|q|ka − βa,i y and Φ0a,i (q, y) := γr,v Φa,i (q, y)

In the following we extend Lemma 5.1.7.

Theorem 5.2.1:
Let (µ, λ) be a dual transmission flow. Let π ∗ ∈ RV , q ∗ ∈ RA , ∆±
v ≥ 0, v ∈ V and
∆±
a ≥ 0, a ∈ A be vectors such that the flow conservation

(qa∗ − (∆+
a − ∆a )) − (qa∗ − (∆+ +
a − ∆a )) − (∆v − ∆v ) = dv
X X
− − −

+
a∈δA (v) −
a∈δA (v)

is fulfilled for each node v ∈ V . Furthermore let ζ ∈ [0, 1]. Then for constants
τa,i (ya,i ) := inf{f˜ζ,q∗ ,µ (qa,i , ya,i ) − `˜ζ,π∗ ,µ,λ (qa,i ) | q a,i ≤ qa,i ≤ q a,i } for each arc
(a, i) ∈ AX , i 6= 0 and each value ya,i the inequality in binary and continuous but
fixed variables x and y

xa,i τa,i (ya,i ) ≤ −ζ dv πv0 (π ∗ ) (5.2.1)

X X

(a,i)∈AX v∈V
i6=0
 

+ 0 + 0
+ζ  (∆−
v π v − ∆v π v ) + xa,i (∆−
a Φa,i (q a,i , ya,i ) − ∆a Φa,i (q a,i , ya,i ))
X 0
X 0



v∈V (a,i)∈AX
i6=0
 
X  
+(1 − ζ)  λ+
v π v − λv π v + xa,i λ+
−
X −

X 
 a q a,i − λa q a,i − d v µv 

v∈V (a,i)∈AX v∈V
i6=0

+ζ xa,0 max (qa∗ − (∆+ +

a − ∆a ))(π v − π w ), (qa − (∆a − ∆a ))(π v − π w )
X  − 0 0 ∗ − 0 0

a=(v,w)∈A
128 Chapter 5 An Improved Benders Cut

+(1 − ζ) xa,0 max µa (π v − π w ), µa (π v − π w )

X 
a=(v,w)∈A

is valid for the topology optimization problem (3.2.1).

Remark 5.2.2:
Note that the basic difference of this inequality (5.2.1), which is valid for the topology
optimization problem (3.2.1), and inequality (5.1.8), which is valid for the passive
transmission problem (4.1.1), are the last two lines of (5.2.1). The other parts of
these inequalities almost coincide except the addition of the binary variables x and
the continuous but fixed variables y.

Proof. We are going to apply Lemma 5.1.7 that yields a valid inequality for the
passive transmission problem corresponding to the arc set A0 . We write A0 = A0 (x)
depending on the binary vector x by A0 (x) := {(a, i) ∈ AX | xa,i = 1, i > 0}. In order
to apply Lemma 5.1.7 we define ∆ variables in dependence on A0 (x) by considering
the following equality which is valid for each active configuration x:

dv = xa,0 (qa∗ − (∆+

a − ∆a )) − xa,0 (qa∗ − (∆+
a − ∆a ))
X −
X −

+
a∈δA (v) −
a∈δA (v)

+ (qa∗ − (∆+
a − ∆a )) − (qa∗ − (∆+ +
a − ∆a )) − (∆v − ∆v ).
X −
X − −

(a,i)∈δ +0 (v) (a,i)∈δ −0 (v)

A (x) A (x)

˜ + (x) ≥ 0 and ∆
We define values ∆ ˜ − (x) ≥ 0 for each node v ∈ V where at least one
v v
of both values equals zero by
 
 
˜ v (x) := max
 
+ +
∆ 0, xa,0 (qa − (∆a − ∆a )) − xa,0 (qa − (∆a − ∆a ))
−
X ∗ −
X ∗ −
,
+
 
a∈δA (v) −
(v)
 a∈δA

 
 
˜+
 
∆ v (x) := max 0, xa,0 (qa∗ − (∆+
a − ∆a )) − xa,0 (qa∗ − (∆+
a − ∆a ))
X −
X −
.
+
 
−
(v) a∈δA (v)
 a∈δA


From this we obtain

(qa∗ − (∆+
a − ∆a )) − (qa∗ − (∆+
a − ∆a ))
X X
− −

+
(a,i)∈δA0 (x) (v) (a,i)∈δA0 (x) (v) (5.2.2)
−

˜ + (x) + ∆+ ) + (∆
−(∆ ˜ − (x) + ∆− ) = dv
v v v v
5.2 A Valid Inequality for the Topology Optimization Problem 129

for every node v ∈ V . We proceed analogously for µ in order to obtain a dual

transmission flow for A0 (x). For each active configuration x it holds

xa,0 µa + µa = λ+
X X X X
−
xa,0 µa − µa − v − λv
+ +
a∈δA (v) −
a∈δA (v) (a,i)∈δA0 (x) (v) (a,i)∈δA
−
0 (x) (v)

for every node v ∈ V . We define values λ̃+

v (x) ≥ 0 and λ̃v (x) ≥ 0 for each node
−

v ∈ V with at least one of both values being equal to zero by

 

 

v (x) := max 0,
X X
λ̃− xa,0 µa −
xa,0 µa
 a∈δ+ (v)

a∈δA (v)
−


A
 
 
+
 
λ̃v (x) := max 0,
X X
xa,0 µa − xa,0 µa
a∈δ + (v)
 a∈δ− (v)
 

A A

and obtain

µa = λ̃+ +
   
v (x) + λv v (x) − λv . (5.2.3)
X X
µa − − λ̃− −

+
(a,i)∈δA 0 (x) (v) (a,i)∈δA
−
0 (x) (v)

We take the definition of the constant τa,i for each arc (a, i) = (v, w, i) ∈ AX , i 6= 0
based on Lemma 5.1.7 as follows:

τa,i (ya,i ) := inf{f˜ζ,q∗ ,µ (qa,i , ya,i ) − `˜ζ,π∗ ,µ,λ (qa,i ) | q a,i ≤ qa,i ≤ q a,i }

Now assume a fixation of the binary vector x and recall that y is fixed. Inequal-
ity (5.1.8) is valid for the corresponding passive transmission problem by Lemma 5.1.8.
Using equation (5.2.2) and (5.2.3) the following inequality is equal to (5.1.8) and
hence valid for the topology optimization problem (3.2.1):

τa,i (ya,i ) ≤
X

(a,i)∈A0 (x)
 

((∆− ˜− + ˜+
v + ∆v (x))π v − (∆v + ∆v (x))π v )
X
0 0 
ζ
v∈V
 
+ 0
+ζ  (∆−
a Φa,i (q a,i , ya,i ) − ∆a Φa,i (q a,i , ya,i )) − dv πv0 (π ∗ )
X X
0

(a,i)∈A0 (x) v∈V

130 Chapter 5 An Improved Benders Cut

 

(λ+ +
X 
+(1 − ζ)  v + λ̃v (x)) π v − (λv + λ̃v (x)) π v
− − 
v∈V
 

λ+
 
+(1 − ζ) 
X X
−
a q a,i − λa q a,i − dv µv  .
(a,i)∈A0 (x) v∈V

We use
τa,i (ya,i ) = xa,i τa,i (ya,i ).
X X

(a,i)∈A0 (x) (a,i)∈AX

i6=0

We complete the proof by rewriting and estimating the right-hand side. We write

λ+ xa,i λ+
   
a q a,i − λa q a,i =
X X
− −
a q a,i − λa q a,i
(a,i)∈A0 (x) (a,i)∈AX
i6=0

and

+ 0 + 0
a Φa,i (q a,i ) − ∆a Φa,i (q a,i )) =
(∆− xa,i (∆−
a Φa,i (q a,i ) − ∆a Φa,i (q a,i )).
X 0
X 0

(a,i)∈A0 (a,i)∈AX
i6=0

˜ + (x) and ∆
Using the previous definitions of ∆ ˜ − (x) and setting
v v

π 0

if ∆ ˜ + (x) ≥ 0,
˜ − (x) − ∆
π̃v :=
v v v

π 0v
 else,

we obtain

˜−
∆ ˜+
v (x)π v − ∆v (x)π v =
˜−
∆ ˜+
v (x) − ∆v (x) π̃v
X 0 0

X

v∈V v∈V

= xa,0 (qa∗ (∆+ − ∆−

a ))(π̃v − π̃w ) ≤
X
− a
a=(v,w)∈A

xa,0 max (qa∗ − (∆+ +

a − ∆a ))(π v − π w ), (qa − (∆a − ∆a ))(π v − π w ) .
X  − 0 0 ∗ − 0 0

a=(v,w)∈A

Similarly, using the definition of λ+

v (x) and λv (x) we derive the estimation
−

λ̃+
X 
v (x)π v − λ̃v (x)π v ≤
−

v∈V
n o
xa,0 max µa (π v − π w ), µa (π v − π w ) .
X

a=(v,w)∈A
5.2 A Valid Inequality for the Topology Optimization Problem 131

Now we turn to the domain relaxation (5.1.1) and explain how the parameters
of inequality (5.2.1) have to be set such that (5.2.1) represents the feasibility of a
passive transmission problem. We derive the dual transmission flow (µ, λ) from a
KKT point of relaxation (5.1.1) and apply Lemma 5.1.12. The binary and continuous
values x∗ and y ∗ which correspond to this relaxation are then infeasible for the linear
inequality (5.2.1) if and only if the corresponding passive transmission problem is
infeasible.

Theorem 5.2.3:
Let (q ∗ , π ∗ , ∆∗ , µ∗ , λ∗ ) be a KKT point of the domain relaxation (5.1.1) for arc set A0
and let (µ, λ) be a dual transmission flow derived from this KKT point. Denote by x∗
and y ∗ the binary and continuous values which yield the domain relaxation (5.1.1).
Let ζ ∈]0, 1[ as mentioned in Lemma 5.1.9. Then for constants τa,i = τa,i (ya,i ) defined
by τa,i (ya,i ) := inf{fζ,q∗ ,µ (qa,i , ya,i ) − `ζ,π∗ ,µ,λ (qa,i ) | q a,i ≤ qa,i ≤ q a,i } for each arc
(a, i) ∈ AX , i 6= 0 and the value ya,i = ya,i ∗
the inequality in binary variables x

xa,i τa,i (ya,i ) ≤ −ζ dv πv0 (π ∗ ) (5.2.4)

X X

(a,i)∈AX v∈V
i6=0
 
X  
+(1 − ζ)  λ+
v π v − λv π v + xa,i λ+
−
X −

X 
 a q a,i − λa q a,i − d v µv 

v∈V (a,i)∈AX v∈V
i6=0

+ζ xa,0 max qa∗ (π 0v − π 0w ), qa∗ (π 0v − π 0w )

X 
a=(v,w)∈A0

+(1 − ζ) xa,0 max µa (π v − π w ), µa (π v − π w )

X 
a=(v,w)∈A

is valid for the topology optimization problem (3.2.1). This inequality cuts off the
passive transmission problem corresponding to the arc set A0 if and only if it is
infeasible. For the corresponding decision vector x = x∗ and y = y ∗ = y = y the
violation of inequality (5.2.4) is greater than or equal to (1 − ζ) times the optimal
objective value of the domain relaxation.

Proof. We define qa∗ := 0 for all arcs a ∈ A \ A0 and obtain

qa∗ = dv .
X X
qa∗ −
+
a∈δA (v) −
a∈δA (v)

Furthermore we set ∆+ +
v := ∆v := 0 for all v ∈ V and ∆a := ∆a := 0 for all a ∈ A.
− −

Now the validity of (5.2.4) as a globally valid inequality for the topology optimization
problem (3.2.1) follows from Theorem 5.2.1.
132 Chapter 5 An Improved Benders Cut

τ (α)
1
α

Figure 5.2: Visualization of τ (α) defined by (5.2.5) for qa ∗

= 5, µa = 1, γa = 1, πv∗ − πw ∗
= 25, µv −
µw = 10, λ+ a − λa = 0, ka = 1, ζ = 0.7, γr,v = 1, β̃ = 0. The depicted point corresponds to the
−

original
Pvalue α = 1 of the passive transmission problem (4.1.1). The function τ (α) corresponds to the
term xa,i τa,i (ya,i ) for a fixed arc a ∈ A in (5.2.4), a valid inequality for the topology
(a,i)∈AX ,i6=0
optimization problem (3.2.1) by Theorem 5.2.3.

If x and y are the binary and continuous values, respectively, that correspond to
the passive transmission problem of the relaxation (5.1.1), for arc set A0 , i.e., x = x∗
and y = y ∗ , then (5.2.4) can be rewritten as (5.1.21). So the theorem follows from
the special choice of ζ and Corollary 5.1.14.

Inequality (5.2.4) forms our improved Benders cut because of the properties
stated in Theorem 5.2.3. A similar result as in Theorem 5.2.3 can be obtained for
the flow conservation relaxation (4.3.1) but we do not give any further details here.
As discussed in Section 4.6 the flow conservation relaxation turned out to be less
efficient than the domain relaxation (4.2.1). That is why we decided to use the
domain relaxation in our computations.
Theorem 5.2.3 states an inequality which represents the infeasibility of a certain
passive transmission problem. More precisely, if it is infeasible, then the inequality is
violated for the specific values of the binary and continuous variables x and y which
lead to the passive transmission problem. Let us now concentrate on the violation.
Therefor we visualize τa,i for an arc a = (v, w) ∈ A. Note that different binary
values for x correspond to different values τa,i in inequality (5.2.4). The value of τa,i
depends on αa,i and ya,i . Thus we consider τa,i (ya,i ) as a function τ (α) for fixed ya,i
defined by

τ (α) := min (ζ γr,v (q − qa∗ ) + (1 − ζ)µa ) (α q|q|ka − β̃)


q∈R
(5.2.5)
− ζ γr,v (πv∗ − γa πw
∗
) q − (1 − ζ)(µv − µw − λ+
a + λa ) q .
−

A visualization of this function is shown in Figure 5.2. From this image we conclude,
that either increasing or decreasing the diameter of a pipe reduces the violation
of inequality (5.2.1) of the previous Theorem 5.2.1. However, this is not coherent
5.3 Integration and Computational Results 133

with the technical properties of a pipeline: decreasing the diameter of a pipe means
to reduce the capacity of the network which, in general, means to increase the
infeasibility of the passive transmission problem.

5.3 Integration and Computational Results

In this chapter we focused on the topology optimization problem (3.2.1) arising from
the second type of network that we consider in this thesis. Recall that these networks
consist only of pipes, loops and valves. In this case it holds y = y for our model (3.2.1).
Let us present our solution framework for the topology optimization problem. We
solve the model (3.2.1) by SCIP as described in Section 2.2. Additionally we generate
the cut (5.2.4) which is valid for the topology optimization problem by Theorem 5.2.3
as follows: Consider a node of the branching tree where the optimal solution of
the LP relaxation yields integral values for the binary variables. If this is the case,
then we consider the corresponding passive transmission problem (4.1.1). Since
this problem might be infeasible, we use the domain relaxation (5.1.1) to solve it.
Recall that this relaxation has better solving performance than the flow conservation
relaxation (4.3.1) as discussed in the previous Chapter 4. Moreover it follows from
Lemma 4.2.10 that the domain relaxation is feasible and convex. After solving this
relaxation to global optimality by IPOpt, one of the following two cases occurs:

1. In the first case, the optimal solution has a zero objective function value. We
derive a feasible solution for the passive transmission problem and add it to
the solution pool of the solver and continue with the branching process.

2. In the second case, the optimal solution has a positive objective function value.
This means that some slack variables are nonzero. In this case we use the
KKT point computed by IPOpt, i.e., primal and dual solution values, and
try to derive inequality (5.2.4) where the parameters are chosen as described
in Theorem 5.2.3. Note that a KKT point exists by Lemma 4.2.11. It might
be impossible to generate inequality (5.2.4) if there does not exist ζ ∈]0, 1[
fulfilling the conditions of Lemma 5.1.9. Otherwise, if there exists such a value
ζ, then we take a minimal ζ and generate inequality (5.2.4), which is valid for
the topology optimization problem (3.2.1) by Theorem 5.2.3. Furthermore it is
violated for the current choice of binary values. We integrate the inequality in a
cut pool of the solver SCIP. If all binary variables are fixed, then we prune the
current node of the branch-and-bound tree and continue with the branching
process. This is feasible due to the convexity of the domain relaxation by
Lemma 4.2.10. Otherwise we continue without pruning.
134 Chapter 5 An Improved Benders Cut

Branch-and-bound, separation, and spatial branching

node of branching tree

Cut
Cut

feasible node with feasible node infeasible node

MILP solution with fixed x, y

infeasible NLP feasible NLP infeasible NLP

LP
N

feasible node
infeasible node
globally solved

Figure 5.3: Solution framework presented in Section 5.3. The topology optimization problem (3.2.1) is
solved with SCIP essentially by branch-and-bound, separation, and spatial branching, see Section 2.2.
We adapt this framework and solve globally the passive transmission problem (4.1.1) as discussed in
Section 4.5, classify the current node of the branching tree and prune it if possible. If the current passive
transmission problem is infeasible, then we generate (as described in Section 5.3) a linear inequality
in the binary and continuous but fixed variables x and y which is dynamically added to the topology
optimization problem. This inequality is violated for the binary values which lead to the considered
passive transmission problem.

We conclude that the described method is an extension of the solution method

presented in Chapter 4 based on the domain relaxation.
The algorithmic scheme of this solution approach is shown in Figure 5.3. This
figure includes the computation of inequalities compared to Figure 4.5. Recall that
Figure 4.5 shows the use of the relaxations (4.2.1) and (4.3.1) within the branch-and-
bound tree for solving the topology optimization problem (3.2.1).
We implemented the cut generation algorithm in C. In our initial implementation
we added all obtained cuts directly to the branch-and-bound process. It turned out
that this strategy was not very efficient. In the final implementation we do not add
them immediately. They are instead stored in a cut pool until a predefined number of
inequalities is reached (experimentally, a pool size of 40 inequalities turned out to be
a good value). Then we restart the branch-and-bound solution process and multiply
the cut pool size by 1.5. Additionally, we also restart if a new primal feasible solution
is found with a better objective function value compared to the current best solution.
For the restart, only the best feasible solution and the valid inequalities are kept.
When restarting the solver SCIP the cuts are provided from the very beginning
of the solution process. Hence they are available during presolve and SCIP might
further strengthen our cuts (in combination with all other model inequalities).
We compare five different strategies for solving the topology optimization problem:
5.3 Integration and Computational Results 135

1. The first strategy is to use SCIP for solving our model (3.2.1) and to enforce a
certain branching priority rule, so that SCIP first branches on binary decision
variables x. Only after all discrete variables are fixed it is allowed to perform
spatial branching on continuous variables.

2. The second strategy is similar to the first strategy. Additionally, whenever a

node of the branching tree has fixed binary variables, and a global solution
for the corresponding domain relaxation (5.1.1) is computed by IPOpt, then
we prune the current node. Note that this strategy equals strategy 3 from
Section 4.6.

3. The third strategy implements the cut generation as described above, but
restarts are not enforced. Additionally we set branching priorities according to
the first strategy and prune nodes of the branching tree as in 2.

4. The fourth strategy implements the cut generation together with restarts as
described above. Additionally we set branching priorities according to the first
strategy. Further, we prune nodes of the branching tree as in 2.

5. The fifth strategy is equal to the fourth one, but the added cut simply forbids
the current binary vector. For the current integral solution values x∗ this
inequality is
xa,i + (1 − xa,i ) ≥ 1.
X X

(a,i)∈AX (a,i)∈AX
x∗ =0 x∗ =1
a,i a,i

Hence this fifth strategy basically represents the technique of the previous
Chapter 4 using the domain relaxation (5.1.1) implemented by strategy 3 from
Section 4.6 and storing the information about infeasible passive transmission
problems when restarting.

Computational Results
For the computational study we consider the transport networks net1 and net2. For
net2 we contract all compressors and control valves. These networks have different
topologies and consist of pipelines only. All pipes (a, i) have constants αa,i > 0,
βa,i = 0 and ka,i = 1. A visualization of these networks is shown in Figure 3.5
and Figure 3.6. All extensions are parallel arcs (or loops) as described in the next
paragraph. Therefore they are not visible in the pictures. The dimensions of the
networks are given in Table 3.1. The considered networks belong to the second type
of network which we consider in this thesis. They only contain pipelines, loops and
valves.
136 Chapter 5 An Improved Benders Cut

strategy 1 2 3 4 5 all
solved instances 53 53 54 64 51 64

Table 5.1: Summary of the Tables A.6 – A.9 showing the globally solved instances out of 82 nominations
in total. All instances solved by strategy 2 are also solved by strategy 3.

(A,B) = (1,2)
solved(52) incomp.(28)
time [s] nodes gap [%]
strategy A 161.6 44,146 113
strategy B 174.3 38,194 114
shifted geom. mean +8 % −13 % +1 %

Table 5.2: Run time, number of branch-and-bound nodes and gap comparison for the strategies 1 and 2
(aggregated results). The columns “solved” contain mean values for those instances globally solved by
both strategies A and B. The columns “incomplete” show mean values for those instances having a primal
feasible solution available but were not globally solved by both strategies A and B. The underlying data
are available in Tables A.6 – A.9.

For each network we are given a balanced flow demand at the entry and the exit
nodes. There exists a feasible flow in the network for this given demand. Now we
scale up this demand, that is, we multiply each entry and exit value by the same
scalar > 1. For a certain value 2.0 the instance is no longer feasible, i.e., there is no
valid flow which fulfills all model constraints. In order to obtain a feasible flow again,
the network capacity needs to be extended, for which we introduced a number of
parallel arcs (loops). For the network net1 we consider different instances where
we allow to build up to 7 loops, up to 8 loops, continuing up to 11 loops for each
arc of the network. For the instance net2 we similarly allow between 2 and 4 loops,
respectively. That is, each original pipeline can be extended at most by this number
of pipelines having the same characteristics as the original one.
We set a time limit of 39 600 s and used the computational setup described in
Section 3.5. The computational results are shown in Tables A.6 – A.9. Especially
the number of inequalities (5.2.4) that are generated and added to the topology opti-
mization problem is available in these tables. A summary is given in Tables 5.1 – 5.3.
We use the geometric mean of run time, number of branch-and-bound nodes and gap
as described in Section 3.5.
The first and second strategy almost show the same solving performance in terms
of number of solved instances. We remark that the number of nodes is reduced by
13 %, while the solving time increases by 8 %, see Table 5.2. We conclude that the
domain relaxation presented in Chapter 4 allows to reduce the necessary number of
branch-and-bound nodes, but the computation time increases. The first and the third
5.3 Integration and Computational Results 137

(A,B) = (1,3) (A,B) = (1,4)

solved(52) incomp.(27) solved(53) incomp.(18)
time [s] nodes gap [%] time [s] nodes gap [%]
strategy A 161.6 44,146 116 180.0 49,219 159
strategy B 175.5 31,661 102 120.0 8,681 31
shifted geom. mean +9 % −28 % −12 % −33 % −82 % −81 %

Table 5.3: Run time, number of branch-and-bound nodes and gap comparison for the strategies 1 and
3 and additionally 1 and 4 (aggregated results). The columns “solved” contain mean values for those
instances globally solved by both strategies A and B. The columns “incomplete” show mean values for
those instances having a primal feasible solution available but were not globally solved by both strategies
A and B. The underlying data are available in Tables A.6 – A.9.

strategy also show almost the same solving performance in terms of number of solved
instances. The number of nodes is reduced by 28 %, while the solving time increases
by 9 %, see Table 5.3. Hence the inequalities presented in this chapter allow to reduce
the number of branch-and-bound nodes. But we conclude that the computation time
being saved due to the reduction of branch-and-bound nodes is used to generate the
inequalities. This is different when restarts are allowed as implemented by strategy 4.
A comparison of the fourth and the first strategy shows that additionally 13 % more
instances of the test set (11 out of 82) are solved to global optimality within the
time limit of 39 600 s. Furthermore the run time is reduced by 33 %, the number of
nodes by 82 % and the gap by 81 % for those instances which remain with a positive
gap value following both strategies. Again the number of nodes decreases more than
the run time. This is due to the time which is necessary for solving the domain
relaxation and generating the inequalities.
Figure 5.4 shows an aggregated performance plot of the five strategies. The
fifth strategy clearly shows the worst results. It does not even improve the first
strategy. The graphs of the third and fourth strategy are coherent with the previous
discussion. The graph of the fourth strategy dominates the other strategies for those
instances that have a solution time of more than 200 s following strategy 4. This
order is different for the other instances. Here the first strategy dominates. This
solving behavior can be clearly attributed to the restarts that we apply whenever
a primal solution is found. Figure 5.5 shows a coherent result. It is a scatterplot
of the run time comparing the second and the fourth strategy. Finally Figure 5.6
shows a scatterplot of the gap comparing both strategies. The gap is reduced for
every instance.
We do not generate the cut for the first type of network which we consider in
this thesis. Recall that they consist of pipes and valves only. In this case the binary
variables are associated with valves only. If an inequality (5.2.4) was generated for a
138 Chapter 5 An Improved Benders Cut

100
strategy 4
strategy 1
strategy 3
Number of solved instances (%) strategy 2
75
strategy 5

50

25

0
10−1 100 101 102 103 104
Run time (sec)

Figure 5.4: Performance plot for different nominations on the networks net1 and net2 (aggregated) and a
time limit of 39 600 s. The different strategies are described in Section 5.3. Strategy 1 mainly consist of
SCIP. Strategies 2 and 3 also correspond to SCIP in combination with our solution methods presented in
this chapter, i.e., generating the cuts (5.2.4) and adding them to the problem formulation. Both strategies
only differ in the restarting policy. The underlying data are available in Tables A.6 – A.9.

closed valve a, then it would be easy to see that the inequality might get inactive
when a is opened. This is due to a possible large coefficient of xa,1 . The same
observation holds for an opened valve getting closed. In this case it is due to a
possible large coefficient of xa,0 . Both cases imply that the inequality typically has a
similar characteristic as the cut used in strategy 5.

Summary

In this chapter we presented a valid inequality for the topology optimization prob-
lem (3.2.1) restricted to the case y = y. This case includes the second type of
network of our test instances which we consider in this thesis, namely those which
contain pipes, loops and valves only. A first step for the definition of the cut was the
inequality of Corollary 5.1.6 being a linear combination of two different nonlinear
inequalities. In Section 5.1.4 we explained that the first one can be obtained from the
Lagrange function of the domain relaxation (4.2.1). The second one is the so-called
pc-regularization which was necessary to derive a globally valid cut. In a second
step we gave a linear underestimator in Section 5.1.2 which then led to an inequality
without any variables representing the infeasibility of a passive transmission problem,
5.3 Integration and Computational Results 139

105

104
Run time strategy 4 (sec)
103

102

101

100

10−1

10−2

10−2 10−1 100 101 102 103 104 105

Run time strategy 2 (sec)
Figure 5.5: Run time comparison for different nominations on the networks net1 and net2. Each cross (×)
corresponds to a single instance of the test set. Note that multiple crosses are drawn in the upper right
corner of the plots that cannot be differentiated. They represent those instances that ran into the time
limit of 39 600 s for SCIP without and with using cuts. The underlying data are available in Tables A.6
– A.9.

400
Gap strategy 4 (%)

300

200

100

0 100 200 300 400

Gap strategy 2 (%)
Figure 5.6: Gap comparison for different nominations on the networks net1 and net2. Each cross (×)
corresponds to a single instance of the test set. Those instances where at least one gap value is infinity
are not depicted here. The underlying data are available in Tables A.6 – A.9.
140 Chapter 5 An Improved Benders Cut

see Section 5.1.3. In Section 5.2 we explained how to extend this inequality to a valid
cut for the topology optimization problem. Adding the inequality to the problem
is reminiscent of a Benders decomposition approach as the inequality represents
the infeasibility of a certain passive transmission problem. That is why we call the
inequality an improved Benders cut. From the computational results we conclude
that the separation of the new inequalities (5.2.4) leads to significant smaller branch-
and-bound trees and thus lower overall running times. On average the run time is
reduced by 33 % in comparison to SCIP. In addition the test set is solved to global
optimality in 13 % more instances.
Chapter 6

Sufficient Conditions for Infeasibility

of the Active Transmission Problem

In the previous Chapters 4 and 5 we focused on networks without compressor

and control valves. We considered the topology optimization problem (3.2.1) in
the case y = y. Especially in Chapter 4 we presented different relaxations of
the passive transmission problem (4.1.1). They allow to solve (4.1.1) efficiently.
The passive transmission problem occurs at nodes of the branching tree of the
topology optimization problem being solved by SCIP after branching on discrete
variables. Time consuming spatial branching can be reduced when solving these
nodes by applying the relaxations. Hence, the relaxations allow to solve the topology
optimization problem efficiently. In this chapter we turn to the more general case
of networks containing compressors and control valves, i.e., y 6= y. They belong to
the third type of network which we consider in this thesis. We present an algorithm
which allows to proceed in a similar fashion as in the restricted case of y = y, i.e.,
we solve subproblems of the topology optimization problem (3.2.1) by a specialized
method.
The subproblem considered in this chapter is the active transmission prob-
lem (3.4.2). As previously mentioned this problem arises from the topology op-
timization problem after fixing all binary variables. Our strategy for solving the
active transmission problem is as follows: We relax it in a similar way as done for the
passive transmission problem (4.1.1) as described in Chapter 4. Then we compute
a locally optimal solution for the upcoming relaxation by IPOpt. Now this may
provide a solution that is also feasible for the active transmission problem. In this
case we solved it to global optimality. However, if it does not yield a feasible solution,
then we try to verify that there exists no solution for the active transmission problem.
Therefor we present conditions for the infeasibility of the active transmission problem.

141
142 Chapter 6 Sufficient Conditions for Infeasibility

In the end, our approach consists of a primal heuristic together with infeasibility
conditions of the active transmission problem.
We include this solution method into the branch-and-bound tree which is used for
solving the topology optimization problem (3.2.1). Whenever an active transmission
problem arises, we use the heuristic to solve it. If the primal heuristic yields a feasible
solution for the topology optimization problem, then we add this solution to the
solution pool of the solver. If we can prove that the active transmission problem is
infeasible, then we prune the corresponding node of the branch-and-bound tree. This
allows to cut off nodes from the branch-and-bound tree. This in turn leads to an
increase of 20 % for globally solved topology expansion instances on the third type of
network considered in this thesis.
The outline of this chapter is as follows: In Section 6.1 we present two relaxations
of the active transmission problem which are obtained in a similar way as the domain
relaxation (4.2.1) and the flow conservation relaxation (4.3.1). We prove that both
relaxations are non-convex nonlinear optimization problems. In Section 6.2 we present
an MILP which represents the conditions of infeasibility of the active transmission
problem. Its definition depends on a local optimum of the relaxations. In Section 6.3
we present our initial motivation for the definition of the MILP. Computational
results are given in Section 6.4.

6.1 Non-Convex Relaxations for the Active

Transmission Problem

The active transmission problem considered in this chapter is obtained from the
topology optimization problem (3.2.1) by fixing all binary variables x. Recall that y
is a vector of real variables, hence fixing x means to concentrate on the subproblem of
(3.2.1) where all discrete decisions are fixed. Let A0 := {(a, i) ∈ AX : xa,i = 1, i > 0}
denote the set of arcs so that the flow is not fixed to zero. The arising problem is as
follows:

∃ q, π, p, y (6.1.1a)

s. t. αa qa |qa |ka − βa ya − (πv − γa πw ) = 0 ∀ a = (v, w) ∈ A0 , (6.1.1b)

qa = dv (6.1.1c)
X X
qa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA
6.1 Non-Convex Relaxations for the Active Transmission Problem 143

Aa (qa , pv , pw ) ≤ ba ∀ a = (v, w) ∈ A0 , (6.1.1d)

T

pv |pv | − πv = 0 ∀ v ∈ V, (6.1.1e)

πv ≤ π v ∀ v ∈ V, (6.1.1f)

πv ≥ π v ∀ v ∈ V, (6.1.1g)

qa ≤ q a ∀ a ∈ A0 , (6.1.1h)

qa ≥ q a ∀ a ∈ A0 , (6.1.1i)

ya ≤ y a ∀ a ∈ A0 , (6.1.1j)

ya ≥ y a ∀ a ∈ A0 , (6.1.1k)

pv , πv ∈ R ∀ v ∈ V, (6.1.1l)

qa , ya ∈ R ∀ a ∈ A0 . (6.1.1m)

In order to solve the active transmission problem (6.1.1) we proceed in a similar

way as for the passive transmission problem (4.1.1) described in Chapter 4. For
this we consider two different relaxations which are presented in Section 6.1.1 and
Section 6.1.2. Unfortunately it turns out that both relaxations are non-convex
optimization problems. So far we are not aware of a convex one. In the following we
assume that the graph (V, A0 ) is connected because otherwise we consider an active
transmission problem for every connected component separately.

6.1.1 Relaxation of Domains

The first relaxation we consider is similar to the domain relaxation (4.2.1) of the
passive transmission problem. It turns out to be a non-convex optimization problems
and writes as follows:

min ∆v + (∆a + k∆0a k) (6.1.2a)

X X

v∈V a∈A0

s. t. αa qa |qa |ka − βa ya − (πv − γa πw ) = 0 ∀ a = (v, w) ∈ A0 , (6.1.2b)

qa = dv (6.1.2c)
X X
qa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

Aa (qa , pv , pw ) − ∆0a ≤ ba ∀ a = (v, w) ∈ A0 , (6.1.2d)

T
144 Chapter 6 Sufficient Conditions for Infeasibility

pv |pv | − πv = 0 ∀ v ∈ V, (6.1.2e)

π v − ∆v ≤ π v ∀ v ∈ V, (6.1.2f)

πv + ∆ v ≥ π v ∀ v ∈ V, (6.1.2g)

qa − ∆a ≤ q a ∀ a ∈ A0 , (6.1.2h)

qa + ∆ a ≥ q a ∀ a ∈ A0 , (6.1.2i)

ya ≤ y a ∀ a ∈ A0 , (6.1.2j)

ya ≥ y a ∀ a ∈ A0 , (6.1.2k)

pv , πv ∈ R ∀ v ∈ V, (6.1.2l)

qa , ya ∈ R ∀ a ∈ A0 , (6.1.2m)

∆v ∈ R≥0 ∀ v ∈ V, (6.1.2n)

∆a ∈ R≥0 ∀ a ∈ A0 , (6.1.2o)

∆0a ∈ Rν≥0
a
∀ a ∈ A0 . (6.1.2p)

Lemma 6.1.1:
The optimization problem (6.1.2) is a feasible relaxation of the active transmission
problem (6.1.1).

Proof. Every feasible solution of the active transmission problem (6.1.1) can be
extended to a feasible solution of (6.1.2) by considering ∆ := 0 as slack values. Hence
(6.1.2) is a relaxation of the active transmission problem (6.1.1).
When fixing y and neglecting constraints (6.1.2d) and (6.1.2e) the optimization
problem (6.1.2) is equal to the domain relaxation (4.2.1). Contracting those arcs
a = (v, w) ∈ A0 that are modeled by πv = πw yields a modified version of the
domain relaxation with α > 0. This problem is always feasible by Theorem 4.2.9.
Transforming a solution to the original problem and removing the fixation of y keeps
the feasibility. By adding the neglected constraints (6.1.2e) the feasibility is kept
again because there are no bounds on the pressure variables p. The same holds for
constraints (6.1.2d) because the values ∆0a can be chosen such that (6.1.2d) is valid
for any values qa , pv , pw for each arc a = (v, w) ∈ A0 .

Let us consider a local optimal solution of this problem (6.1.2). In the case that
the objective value of this solution equals zero, we derive an optimal solution for
the active transmission problem by neglecting the slack variables ∆. Otherwise,
6.1 Non-Convex Relaxations for the Active Transmission Problem 145

αa9 = 20

αa8 = 5
βa4 = 1
αa1 = 40 αa2 = 5/6 αa3 = 20 αa4 = 1 αa5 = 40 αa6 = 5/6

αa7 = 10
αa10 = 10
dv1 = −2.5

dv2 = −6.5

dv5 = 1.5

dv6 = 6.5

dv7 = −7
dv3 = 7

d v4 = 1

Figure 6.1: An example of a planar network. It is used in the discussion in Example 6.1.2 for showing that
the domain relaxation (6.1.2) is a non-convex optimization problem. The dashed arc is a control valve
with constraint αa4 qa4 |qa4 | − ya4 = πv4 − πv5 and −50 ≤ ya4 ≤ 0. All the other arcs a = (v, w) ∈ A0
are passive pipes, i.e., αa qa |qa | = πv − πw . All primal node potential values are bounded by 0 and 100,
except v3 and v7 , where upper and lower bound are π v7 = 29 and π v = 91, respectively. The absolute
3
value of the arc flow is bounded by 100.

i.e., if the objective value is greater than zero, then it is not guaranteed that the
local optimal solution is a global one. Unfortunately it turns out, as discussed in
Example 6.1.2, that the relaxation (6.1.2) is a non-convex optimization problem. So
we cannot conclude that the active transmission problem (6.1.1) is infeasible in this
case.

Example 6.1.2:
The domain relaxation (6.1.2) for the network shown in Figure 6.1 has two different
KKT points. The network is planar and consists of 7 nodes and 10 arcs (9 pipes
and 1 active control valve). The parameters of the active transmission problem are
available from Figure 6.1. Those that are not shown are defined as follows:


1

if a = a4 ,
βa := ∀ a ∈ A0 ,
0
 else

γa := 1 ∀ a ∈ A0 ,
ka := 1 ∀ a ∈ A0 ,
146 Chapter 6 Sufficient Conditions for Infeasibility


29

if v = v7 ,
π v := ∀ v ∈ V,
100 else


91 if v = v ,

3
π v := ∀ v ∈ V,
0
 else

q a := 100 ∀ a ∈ A0 ,
q a := −100 ∀ a ∈ A0 ,
y a := 0 ∀ a ∈ A0 ,

−50 if a = a4 ,

y a := ∀ a ∈ A0 .
0
 else

There are no additional constraints for the arcs, i.e., Aa = 0, ba = 0 for every arc
a ∈ A0 . Hence we can neglect the coupling constraints (6.1.1e) and the pressure
variables p and constraints (6.1.2d).

In this context the domain relaxation (6.1.2) is as follows:

min ∆v + ∆a (6.1.3)
X X

v∈V a∈A0

s. t. qa = dv
X X
qa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

αa qa |qa | − (πv − πw ) = 0 ∀ a = (v, w) ∈ A0 , a 6= a4 ,

αa4 qa4 |qa4 | − ya4 − (πv4 − πv5 ) = 0,

qa − ∆a ≤ 100 ∀ a ∈ A0 ,

qa + ∆a ≥ −100 ∀ a ∈ A0 ,

πv − ∆v ≤ 100, ∀ v ∈ V, v 6= v7

πv + ∆v ≥ 0, ∀ v ∈ V, v 6= v3

πv7 − ∆v7 ≤ 29,

πv3 + ∆v3 ≥ 91,

0 ≤ ya ≤ 0, ∀ a ∈ A0 , a 6= a4
6.1 Non-Convex Relaxations for the Active Transmission Problem 147

−50 ≤ ya4 ≤ 0,

πv ∈ R ∀ v ∈ V,

qa , ya ∈ R ∀ a ∈ A0 ,

∆v ∈ R≥0 ∀ v ∈ V,

∆a ∈ R≥0 ∀ a ∈ A0 .

In order to show that there exist different KKT points of (6.1.3) we consider the
KKT conditions for this problem. The objective and all constraints of (6.1.3) are
continuously differentiable. We write the Lagrange function of problem (6.1.3) in
order to derive the KKT conditions for (6.1.3):

L(q, π, y, ∆, µ, λ) = ∆v + ∆a
X X

v∈V a∈A0
 
+ µa αa qa |qa |ka − βa ya − (πv − πw )
X

a=(v,w)∈A0
 

+ qa +
X X X
µv dv − qa 
 
+
0 (v) 0 (v)
v∈V −
a∈δA a∈δA

λ+
X 
+ v (πv − ∆v − π v ) + λv (π v − πv − ∆v )
−

v∈V

λ̃+
X 
+ a (qa − ∆a − q a ) + λ̃a (q a − qa − ∆a )
−

a∈A0

λ+
X 
+ a (y a − y a ) + λ −
a (y a
− y a )
a∈A0

λ v ∆v − λ̃a ∆a .
X X
−
v∈V a∈A0

We obtain from the complementarity constraints (2.4.2a) of the KKT conditions

(2.4.2) of the domain relaxation (6.1.2) that a KKT point (q ∗ , π ∗ , y ∗ , ∆∗ , µ∗ , λ∗ ) of
(6.1.3) is feasible for

∂L
= 0 ⇒ µa 2αa |qa | + λ̃+
a − λ̃a = µv − µw
−
∀ a = (v, w) ∈ A0 ,
∂qa
∂L
=0⇒ µa = λ+
X X
−
µa − v − λv ∀ v ∈ V,
∂πv +
a∈δA0 (v) −
a∈δA0 (v)
148 Chapter 6 Sufficient Conditions for Infeasibility

a1 a2 a3 a4 a5 a6 a7 a8 a9 a10
qa 0.5 6 1 0 0.5 6 1 2 1 1
µa −1/3 −2/3 −1/3 0 −1/3 −2/3 1/3 −1/3 −1/3 1/3
λ̃a 1 1 1 1 1 1 1 1 1 1
v1 v2 v3 v4 v5 v6 v7
πv 50 60 90 70 70 60 30
∆v 0 0 1 0 0 0 1
µv 80/3 40/3 20/3 60/3 60/3 100/3 120/3
λ+
v 0 0 0 0 0 0 1
λ−
v 0 0 1 0 0 0 0
λv 1 1 0 1 1 1 0

Table 6.1: A KKT point of the domain relaxation (6.1.2) with positive objective value. Those variables
not depicted have zero values, i.e., ya , ∆a , λ±
a , λ̃a = 0, a ∈ A. The KKT point is used for a discussion in
±

Example 6.1.2.

a1 a2 a3 a4 a5 a6 a7 a8 a9 a10
qa 0.34 6.38 0.61 -0.87 0.16 5.91 0.45 2.49 1.08 0.75
ya 0 0 0 -25 0 0 0 0 0 0
λ̃a 1 1 1 1 1 1 1 1 1 1
v1 v2 v3 v4 v5 v6 v7
πv 52.46 57.13 91.08 83.47 59.23 58.11 28.95
λv 1 1 1 1 1 1 1

Table 6.2: A KKT point of the domain relaxation (6.1.2) with zero objective value. All numbers are
depicted up to a precision of 2 digits. Those variables not depicted have zero values, i.e., µa , ∆a , λ±
a , λ̃a =
±

0, a ∈ A and ∆v , µv , λ±
v = 0, v ∈ V . The KKT point is used for a discussion in Example 6.1.2.

∂L
=0⇒ λ+
v + λv + λv = 1
−
∀ v ∈ V, (6.1.4)
∂∆v
∂L
=0⇒ λ̃+
a + λ̃a + λ̃a = 1
−
∀ a ∈ A0 ,
∂∆a
∂L
= 0 ⇒λ+
a − λa − βa µa = 0
−
∀ a ∈ A0 .
∂ya

Similar to Lemma 4.2.11 and the discussion in Section 4.2.3 we observe that the
Lagrange multipliers (µ∗a )a∈A0 of a KKT point (q ∗ , π ∗ , y ∗ , ∆∗ , µ∗ , λ∗ ) of (6.1.3) form
a network flow in (V, A0 ).
Now it is easy to see that the solution values shown in Table 6.1 and Table 6.2
fulfill the KKT conditions (2.4.2) of (6.1.3), i.e., they are feasible for (6.1.3), (6.1.4)
and the complementarity constraints (2.4.2e), and hence are KKT points of the
6.1 Non-Convex Relaxations for the Active Transmission Problem 149

domain relaxation (6.1.3). The primal part shown in Table 6.2 has zero objective
value while this is not the case for the primal part shown in Table 6.1. A convex
combination of both primal feasible solutions is not feasible for the relaxation (6.1.3)
because of constraint (6.1.2b). We conclude that (6.1.2) is a non-convex optimization
problem. As both solutions have different objective values this shows that a nonlinear
solver like IPOpt, which computes KKT points, cannot guarantee to compute the
global optimal solution of (6.1.2).
150 Chapter 6 Sufficient Conditions for Infeasibility

6.1.2 Relaxation of Flow Conservation Constraints

Unfortunately the domain relaxation (6.1.3) presented in Section 6.1.1 is a non-convex

optimization problem. Now we try to find another relaxation of the active transmission
problem (6.1.1). Therefor we extend the flow conservation relaxation (4.3.1) presented
in Chapter 4 to the more general case of this chapter. Again it turns out that the
obtained relaxation is non-convex. It is as follows:

∆+ +
X  X 
min v + ∆ −
v + ∆ a + ∆ −
a + k∆ 0
a k (6.1.5)
v∈V a∈A0

s. t. αa qa |qa |ka − βa ya − (πv − γa πw ) = 0 ∀ a = (v, w) ∈ A0 ,

(qa − (∆+
a − ∆a ))
X
−

+
a∈δA0 (v)

(qa − (∆+ +
a − ∆a )) − (∆v − ∆v ) = dv
X
− −
− ∀ v ∈ V,
0 (v)
−
a∈δA

Aa (qa,i , pv , pw ) − ∆0a ≤ ba ∀ a = (v, w) ∈ A0 ,

T

pv |pv | − πv = 0 ∀ v ∈ V,

πv ≤ π v ∀ v ∈ V,

πv ≥ π v ∀ v ∈ V,

qa ≤ q a ∀ a ∈ A0 ,

qa ≥ q a ∀ a ∈ A0 ,

ya ≤ y a ∀ a ∈ A0 ,

ya ≥ y a ∀ a ∈ A0 ,

∆−
v (π v − πv ) ≤ 0 ∀ v ∈ V,

∆+
v (πv − π v ) ≤ 0 ∀ v ∈ V,

∆−
a (q a − qa ) ≤ 0 ∀ a ∈ A0 ,

∆+
a (qa − q a ) ≤ 0 ∀ a ∈ A0 ,

pv , πv ∈ R ∀ v ∈ V,

qa ∈ R ∀ a ∈ A0 ,
6.1 Non-Convex Relaxations for the Active Transmission Problem 151

∆±
v ∈ R≥0 ∀ v ∈ V,

∆±
a ∈ R≥0 ∀ a ∈ A0 ,

∆0a ∈ Rν≥0
a
∀ a ∈ A0 .

Lemma 6.1.3:
The optimization problem (6.1.5) is a relaxation of the active transmission prob-
lem (6.1.1).

Proof. Every feasible solution of the active transmission problem (6.1.1) yields a
feasible solution of (6.1.5) in combination with ∆ := 0.

In the next example we show that this relaxation is a non-convex optimization

problem. Therefore, from a local optimal solution (6.1.5) with positive slack we
cannot conclude that the active transmission problem (6.1.1) is infeasible.

Example 6.1.4:
We consider the network shown in Figure 6.2. It is similar to the example network in
Example 6.1.2 but we consider a different nomination and different node potential
bounds here. More precisely the differences are obtained by changing the node potential
bounds in v3 and v7 to π v3 = 90 and π v7 = 30. Furthermore the considered nomination
is adapted by setting dv3 := 7.1 and dv7 := −7.1. We neglect the coupling constraints
pv |pv | = πv , v ∈ V and the pressure variables p. Recall that it holds Aa = 0, ba = 0.
Then relaxation (6.1.5) writes as follows:

min (∆+
v + ∆v ) + (∆+
a + ∆a ) (6.1.6)
X X
− −

v∈V a∈A0

s. t.
(qa − (∆+
a − ∆a ))
X
−

+
a∈δA0 (v)

(qa − (∆+ +
a − ∆a )) − (∆v − ∆v ) = dv
X
− −
− ∀ v ∈ V,
0 (v)
−
a∈δA

∀ a = (v, w) ∈ A0
αa qa |qa | − (πv − πw ) = 0
a 6= a4 ,

αa4 qa4 |qa4 | − ya4 − (πv4 − πv5 ) = 0,

πv ≤ 100, ∀ v ∈ V, v 6= v7
152 Chapter 6 Sufficient Conditions for Infeasibility

αa9 = 20

αa8 = 5
βa4 = 1
αa1 = 40 αa2 = 5/6 αa3 = 20 αa4 = 1 αa5 = 40 αa6 = 5/6

αa7 = 10
αa10 = 10
dv1 = −2.5

dv2 = −6.5

dv7 = −6.9
dv3 = 6.9

dv5 = 1.5

dv6 = 6.5
d v4 = 1

Figure 6.2: An example of a planar network. It is used in the discussion in Example 6.1.4 for showing
that the flow conservation relaxation (6.1.5) is a non-convex optimization problem. The dashed arc is a
control valve with constraint αa4 qa4 |qa4 | − ya4 = πv4 − πv5 and −50 ≤ ya4 ≤ 0. All the other arcs
a = (v, w) ∈ A0 are passive pipes, i.e., αa qa |qa | = πv − πw . All primal node potential values are bounded
by 0 and 100, except v3 and v7 , where upper and lower bound are π v7 = 30 and π v = 90, respectively.
3
The absolute value of the arc flow is bounded by 100. The network is equal to the one shown in Figure 6.1,
but the nomination and node potential bounds differ.

πv ≥ 0, ∀ v ∈ V, v 6= v3

πv3 ≥ 90,

πv7 ≤ 30,

0 ≤ ya ≤ 0, ∀ a ∈ A0 , a 6= a4

−50 ≤ ya4 ≤ 0,

∆−
v7 (30 − πv7 ) ≤ 0,

∆+
v3 (πv3 − 90) ≤ 0,

∆−
v (100 − πv ) ≤ 0 ∀ v ∈ V, v 6= v7

∆+
v (πv − 0) ≤ 0 ∀ v ∈ V, v 6= v3

∆−
a (100 − qa ) ≤ 0 ∀ a ∈ A0 ,

∆+
a (qa + 100) ≤ 0 ∀ a ∈ A0 ,

πv ∈ R ∀ v ∈ V,
6.1 Non-Convex Relaxations for the Active Transmission Problem 153

qa , ya ∈ R ∀ a ∈ A0 ,

∆+
v , ∆v ∈ R≥0
−
∀ v ∈ V,

∆+
a , ∆a ∈ R≥0
−
∀ a ∈ A0 .

In order to show that there exist different KKT points of (6.1.6) we consider the
KKT conditions for this problem. The objective and all constraints of (6.1.6) are
continuously differentiable. We write the Lagrange function of problem (6.1.6) in
order to derive the KKT conditions for (6.1.6):

∆+ ∆+
X  X 
L(q, π, y, ∆, µ, λ) = v + ∆v
−
+ a + ∆a
−

v∈V a∈A0

+ µa αa qa |qa | − βa ya − (πv − πw )
X

a=(v,w)∈A0


+ µv dv + (∆+
v − ∆v )
X  −

v∈V

(qa − ∆+
a + ∆a )
X
−
−
+
a∈δA0 (v)


+ (qa − ∆+
a + ∆a )
X
− 

0 (v)
−
a∈δA

λ+
X 
+ v (π v − π v ) + λ −
v (π v − π v )
v∈V

λ+
X 
+ a (qa − q a ) + λa (q a − qa )
−

a∈A0

λ̂+
X 
+ a (ya − y a ) + λ̂a (y a − ya )
−

a∈A0

µ+ +
X 
+ v (π v − πv ) ∆v + µv (πv − π v ) ∆v
− −

v∈V

µ+ +
X 
+ a (q a − qa ) ∆a + µa (qa − q a ) ∆a
− −

a∈A0

λ̃+ + + +
X  X 
− v ∆ v + λ̃ v ∆
− −
v − λ̃a ∆ a + λ̃a ∆
− −
a .
v∈V a∈A0

We obtain from (2.4.2a) of the KKT conditions (2.4.2) of the flow conservation
relaxation (6.1.5) that a KKT point (q ∗ , π ∗ , y ∗ , ∆∗ , µ∗ , λ∗ ) of (6.1.6) is feasible for
154 Chapter 6 Sufficient Conditions for Infeasibility

.
a1 a2 a3 a4 a5 a6 a7 a8 a9 a10
qa 0.5 6 1 0 0.5 6 1 2 1 1
µa −1/50 −2/50 −1/50 0 −1/50 −2/50 1/50 −1/50 −1/50 1/50
λ̃+
a 1/5 3/5 1/5 1 1/5 3/5 7/5 3/5 1/5 7/5
λ̃−
a 9/5 7/5 9/5 1 9/5 7/5 3/5 7/5 9/5 3/5
v1 v2 v3 v4 v5 v6 v7
πv 50 60 90 70 70 60 30
∆+v 0 0 0.1 0 0 0 0
∆−v 0 0 0 0 0 0 0.1
µv 1/5 −3/5 −1 −1/5 −1/5 3/5 1
λ+v 0 0 0 0 0 0 3/50
λ−v 0 0 3/50 0 0 0 0
λ̃+
v 6/5 2/5 0 4/5 4/5 8/5 2
λ̃−
v 4/5 8/5 2 6/5 6/5 2/5 0

Table 6.3: A KKT point of the flow conservation relaxation (6.1.5) with positive objective value. Those
variables not depicted have zero values, i.e., ya , ∆±
a , λa , µa = 0, a ∈ A and µv = 0, v ∈ V . The KKT
± ± ±

point is used for a discussion in Example 6.1.4.

∂L
=0⇒ µa 2αa |qa | + λ+
a − λa = µv − µw
−
∀ a = (v, w) ∈ A0 ,
∂qa
∂L
=0⇒ −(µ+ − +
v ∆v − µv ∆v )
−
∂πv
+ µa = λ+
X X
−
µa − v − λv ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

∂L +
=0⇒ v (πv − π v ) − λ̃v = −1
µv + µ− ∀ v ∈ V,
∂∆+v

∂L
=0⇒ −µv + µ+
v (π v − πv ) − λ̃v = −1
−
∀ v ∈ V,
∂∆−v

∂L +
= 0 ⇒ µv − µw + µ−
a (qa − q a ) − λ̃a = −1 ∀ a = (v, w) ∈ A0 ,
∂∆+a

∂L
= 0 ⇒ µw − µv + µ+
a (q a − qa ) − λ̃a = −1
−
∀ a = (v, w) ∈ A0 ,
∂∆−a

∂L
=0⇒ λ̂+
a − λ̂a − βa µa = 0.
−
∀ a ∈ A0 .
∂ya

Now it is easy to see that the solution values shown in Table 6.3 and Table 6.4 fulfill
the KKT conditions (2.4.2), i.e., they form a feasible solution for (6.1.6), (2.4.2a) as
stated above for the considered problem and the complementarity constraints (2.4.2e),
and hence are KKT points of the flow conservation relaxation (6.1.6). The primal part
6.2 Detecting Infeasibility of the Active Transmission Problem by MILP 155

a1 a2 a3 a4 a5 a6 a7 a8 a9 a10
qa 0.18 7.11 -0.21 -2.20 -0.26 5.76 -0.43 2.98 1.13 0.46
ya 0 0 0 -50 0 0 0 0 0 0
λ̃+
a 1 1 1 1 1 1 1 1 1 1
λ̃−
a 1 1 1 1 1 1 1 1 1 1
v1 v2 v3 v4 v5 v6 v7
πv 49.71 51.02 93.26 94.23 49.09 51.88 24.15
λ̃+
v 1 1 1 1 1 1 1
λ̃−
v 1 1 1 1 1 1 1

Table 6.4: A KKT point of the flow conservation relaxation (6.1.5) with zero objective value. All num-
bers are depicted up to a precision of 2 digits. Those variables not depicted have zero values, i.e.,
µa , ∆±
a , λa , µa = 0, a ∈ A and ∆v , µv , µv , λv = 0, v ∈ V . The KKT point is used for a discussion in
± ± ± ± ±

Example 6.1.4.

shown in Table 6.4 has zero objective value while this is not the case for the primal
part shown in Table 6.3. A convex combination of both primal feasible solutions is
not feasible for the relaxation (6.1.6) because of the constraints

αa qa |qa |ka − βa ya − (πv − γa πw ) = 0 ∀ a = (v, w) ∈ A0 .

We conclude that (6.1.5) is a non-convex optimization problem. As both solutions

have different objective values this shows that a nonlinear solver like IPOpt, which
computes KKT points, cannot guarantee to compute the global optimal solution of
(6.1.5).

6.2 Detecting Infeasibility of the Active Transmission

Problem by MILP

In the previous section we presented two relaxations (6.1.2) and (6.1.5) of the active
transmission problem (6.1.1), which are non-convex optimization problems. Let us
˜ ỹ) of these relaxations such that (q̃, π̃, p̃, ỹ) is
consider a feasible solution (q̃, π̃, p̃, ∆,
not feasible for the active transmission problem. In this section we are going to
present an MILP for evaluating the infeasibility of the active transmission problem.
This MILP is called infeasibility detection MILP. Its definition depends on the vector
(q̃, π̃, p̃, ỹ) and is stated in Definition 6.2.1. The main Theorem 6.2.2 of this section
states that if the infeasibility detection MILP is infeasible or has optimal objective
value zero, then the active transmission problem (6.1.1) is infeasible.
156 Chapter 6 Sufficient Conditions for Infeasibility

For a motivation of the definition of the infeasibility detection MILP we consider

a feasible solution (q ∗ , π ∗ , p∗ , y ∗ ) for the active transmission problem (6.1.1). From
constraint (6.1.1b) and (6.1.2b) and the corresponding constraint in (6.1.5) we obtain

αa qa∗ |qa∗ |ka − βa ya∗ − (πv∗ − γa πw

∗
) = 0,
(6.2.1)
αa q̃a |q̃a |ka − βa ỹa − (π̃v − γa π̃w ) = 0,

for every arc a = (v, w) ∈ A0 . Throughout this section we use γr,v from Definition 4.2.3
and the function πv0 (π) = γr,v πv defined in (4.2.4) which allows the relation (4.2.5),
i.e.,
πv0 (π) − πw
0
(π) = γr,v (πv − γa πw ).

Using π 0 we derive from (6.2.1):

(πv0 (π ∗ ) − πv0 (π̃)) − (πw

0
(π ∗ ) − πw
0
(π̃)) + γr,v βa (ya∗ − ỹa )
= γr,v (πv∗ − γa πw
∗
+ βa ya∗ − (π̃v − γa π̃w + βa ỹa ))
= γr,v αa (qa∗ |qa∗ |ka − q̃a |q̃a |ka )

> 0 if α (q ∗ − q̃ ) > 0,


 a a a

= 0 if αa (qa − q̃a ) = 0,
∗

< 0 if α (q ∗ − q̃ ) < 0,



a a a

for every arc a = (v, w) ∈ A0 . We write this inequality for short as


> 0 if αa (qa∗ − q̃a ) > 0,





sv − sw + sa = 0 if αa (qa∗ − q̃a ) = 0, (6.2.2)

< 0 if α (q ∗ − q̃ ) < 0.


 a a a

Here sa corresponds to γr,v βa (ya∗ − ỹa ) and sv to πv0 (π ∗ ) − πv0 (π̃).

For the node potential π 0 we observe after identifying xv with πv0 (π ∗ ) − πv0 (π̃) the
conditions

xv ∈ R if π v < π̃v < π v ,

xv ≤ 0 if π̃v = π v ,
xv < 0 if π̃v > π v , (6.2.3)
xv ≥ 0 if π v = π̃v ,
xv > 0 if π v > π̃v .
6.2 Detecting Infeasibility of the Active Transmission Problem by MILP 157

We write these conditions (6.2.3) as a single constraint

x+
v − xv − xv − κv z = 0
−
with 0 ≤ x+ +
v ≤ xv , 0 ≤ xv ≤ xv , z > 0
− −
(6.2.4)

where the variable bounds and κ are defined as


0− if π̃ ≥ π ,

+ v v
xv :=
∞ else,


0

if π̃v ≤ π v ,
x−
v :=
∞ else,


1 if π̃v > π v ,




κv := −1 if π̃v < π v ,

0 else.




The flow conservation constraint (6.1.1c), constraint (6.2.4) and (in)equality

(6.2.2) form the main part of the infeasibility detection MILP. Keeping this idea in
mind, the MILP, which contains indicator constraints, is defined as follows:

Definition 6.2.1:
Let (q̃, π̃, p̃, ỹ) be a solution of the active transmission problem (6.1.1) fulfilling at least
constraint (6.1.2b) and (6.1.2e). The infeasibility detection MILP is defined as
follows:

max z (6.2.5a)

s. t. qa = dv (6.2.5b)
X X
qa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

x+
v − xv − xv − κv z = 0
−
∀ v ∈ V, (6.2.5c)
κ̃a (qa − q̃a ) > 0 ⇒ xv − xw + xa ≥ 0 ∀ a = (v, w) ∈ A ,
0
(6.2.5d)
κ̃a (qa − q̃a ) = 0 ⇒ xv − xw + xa = 0 ∀ a = (v, w) ∈ A0 , (6.2.5e)
κ̃a (qa − q̃a ) < 0 ⇒ xv − xw + xa ≤ 0 ∀ a = (v, w) ∈ A ,
0
(6.2.5f)
αa (qa − q̃a ) > 0 ⇒ sv − sw + sa ≥ κa z ∀ a = (v, w) ∈ A0 , (6.2.5g)
αa (qa − q̃a ) = 0 ⇒ sv − sw + sa = 0 ∀ a = (v, w) ∈ A ,
0
(6.2.5h)
αa (qa − q̃a ) < 0 ⇒ sv − sw + sa ≤ κa z ∀ a = (v, w) ∈ A0 , (6.2.5i)
158 Chapter 6 Sufficient Conditions for Infeasibility

q a ≤ qa ≤ q a ∀ a ∈ A0 , (6.2.5j)
sa ≤ sa ≤ sa ∀a ∈ A , 0
(6.2.5k)
xa ≤ xa ≤ xa ∀ a ∈ A0 , (6.2.5l)
x+
v ≤ x+
v ∀ v ∈ V, (6.2.5m)
x− −
v ≤ xv ∀ v ∈ V, (6.2.5n)
xv , sv ∈ R ∀ v ∈ V, (6.2.5o)
x+ −
v , xv ∈ R≥0 ∀ v ∈ V, (6.2.5p)
xa , sa , qa ∈ R ∀a ∈ A , 0
(6.2.5q)
z ∈ R≥0 . (6.2.5r)

For simplicity we do not state this problem at once as a mixed-integer nonlinear

optimization problem, but give an equivalent reformulation in Remark 6.2.3.

Constraint (6.2.5c) originates from (6.2.4) by expressing z > 0 as objective.

Similarly (6.2.5g)–(6.2.5i) originate from (6.2.2). Constraints (6.2.5d)–(6.2.5f) form
a weaker version of (6.2.2).

For the definition of this MILP we make use of different constants which are
defined below. Especially the bounds on sa for an arc a = (v, w) ∈ A0 originate from
the previously described relation that sa corresponds to γr,v βa (ya∗ − ỹa ):

0−

if βa ỹa = max{βa y a , βa y a }, ∀ a ∈ A0 ,
sa := (6.2.6a)
∞
 else a = (v, w),


0

if βa ỹa = min{βa y a , βa y a }, ∀ a ∈ A0 ,
sa := (6.2.6b)
−∞
 else a = (v, w),


0− if βa ỹa = max{βa y a , βa y a },





0− if ∃k : [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k ,




∀ a ∈ A0 ,


xa := (Aa )(k,1) ≥ 0, (Aa )(k,2) < 0, (6.2.6c)
 a = (v, w),
(Aa )(k,3) > 0,







∞− else


6.2 Detecting Infeasibility of the Active Transmission Problem by MILP 159


0 if βa ỹa = min{βa y a , βa y a },





0 if ∃k : [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k ,




∀ a ∈ A0 ,


xa := (Aa )(k,1) ≤ 0, (Aa )(k,2) > 0, (6.2.6d)
 a = (v, w),
(Aa )(k,3) < 0,







−∞ else




0−

if π̃v ≥ π v ,
x+
v := ∀ v ∈ V, (6.2.6e)
∞−
 else


0

if π̃v ≤ π v ,
v :=
x− ∀ v ∈ V, (6.2.6f)
∞−
 else


1 if π̃v > π v ,





κv := −1 if π̃v < π v , ∀ v ∈ V, (6.2.6g)

00−

else




−1 if αa > 0, q̃a > q a ,





κa := 1 if αa > 0, q̃a < q a , ∀ a ∈ A0 , (6.2.6h)

00−

else




1− if αa > 0,




1 if ∃k : [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k ,


κ̃a := ∀ a ∈ A0 . (6.2.6i)


 (Aa )(k,1) 6= 0,

00−

else



Theorem 6.2.2:
Let (q̃, π̃, p̃, ỹ) be a solution for the active transmission problem (6.1.1) fulfilling at
least constraint (6.1.2b) and (6.1.2e). If the infeasibility detection MILP (6.2.5) is
infeasible or has optimal objective value zero, then the active transmission prob-
lem (6.1.1) is infeasible.

Proof. Let (q̃, π̃, p̃, ỹ) be a solution for the active transmission problem (6.1.1) fulfill-
ing at least constraint (6.1.2b) and (6.1.2e). If the infeasibility detection MILP (6.2.5),
160 Chapter 6 Sufficient Conditions for Infeasibility

which depends on (q̃, π̃, p̃, ỹ), is infeasible, then there does not exist a flow vector
q 0 ∈ RA which fulfills
0

qa = dv (6.2.7)
X X
qa − ∀ v ∈ V, q a ≤ qa ≤ q a ∀ a ∈ A0 .
+
0 (v) 0 (v)
−
a∈δA a∈δA

This can be seen as follows: Otherwise, if there exists a vector q 0 fulfilling (6.2.7),
then (q 0 , 0) is a feasible solution for (6.2.5). We conclude that the active transmission
problem (6.1.1) is infeasible if MIP (6.2.5) is infeasible.

Now assume that the MIP (6.2.5) has an optimal solution with objective value
zero. We prove that this implies that the active transmission problem (6.1.1) is
infeasible. Therefor we assume that the active transmission problem has a feasible
solution (q ∗ , π ∗ , p∗ , y ∗ ) and show that there exists a feasible solution (q ∗ , x∗ , s∗ , z ∗ )
to MIP (6.2.5) with positive objective, i.e., z ∗ > 0. In the following we describe how
this solution (q ∗ , x∗ , s∗ , z ∗ ) is defined. First we give the definition of s∗ and z ∗ and
show that (6.2.5g)–(6.2.5i) and (6.2.5k) are fulfilled. Then we turn to the definition
of x∗ and prove that (6.2.5c)–(6.2.5f) and (6.2.5l)–(6.2.5n) are fulfilled. As the flow
vector q ∗ is feasible for the flow conservation (6.2.5b) and the bound constraints
(6.2.5j) we conclude that (q ∗ , x∗ , s∗ , z ∗ ) is feasible for MIP (6.2.5).
Recall that r was used for the definition of γr,v for every node v ∈ V . The vector
(s∗ , z ∗ ) is defined as follows:

s∗v := πv0 (π ∗ ) − πv0 (π̃) ∀ v ∈ V,

s∗a := γr,v βa (ya∗ − ỹa ) ∀ a = (v, w) ∈ A0 ,
z ∗ := min 1, min |γr,v αa (qa∗ |qa∗ |ka − q̃a |q̃a |ka )| a = (v, w) ∈ A0 : αa qa∗ 6= αa q̃a
  
.

Let us now prove that (s∗ , z ∗ ) is feasible for (6.2.5g)–(6.2.5i) and (6.2.5k). First we
prove that s∗a ≤ sa holds for every arc a ∈ A0 . Let a ∈ A0 . By definition (6.2.6a)
we have to show s∗a ≤ 0 if βa ỹa = max{βa y a , βa y a }. This means that one of the
following three cases applies:

1. ỹa = y a , βa > 0 ⇒ ya∗ ≤ ỹa ⇒ s∗a ≤ 0,

2. ỹa = y a , βa < 0 ⇒ ya∗ ≥ ỹa ⇒ s∗a ≤ 0,

3. βa = 0 ⇒ s∗a = 0.
6.2 Detecting Infeasibility of the Active Transmission Problem by MILP 161

Hence s∗a ≤ 0 if βa ỹa = max{βa y a , βa y a }. Similarly we prove that s∗a ≥ sa holds for
every arc a ∈ A0 . We conclude that s∗ is feasible for (6.2.5k). Now we turn to the
constraints (6.2.5g)–(6.2.5i). We consider an arc a = (v, w) ∈ A0 and obtain:

s∗v − s∗w + s∗a = πv0 (π ∗ ) − πw

0
(π ∗ ) + γr,v βa ya∗ − (πv0 (π̃) − πw
0
(π̃) + γr,v βa ỹa )
= γr,v αa (qa∗ |qa∗ |ka − q̃a |q̃a |ka )

> 0 if αa (qa∗ − q̃a ) > 0,




= 0 if αa (qa∗ − q̃a ) = 0,

< 0 if α (q ∗ − q̃ ) < 0.



a a a

We conclude that (s∗ , z ∗ ) is feasible for (6.2.5g)–(6.2.5i) and (6.2.5k).

By Lemma 6.2.5 there exists a vector x∗ for z ∗ such that (x∗ , z ∗ ) is feasible for
(6.2.5c)–(6.2.5f) and (6.2.5l)–(6.2.5n). Furthermore the flow conservation constraint
(6.2.5b) and the bound constraints (6.2.5j) are fulfilled by q ∗ as (q ∗ , π ∗ , p∗ , y ∗ ) is a
feasible solution for the active transmission problem (6.1.1). Hence (q ∗ , x∗ , s∗ , z ∗ ) is
a feasible solution for MIP (6.2.5).

We finally show that z ∗ > 0. Because of γr,v > 0 for all nodes v ∈ V we have
that αa qa∗ =
6 αa q̃a for an arc a = (v, w) ∈ A0 implies γr,v αa (qa∗ |qa∗ |ka − q̃a |q̃a |ka ) 6= 0.
This proves z ∗ > 0.

Remark 6.2.3:
As discussed below in Section 6.4 we will consider the infeasibility detection MILP
(6.2.5) only for a solution (q̃, π̃, p̃, ỹ) of (6.1.1) which does not violate the flow
conservation constraint (6.1.2c). We roughly describe how the MILP formulation of
(6.2.5) is obtained for this case. At first we replace (qa − q̃a ) by ∆a for every arc
a ∈ A0 . As q̃ fulfills the flow conservation we obtain that ∆ is a circulation, i.e., we
replace the flow conservation (6.2.5b) by

∆a − ∆a = 0
X X
∀ v ∈ V.
+
0 (v) 0 (v)
−
a∈δA a∈δA

We define bounds for ∆ by

 
∞

if q̃a < q a , 0

if q̃a = q a ,
∆a := ∆a :=
0
 if q̃a = q a , −∞
 if q̃a > q a ,
162 Chapter 6 Sufficient Conditions for Infeasibility

and replace (6.2.5j) by ∆ ≤ ∆ ≤ ∆. Then it is easy to see, as ∆ is a circulation,

that ∆ can be chosen such that either ∆a = 0 or |∆a | ≥ 1 holds. We introduce binary
variables xF
a
W
, xBW
a , sF
a
W
, sBW
a ∈ {0, 1} in combination with indicator constraints as
follows:

xF
a
W
= 1 ⇒ κ̃a ∆a ≥ 1 ∀ a ∈ A0 ,
xF
a
W
= 0 ⇒ κ̃a ∆a ≤ 0 ∀ a ∈ A0 ,

xBW
a = 1 ⇒ κ̃a ∆a ≤ −1 ∀ a ∈ A0 ,
xBW
a = 0 ⇒ κ̃a ∆a ≥ 0 ∀ a ∈ A0 ,

sF
a
W
= 1 ⇒ αa ∆a ≥ 1 ∀ a ∈ A0 ,
sF
a
W
= 0 ⇒ αa ∆a ≤ 0 ∀ a ∈ A0 ,

sBW
a = 1 ⇒ αa ∆a ≤ −1 ∀ a ∈ A0 ,
sBW
a = 0 ⇒ αa ∆a ≥ 0 ∀ a ∈ A0 .

Then (6.2.5d)–(6.2.5i) are replaced by

xF
a
W
= 1 ⇒ xv − xw + xa ≥ 0 ∀ a = (v, w) ∈ A0 ,
xF
a
W
= 0, xBW
a = 0 ⇒ xv − xw + xa = 0 ∀ a = (v, w) ∈ A0 ,
xBW
a = 1 ⇒ xv − xw + xa ≤ 0 ∀ a = (v, w) ∈ A0 ,
sF
a
W
= 1 ⇒ sv − sw + sa ≥ κa z ∀ a = (v, w) ∈ A0 ,
sF
a
W
= 0, sBW
a = 0 ⇒ sv − sw + sa = 0 ∀ a = (v, w) ∈ A0 ,
sBW
a = 1 ⇒ sv − sw + sa ≤ κa z ∀ a = (v, w) ∈ A0 .

All these reformulations yield an MILP with indicator constraints which is equivalent
to (6.2.5).

In the remaining part of this section we prove Lemma 6.2.5 which was used in
the previous proof of Theorem 6.2.2. Therefor we prove an auxiliary lemma.

Lemma 6.2.4:
Let (q̃, π̃, p̃, ỹ) be a solution for the active transmission problem (6.1.1) fulfilling at
least constraints (6.1.2b) and (6.1.2e). Furthermore let (q ∗ , π ∗ , p∗ , y ∗ ) be a feasible
solution for (6.1.1). There exists a partition of the node set V = M1 ∪˙ M2 ∪˙ M3
fulfilling the following conditions:
6.2 Detecting Infeasibility of the Active Transmission Problem by MILP 163

+
• ∀ a = (v, w) ∈ δA0 (M1 ) ∪ δA0 (M3 ):
−

(πv0 (π ∗ ) − πw
0
(π ∗ )) < (πv0 (π̃) − πw
0
(π̃)),
@k : [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k , (Aa )(k,1) (qa∗ − q̃a ) ≥ 0,
(Aa )(k,2) < 0, (Aa )(k,3) > 0,

+
• ∀ a = (v, w) ∈ δA
−
0 (M1 ) ∪ δA0 (M3 ):

(πv0 (π ∗ ) − πw
0
(π ∗ )) > (πv0 (π̃) − πw
0
(π̃)),
@k : [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k , (Aa )(k,1) (qa∗ − q̃a ) ≥ 0,
(Aa )(k,2) > 0, (Aa )(k,3) < 0.

Furthermore it holds

{v ∈ V | π̃v > π v } ⊆ M1 , {v ∈ V | π̃v ≤ π v } ∩ M1 = ∅,

{v ∈ V | π̃v < π v } ⊆ M3 , {v ∈ V | π̃v ≥ π v } ∩ M3 = ∅.

Proof. Let (q̃, π̃, p̃, ỹ) be a solution for the active transmission problem (6.1.1)
fulfilling at least constraint (6.1.2b) and (6.1.2e). Furthermore let (q ∗ , π ∗ , p∗ , y ∗ ) be
a feasible solution for (6.1.1). We iteratively construct the sets M1 , M2 , M3 with
V = M1 ∪˙ M2 ∪˙ M3 as follows:

1. Initially we set M1 := {v ∈ V | π̃v > π v }. Then we iteratively extend this set

+
by considering every arc a = (v, w) ∈ δA 0 (M1 ) ∪ δA0 (M1 ). If this arc does not
−

fulfill one of the following cases, then we either add v to M1 if v ∈

/ M1 and set the
predecessor p(v) := w or we add w to M1 if w ∈
/ M1 and set p(w) := v.

+
Case a = (v, w) ∈ δA0 (M1 ) :

(πv0 (π ∗ ) − πw
0
(π ∗ )) < (πv0 (π̃) − πw
0
(π̃)), (6.2.8a)
@k ∈ {1, . . . νa } : [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k , with
(Aa )(k,1) (qa∗ −q̃a ) ≥ 0, (Aa )(k,2) < 0, (6.2.8b)
(Aa )(k,3) > 0.
164 Chapter 6 Sufficient Conditions for Infeasibility

Case a = (v, w) ∈ δA
−
0 (M1 ) :

(πv0 (π ∗ ) − πw
0
(π ∗ )) > (πv0 (π̃) − πw
0
(π̃)), (6.2.9a)
@k ∈ {1, . . . νa } : [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k , with
(Aa )(k,1) (qa∗ −q̃a ) ≥ 0, (Aa )(k,2) > 0, (6.2.9b)
(Aa )(k,3) < 0.

+
This way we obtain the node set M1 such that every arc a = (v, w) ∈ δA0 (M1 )

fulfills (6.2.8) and every arc a = (v, w) ∈ δA

−
0 (M1 ) fulfills (6.2.9). Furthermore it

holds
M1 ∩ {v ∈ V | π̃v ≤ π v } = ∅.

This can be seen as follows: If π̃v ≤ π v holds for every node v ∈ V , then M1 = ∅
by construction. Assume that M1 contains a node t with π̃t ≤ π t . Then we
consider the nodes t, p(t), p(p(t)), . . ., s where s ∈ M1 has no predecessor. These
nodes define the nodes of an edge-disjoint s-t-path P in the undirected graph
(M1 , E 0 [M1 ]) which originates from (M1 , A0 [M1 ]) by removing the orientation of
each arc a ∈ A0 [M1 ]. Note that (M1 , E 0 [M1 ]) might contain multiple parallel edges.
This way each arc a ∈ A0 [M1 ] corresponds uniquely to an edge e ∈ E 0 [M1 ] and
vice versa. Let v1 , . . . , vn+1 be the nodes and e1 , . . . , en with ei = {vi , vi+1 } be the
ordered edges of P and a1 , . . . , an be the corresponding arcs in (V, A0 [M1 ]). We
have that for every arc ai , i = 1, . . . , n neither (6.2.8) nor (6.2.9) applies because
otherwise t would not be contained in M1 by construction of M1 . This means
that one of the following cases holds for every arc ai , i = 1, . . . , n:

Case (6.2.8a) and (6.2.9a) do not apply: Node potential loss estimation derived
as (6.2.8a) and (6.2.9a) do not apply, hence:

πv0 i (π ∗ ) − πv0 i+1 (π ∗ ) ≥ πv0 i (π̃) − πv0 i+1 (π̃).

Case (6.2.8b) and (6.2.9b) do not apply: In this case we differ between the ori-
entation of arc ai .

• If arc ai = (vi , vi+1 ) then, as (6.2.8b) does not apply, there exists an
index k ∈ {1, . . . , νai } such that [Aai (q̃ai , p̃vi , p̃vi+1 )]k ≥ [bai ]k holds with
(Aai )(k,1) (qa∗i − q̃ai ) ≥ 0, (Aai )(k,2) < 0, (Aai )(k,3) > 0. We rewrite this
6.2 Detecting Infeasibility of the Active Transmission Problem by MILP 165

inequality as a1 q̃ai − a3 ≥ p̃vi − a2 p̃vi+1 with a2 ∈ R≥0 and a1 , a3 ∈ R.

Then we derive the estimation

p̃vi − a2 p̃vi+1 ≤ a1 q̃ai − a3 ≤ a1 qa∗i − a3 ≤ p∗vi − a2 p∗vi+1 .

• If arc ai = (vi+1 , vi ) then, as (6.2.9b) does not apply, there exists an

index k ∈ {1, . . . , νai } such that [Aai (q̃ai , p̃vi+1 , p̃vi )]k ≥ [bai ]k holds with
(Aai )(k,1) (qa∗i − q̃ai ) ≥ 0, (Aai )(k,2) > 0, (Aai )(k,3) < 0. We rewrite this
inequality as a1 q̃ai − a3 ≥ p̃vi − a2 p̃vi+1 with a2 ∈ R≥0 and a1 , a3 ∈ R.
Again we derive the estimation

p̃vi − a2 p̃vi+1 ≤ a1 q̃ai − a3 ≤ a1 qa∗i − a3 ≤ p∗vi − a2 p∗vi+1 .

Because of the coupling constraint p̃v |p̃v | = π̃v relating the pressure and node
potential variables for each node v ∈ V , by using the previous estimations, we
obtain:

π̃v1 > πv∗1 ⇒ π̃v2 > πv∗2 , . . . , π̃vn+1 > πv∗n+1 ,

π̃vn+1 < πv∗n+1 ⇒ π̃vn < πv∗n , . . . , π̃v1 < πv∗1 .

The path P is chosen such that the start node v1 violates its upper node potential
bound, i.e., π̃v1 > π v1 and for the end node vn+1 it holds π̃vn+1 ≤ π vn+1 . Hence
the first of the above cases applies. We conclude that π ∗ violates a node potential
bound in vn+1 which is a contradiction to the assumption that (q ∗ , π ∗ , p∗ , y ∗ ) is
feasible for the active transmission problem.

2. In a second step we initially set M3 := {v ∈ V | πv < π v }. We now concentrate

on the graph (V \ M1 , A0 [V \ M1 ]). Again we iteratively consider each arc
+
a = (v, w) ∈ δA0 (M3 ) ∪ δA0 (M3 ). If neither (6.2.8) applies for an ingoing arc
−

nor (6.2.9) applies for an outgoing arc, then we add v to M3 if v ∈

/ M3 and w if
/ M3 . By definition it follows M1 ∩ M3 = ∅. By a similar reasoning as in the
w∈
previous item we conclude

M3 ∩ {v ∈ V | π̃v ≥ π v } = ∅.

3. In a third step we define M2 := {v ∈ V | v ∈

/ M1 ∪ M3 }. The previously defined
sets then have the property V = M1 ∪˙ M2 ∪˙ M3 .
166 Chapter 6 Sufficient Conditions for Infeasibility

Lemma 6.2.5:
Let (q̃, π̃, p̃, ỹ) be a solution for the active transmission problem (6.1.1) fulfilling at least
constraint (6.1.2b) and (6.1.2e). Furthermore let (q ∗ , π ∗ , p∗ , y ∗ ) be a feasible solution
for (6.1.1) and 0 ≤ z ∗ ≤ 1. There exists a vector x∗ = (x+
v , xv , xv , xa )v∈V,a∈A0 with
− ∗

x∗v + , x∗v − ∈ R≥0 , x∗v , x∗a ∈ R for v ∈ V and a ∈ A0 with x∗ 6= 0 which is feasible for
(6.2.5c)–(6.2.5f) and (6.2.5l)–(6.2.5n).

Proof. Let (q̃, π̃, p̃, ỹ) be a solution for the active transmission problem (6.1.1)
fulfilling at least constraint (6.1.2b) and (6.1.2e). Furthermore let (q ∗ , π ∗ , p∗ , y ∗ ) be
a feasible solution for the active transmission problem (6.1.1). For 0 ≤ z ∗ ≤ 1 we
describe how to define x∗ = (xv , xa )∗v∈V,a∈A0 which is feasible for (6.2.5d)–(6.2.5f)
and (6.2.5l).
˙ 2 ∪M
By Lemma 6.2.4 there exists a partition V = M1 ∪M ˙ 3 such that the following
holds:

+
• ∀ a = (v, w) ∈ δA0 (M1 ) ∪ δA0 (M3 ):
−

(πv0 (π ∗ ) − πw
0
(π ∗ )) < (πv0 (π̃) − πw
0
(π̃)),
@k ∈ {1, . . . , νa } : [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k , with (6.2.10)
(Aa )(k,1) (qa∗ −q̃a ) ≥ 0, (Aa )(k,2) < 0, (Aa )(k,3) > 0.

+
• ∀ a = (v, w) ∈ δA
−
0 (M1 ) ∪ δA0 (M3 ):

(πv0 (π ∗ ) − πw
0
(π ∗ )) > (πv0 (π̃) − πw
0
(π̃)),
@k ∈ {1, . . . , νa } : [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k , with (6.2.11)
(Aa )(k,1) (qa∗ −q̃a ) ≥ 0, (Aa )(k,2) > 0, (Aa )(k,3) < 0.

From this we obtain an estimation which is needed in the following:

a = (v, w) ∈ δ + (M1 ) ∪ δA
−
0 (M3 )

⇒ αa qa∗ |qa∗ |ka − βa ya∗ = γr,v

−1 0
(πv (π ∗ ) − πw
0
(π ∗ )) (6.2.12a)
−1 0
< γr,v (πv (π̃) − πw
0
(π̃)) = αa q̃a |q̃a |ka − βa ỹa

+
a = (v, w) ∈ δ − (M1 ) ∪ δA0 (M3 )

⇒ αa qa∗ |qa∗ |ka − βa ya∗ = γr,v

−1 0
(πv (π ∗ ) − πw
0
(π ∗ )) (6.2.12b)
−1 0
> γr,v (πv (π̃) − πw
0
(π̃)) = αa q̃a |q̃a |ka − βa ỹa
6.2 Detecting Infeasibility of the Active Transmission Problem by MILP 167

We use the sets M1 , M2 , M3 to define the values x∗v , v ∈ V and x∗a , a = (v, w) ∈ A0
as follows:

−1 if v ∈ M1 ,




xv := 0
∗
if v ∈ M2 , (6.2.13a)

1 if v

∈ M3 ,



+
0 if a ∈ δA 0 (M1 ) ∪ δA0 (M3 ) : κ̃a (qa − q̃a ) < 0,
− ∗




xa := 0
∗
if a ∈ δA
− +
0 (M1 ) ∪ δA0 (M3 ) : κ̃a (qa − q̃a ) > 0,
∗ (6.2.13b)

else.

x∗w − x∗v



We proceed by showing that this definition is feasible for constraints (6.2.5d)–(6.2.5f)

and (6.2.5l). We have x∗v = x∗w and x∗a = 0 for all arcs a = (v, w) ∈ A0 (M1 : M1 )
∪ A0 (M2 : M2 ) ∪ A0 (M3 : M3 ). Thus (6.2.5d)–(6.2.5f) and (6.2.5l) are fulfilled for
these arcs. Let us now turn to the remaining arcs. At first we observe:

• For every arc a ∈ A0 (M1 : M2 ) ∪ A0 (M2 : M3 ) ∪ A0 (M1 : M3 ) it holds

+
0 (M1 ) ∪ δA0 (M3 ).
−
a ∈ δA

• For every arc a ∈ A0 (M2 : M1 ) ∪ A0 (M3 : M1 ) ∪ A0 (M3 : M2 ) it holds

+
0 (M1 ) ∪ δA0 (M3 ).
−
a ∈ δA

These are the two cases that we distinguish in the following:

Case qa∗ < q̃a : We distinguish two cases.

+
Case a = (v, w) ∈ δA 0 (M1 ) ∪ δA0 (M3 ) : In this case we have xv − xw ≤ −1.
− ∗ ∗

• By (6.2.5f) κ̃a 6= 0 means xv − xw + xa ≤ 0 must be fulfilled by x∗ . We have

κ̃a 6= 0 ⇒ κ̃a > 0 ⇒ κ̃a (qa∗ − q̃a ) < 0 ⇒ x∗a = 0 by (6.2.13b). Hence it holds
x∗v − x∗w + x∗a ≤ 0 and xa ≤ x∗a ≤ xa .
• By (6.2.5e) κ̃a = 0 means xv − xw + xa = 0 must be fulfilled by x∗ . We
have x∗a = −(x∗v − x∗w ) > 0 by (6.2.13b). Hence it holds x∗v − x∗w + x∗a = 0.
κ̃a = 0 means αa = 0 by (6.2.6i). This in combination with (6.2.12a) implies
βa ya∗ > βa ỹa . By (6.2.10) there exists no index k so that [Aa (q̃a , p̃v , p̃w )]k ≥
[ba ]k , with (Aa )(k,1) (qa∗ − q̃a ) ≥ 0,(Aa )(k,2) < 0, (Aa )(k,3) > 0. This and the
conclusions that there exists no index k such that [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k ,
with (Aa )(k,1) =
6 0 (because of κ̃a = 0) especially implies that there exists no
index k such that [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k , with (Aa )(k,1) ≥ 0,(Aa )(k,2) < 0
and (Aa )(k,3) > 0. From this we conclude xa = ∞ by (6.2.6c). This yields
xa ≤ 0 < x∗a ≤ xa .
168 Chapter 6 Sufficient Conditions for Infeasibility

+
Case a = (v, w) ∈ δA
−
0 (M1 ) ∪ δA0 (M3 ) : In this case we have xv − xw ≥ 1.
∗ ∗

• From κ̃a ≥ 0 it follows κ̃a (qa∗ − q̃a ) ≤ 0. Hence we obtain from (6.2.13b) that
x∗a = −(x∗v − x∗w ) ≤ −1 < 0 holds. By (6.2.5e) and (6.2.5f) xv − xw + xa ≤ 0
or xv − xw + xa = 0 must be fulfilled by x∗ , which is obviously true. By
(6.2.12b) we have βa ya∗ < βa ỹa . By (6.2.11) there exists no index k such
that [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k , with (Aa )(k,1) (qa∗ − q̃a ) ≥ 0, (Aa )(k,2) > 0,
(Aa )(k,3) < 0. This especially means that there exists no index k such that
[Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k , with (Aa )(k,1) ≤ 0, (Aa )(k,2) > 0, (Aa )(k,3) < 0.
Hence we have xa = −∞ by (6.2.6d). This yields xa < x∗a ≤ 0 ≤ xa .

Case qa∗ = q̃a : We distinguish two cases.

+
Case a = (v, w) ∈ δA 0 (M1 ) ∪ δA0 (M3 ) : In this case we have xv − xw ≤ −1. By
− ∗ ∗

(6.2.12a) we have βa ya∗ > βa ỹa .

• By (6.2.5e) xv −xw +xa = 0 must be fulfilled by x∗ . Because of κ̃a (qa∗ −q̃a ) = 0
we have x∗a = −(x∗v − x∗w ) ≥ 1 > 0 by (6.2.13b). Hence x∗v − x∗w + x∗a = 0.
By (6.2.10) there exists no index k such that [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k ,
with (Aa )(k,1) (qa∗ − q̃a ) ≥ 0, (Aa )(k,2) < 0, (Aa )(k,3) > 0. This especially
implies that there exists no index k such that [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k , with
(Aa )(k,1) ≥ 0, (Aa )(k,2) < 0, (Aa )(k,3) > 0. Hence xa = ∞ by (6.2.6c). So
we obtain xa ≤ 0 < x∗a < xa .
+
Case a = (v, w) ∈ δA
−
0 (M1 ) ∪ δA0 (M3 ) : In this case we have xv − xw ≥ 1. By
∗ ∗

(6.2.12b) we have βa ya∗ < βa ỹa .

• By (6.2.5e) xv −xw +xa = 0 must be fulfilled by x∗ . Because of κ̃a (qa∗ −q̃a ) = 0
we have x∗a = −(x∗v − x∗w ) ≤ −1 < 0 by (6.2.13b). Hence x∗v − x∗w + x∗a = 0.
By (6.2.11) there exists no index k such that [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k ,
with (Aa )(k,1) (qa∗ − q̃a ) ≥ 0, (Aa )(k,2) > 0, (Aa )(k,3) < 0. This especially
implies that there exists no index k such that [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k , with
(Aa )(k,1) ≤ 0, (Aa )(k,2) > 0, (Aa )(k,3) < 0. Hence xa = −∞ by (6.2.6d).
This yields xa < x∗a < 0 ≤ xa .

Case qa∗ > q̃a : We distinguish two cases:

+
Case a = (v, w) ∈ δA 0 (M1 ) ∪ δA0 (M3 ) : In this case we have xv − xw ≤ −1. By
− ∗ ∗

(6.2.12a) we have βa ya∗ > βa ỹa .

• From κ̃a ≥ 0 it follows κ̃a (qa∗ − q̃a ) ≥ 0. Hence we obtain from (6.2.13b) that
x∗a = −(x∗v − x∗w ) ≥ 1 > 0 holds. By (6.2.5d) and (6.2.5e) xv − xw + xa ≥ 0
or xv − xw + xa = 0 must be fulfilled by x∗ , which is obviously true.
By (6.2.10) there exists no index k such that [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k ,
6.2 Detecting Infeasibility of the Active Transmission Problem by MILP 169

with (Aa )(k,1) (qa∗ − q̃a ) ≥ 0, (Aa )(k,2) < 0, (Aa )(k,3) > 0. This especially
implies that there exists no index k such that [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k , with
(Aa )(k,1) ≥ 0, (Aa )(k,2) < 0, (Aa )(k,3) > 0. Hence xa = ∞ by (6.2.6c). This
yields xa ≤ 0 < x∗a < ∞ = xa .
+
Case a = (v, w) ∈ δA
−
0 (M1 ) ∪ δA0 (M3 ) : In this case we have xv − xw ≥ 1.
∗ ∗

• By (6.2.5d) κ̃a 6= 0 means xv − xw + xa ≥ 0 must be fulfilled by x∗ . We

have κ̃a 6= 0 ⇒ κ̃a > 0 ⇒ κ̃a (qa∗ − q̃a ) > 0 ⇒ x∗a = 0 by (6.2.13b). Hence it
holds x∗v − x∗w + x∗a = x∗v − x∗w ≥ 1 ≥ 0 and xa ≤ x∗a ≤ xa .
• By (6.2.5e) κ̃a = 0 means xv − xw + xa = 0 must be fulfilled by x∗ . By
(6.2.13b) we have x∗a = −(x∗v −x∗w ) < 0 which implies x∗v −x∗w +x∗a = 0. κ̃a = 0
means αa = 0. This implies in combination with (6.2.12b) that βa ya∗ < βa ỹa
holds. By (6.2.11) there exists no index k such that [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k ,
with (Aa )(k,1) (qa∗ − q̃a ) ≥ 0, (Aa )(k,2) > 0, (Aa )(k,3) < 0. Additionally
κ̃a = 0 yields that @k : [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k with (Aa )(k,1) 6= 0 by
(6.2.6i). This especially implies that there exists no index k such that
[Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k , with (Aa )(k,1) ≤ 0, (Aa )(k,2) > 0, (Aa )(k,3) < 0.
Hence xa = −∞ by (6.2.6d). This yields xa < x∗a < 0 ≤ xa .

This case discussion proves that (x∗v , x∗a )v∈V,a∈A0 is feasible for (6.2.5d)–(6.2.5f) and
(6.2.5l). We set
x∗v + := max{0, x∗v + κv z ∗ } ∀ v ∈ V,
(6.2.14)
x∗v − := max{0, −x∗v − κv z ∗ } ∀ v ∈ V.

From this definition it follows that (6.2.5c) is fulfilled. We prove that this definition
is feasible for (6.2.5m) and (6.2.5n). Therefor we make use of

{v ∈ V | π̃v > π v } ⊆ M1 {v ∈ V | π̃v ≤ π v } ∩ M1 = ∅, (6.2.15a)

{v ∈ V | π̃v < π v } ⊆ M3 {v ∈ V | π̃v ≥ π v } ∩ M3 = ∅, (6.2.15b)

which holds by Lemma 6.2.4. We consider the bound constraints (6.2.5m) and
(6.2.5n) separately:

• It holds x∗v + ≤ x+
v for every node v ∈ V : Let v ∈ V . According to (6.2.6e)
we have to show x∗v + = 0 if π̃v ≥ π v . In this case it holds v ∈ M1 ∪ M2 by
(6.2.15b). We distinguish two cases:

(6.2.6g)
π̃v = π v ⇒ κv = 0
(6.2.13a),(6.2.15b)
⇒ x∗v + κv z ∗ = x∗v ≤ 0 ⇒ x∗v + = max{0, x∗v + κv zv∗ } = 0.
170 Chapter 6 Sufficient Conditions for Infeasibility

(6.2.6g),(6.2.15a),(6.2.13a)
π̃v > π v ⇒ κv = 1, x∗v = −1
z ∗ ∈[0,1]
⇒ x∗v + κv z ∗ = −1 + z ∗ ≤ 0 ⇒ x∗v + = max{0, x∗v + κv z ∗ } = 0.

• It holds x∗v − ≤ x−
v for every node v ∈ V : Let v ∈ V . According to (6.2.6f)
we have to show x∗v − = 0 if π̃v ≤ π v . In this case it holds v ∈ M3 ∪ M2 by
(6.2.15a). We distinguish two cases:

(6.2.6g)
π̃v = π v ⇒ κv = 0
(6.2.13a),(6.2.15a)
⇒ x∗v + κv z ∗ = x∗v ≥ 0 ⇒ x∗v − = min{0, x∗v + κv z ∗ } = 0.

(6.2.6g),(6.2.15b),(6.2.13a)
π̃v < π v ⇒ κv = −1, x∗v = 1
z ∗ ∈[0,1]
⇒ x∗v + κv z ∗ = 1 − z ∗ ≥ 0 ⇒ x∗v − = min{0, x∗v + κv z ∗ } = 0.

We conclude that the definition (6.2.13a), (6.2.13b), and (6.2.14) yield a vector x∗
which is feasible for (6.2.5c)–(6.2.5f) and (6.2.5l)–(6.2.5n).

6.3 Interpretation of the Infeasibility Detection MILP

In this section we present the initial idea which led to the formulation of the
infeasibility detection MILP (6.2.5). The concept becomes visible when looking at
the dual problem of (6.2.5) for a fixed flow vector q 0 ∈ RA which fulfills the flow
0

conservation and bound constraints (6.2.7), i.e.,

qa0 = dv
X X
qa0 − ∀ v ∈ V, q a ≤ qa0 ≤ q a ∀ a ∈ A0 .
+
0 (v) 0 (v)
−
a∈δA a∈δA

Let (q̃, π̃, p̃, ỹ) be a solution of the active transmission problem (6.1.1) fulfilling at least
constraint (6.1.2b) and (6.1.2e). Assume that the MILP (6.2.5) has optimal objective
value zero. By Theorem 6.2.2 we conclude that there do not exist vectors π 0 ∈ RV ,
p0 ∈ RV , y 0 ∈ RA such that (q 0 , π 0 , p0 , y 0 ) is feasible for the active transmission
0

problem (6.1.1). In the following we show an example demonstrating that especially

π 0 ∈ RV with π ≤ π 0 ≤ π cannot exist. Therefor we assume that (q 0 , π 0 , p0 , y 0 ) is a
feasible solution for (6.1.1) and derive a contradiction by comparing (q 0 , π 0 , p0 , y 0 )
and (q̃, π̃, p̃, ỹ).
6.3 Interpretation of the Infeasibility Detection MILP 171

As the MILP (6.2.5) has optimal objective value zero it is easy to see that the
following linear optimization problem is bounded (recall that q 0 and q̃ are fixed):

max z (6.3.1)

s. t. x+
v − x v − x v − κv z = 0
−
 
λv ∀ v ∈ V,
xv − xw + xa − x+ κ̃a (qa0 − q̃a ) = 0 ∀ a = (v, w) ∈ A0 ,
 
µa a

νa sv − sw + sa − s+
a αa (qa − q̃a ) − κa z = 0
0
∀ a = (v, w) ∈ A0 ,
 

sa ≤ sa ≤ sa ∀ a ∈ A0 ,
xa ≤ xa ≤ xa ∀ a ∈ A0 ,
x+ +
v ≤ xv ∀ v ∈ V,
x−
v ≤ x−
v ∀ v ∈ V,
xv , s v ∈ R ∀ v ∈ V,
x+ −
v , xv ∈ R≥0 ∀ v ∈ V,
xa , sa ∈ R ∀ a ∈ A0 ,
x+ +
a , sa ∈ R≥0 ∀ a ∈ A0 ,
z ∈ R≥0 .

We associated dual variables λv for each node v ∈ V and µa , νa for each arc a ∈ A0 .
As (6.3.1) is bounded it follows that its dual is feasible. This dual is as follows:

∃ λ, µ, ν (6.3.2a)

s. t. νa = 0 (6.3.2b)
X X
νa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

µa − λv = 0 (6.3.2c)
X X
µa − ∀ v ∈ V,
+
0 (v) 0 (v)
−
a∈δA a∈δA

λv + κa νa ≥ 1, (6.3.2d)
X X X
λv −
v∈V :π̃v >π v v∈Vπ :π̃v <π v a∈A0

κ̃a (qa0 − q̃a ) µa ≥ 0 ∀ a ∈ A0 , (6.3.2e)

αa (qa0 − q̃a ) νa ≥ 0 ∀a ∈ A , 0
(6.3.2f)
λv ≤ λv ≤ λv ∀ v ∈ V, (6.3.2g)
µa ≤ µa ≤ µa ∀a ∈ A , 0
(6.3.2h)
ν a ≤ νa ≤ ν a ∀ a ∈ A0 , (6.3.2i)
µa , νa ∈ R ∀a ∈ A , 0
(6.3.2j)
172 Chapter 6 Sufficient Conditions for Infeasibility

λv ∈ R ∀ v ∈ V. (6.3.2k)

Here the variable bounds are defined as


∞− if π̃ ≥ π ,

v v
λv := ∀ v ∈ V, (6.3.3a)
0
 else

−∞ if π̃ ≤ π ,

v
λv := (6.3.3b)
v
∀ v ∈ V,
0
 else

∞− if β ỹ = max{β y , β y }

∀ a ∈ A0 ,
a a a a a a
ν a := (6.3.3c)
0
 else a = (v, w),

−∞ if β ỹ = min{β y , β y }

∀ a ∈ A0 ,
a a a a a a
ν a := (6.3.3d)
0
 else a = (v, w).

∞− if βa ỹa = max{βa y a , βa y a }





∞− if ∃k : [Aa (q̃a , p̃v , p̃w )]k ≥ [ba ]k ,




∀ a ∈ A0 ,


µa := (Aa )(k,1) ≥ 0, (Aa )(k,2) < 0, (6.3.3e)
 a = (v, w),
(A ) 0,

>

a (k,3)




0 else




−∞ if βa ỹa = min{βa y , βa y a }



 a
if : [A (q̃ a v w )]k ≥ [ba ]k ,

−∞ ∃k , p̃ , p̃


 a
∀ a ∈ A0 ,


µa := (Aa )(k,1) ≤ 0, (Aa )(k,2) > 0, (6.3.3f)
 a = (v, w),
(A ) 0,

<

a (k,3)




0 else




Now let (λ∗ , µ∗ , ν ∗ ) be a feasible solution for (6.3.2). The vectors ν ∗ and µ∗ form
a network flow by constraints (6.3.2b) and (6.3.2c). For the following discussion we
focus on the case that either µ∗ ≥ 0 if λ∗ 6= 0 or ν ∗ ≥ 0 holds. The case where these
assumptions are not fulfilled can be led back to the case fulfilling the assumptions by
changing the orientation of some arcs. Our initial motivation for the definition of
MILP (6.2.5) was to look either for a path or a circuit as discussed in the following
two cases:
6.3 Interpretation of the Infeasibility Detection MILP 173

Case λ∗ 6= 0: We split the network flow µ∗ into sets of flow along paths P1 , . . . , Pm
and flow along circuits C1 , . . . , Cn , see Theorem 4.2.5. This way we obtain from µ∗ ≥ 0
that there exist flow values µPi > 0, i = 1, . . . , m and µCi > 0, i = 1, . . . , n such that

µ∗a = µPi +
X X
µCi ∀ a ∈ A0 .
i=1,...,m: i=1,...,n:
a∈A0 (Pi ) a∈A0 (Ci )

Consider a path P` that starts in node v and ends in node w. Because of con-
straint (6.3.2d) the index ` can be chosen such that either π̃v > π v and π̃w ≤ π w
or π̃v ≥ π v and π̃w < π w holds. Let the nodes of P` be given by v1 , . . . , vn+1 where
v1 = v and vn+1 = w and connecting arcs by a1 , . . . , an .

In order to show that (q 0 , π 0 , p0 , y 0 ) is not feasible for the active transmission prob-
lem (6.1.1) we distinguish two cases for each arc ai of the path P` :

1. In the case that βai ỹai = max{βai y a , βai y ai } we obtain the following estimation
i
from q̃ai ≤ qa0 i if αai 6= 0 (by (6.3.2e)) and βai ya0 i ≤ βai ỹai :

π̃vi − γai π̃vi+1 = αai q̃ai |q̃ai |kai − βai ỹai

≤ αai qa0 i |qa0 i |kai − βai ya0 i = πv0 i − γai πv0 i+1 .

2. In the case that [Aai (q̃ai , p̃vi , p̃vi+1 )]k ≥ [bai ]k holds for an index k with
(Aai )(k,1) ≥ 0, (Aai )(k,2) < 0, (Aai )(k,3) > 0 we rewrite this inequality as
a1 q̃ai − a3 ≥ p̃vi − a2 p̃vi+1 with a1 ∈ R≥0 , a2 ∈ R>0 and a3 ∈ R. Then we derive
the estimation (using (6.3.2e) and (6.2.6i), which yields a1 > 0 ⇒ q̃ai ≤ qa0 i ):

p̃vi − a2 p̃vi+1 ≤ a1 q̃ai − a3 ≤ a1 qa0 i − a3 ≤ p0vi − a2 p0vi+1 .

We note that at least one of the previous cases applies because of 0 < µ∗ai ≤ µai and
(6.3.3e). Because of the coupling constraint p̃v |p̃v | = π̃v relating the pressure and
node potential variables for each node v ∈ V , we obtain:

π̃v1 > πv0 1 ⇒ π̃v2 > πv0 2 , . . . , π̃vn+1 > πv0 n+1 , (6.3.4a)
π̃vn+1 < πv0 n+1 ⇒ π̃vn < πv0 n , . . . , π̃v1 < πv0 1 . (6.3.4b)
174 Chapter 6 Sufficient Conditions for Infeasibility

The path P` is chosen such that either the start node v1 or the end node vn+1 of
path P` violates its node potential bound, i.e., one of the following cases applies:

(6.3.4a) 0
π̃v1 > π v1 and π̃vn+1 ≤ π vn+1 and π v1 ≥ πv0 1 ⇒ πvn+1 < π vn+1 ,
(6.3.4b)
π̃v1 ≥ π v1 and π̃vn+1 < π vn+1 and π vn+1 ≤ πv0 n+1 ⇒ πv0 1 > π v1 .

Hence π 0 violates the node potential bounds which implies that (q 0 , π 0 , p0 , y 0 ) is not
feasible for the active transmission problem (6.1.1).

Case ν ∗ 6= 0, λ∗ = 0: Similar as in the previous case we split the network flow ν ∗ into
sets of flow along circuits C1 , . . . , Cn , see Theorem 4.2.5. By constraint (6.3.2d) there
exists an arc a ∈ A0 with νa∗ 6= 0 and αa > 0 by (6.2.6h). From our assumption we
obtain νa∗ > 0 for this arc. Let ` be chosen such that C` contains this arc. Let the nodes
of C` be given by v1 , . . . , vn+1 where v1 = vn+1 and connecting arcs by a1 , . . . , an .
We note that βai ỹai = max{βai y a , βai y ai } holds because of 0 < µ∗ai ≤ µai and
i
(6.3.3c). From this observation we derive the following contradiction from q̃ai ≤ qa0 i
if αai 6= 0 (by (6.3.2f)):
 
n i−1
0= γaj  (π̃vi − γai π̃vi+1 )
X Y

i=1 j=1
 
n i−1
= γaj  (αai q̃ai |q̃ai |kai − βai ỹai )
X Y

i=1 j=1
 
n i−1
γaj  (αai qa0 i |qa0 i |kai − βai ya0 i )
X Y
< 
i=1 j=1
 
n i−1
= γaj  (πv0 i − γai πv0 i+1 ) = 0.
X Y

i=1 j=1

The inequality is strict because κa νa > 0 by (6.3.2d) and q̃a < q a ≤ qa0 by (6.2.6h)
and the feasibility of q 0 . This contradiction implies that our assumption was wrong
and hence the solution (q 0 , π 0 , p0 , y 0 ) is not feasible for the active transmission prob-
lem (6.1.1).

We note that at least one of the above cases applies because of constraint (6.3.2d). This
contradicts our assumption that (q 0 , π 0 , p0 , y 0 ) is feasible for the active transmission
problem (6.1.1). This shows that this assumption was wrong.
6.4 Integration and Computational Results 175

6.4 Integration and Computational Results

In this chapter we focused on the topology optimization problem (3.2.1) for the third
type of network considered in this thesis. Recall that every gas network is associated
with this type, i.e., especially compressors and control valves might be contained.
The strategy for solving the topology optimization problem is as follows: We solve
the model (3.2.1) by SCIP as described in Section 2.2. Furthermore we combine the
methods of the previous sections for solving the active transmission problem (6.1.1)
as follows: First we compute a feasible solution of one of the relaxations (6.1.2) or
(6.1.5). We either obtain a feasible solution for (6.1.1) directly, if the slack value
equals zero. Or we solve the infeasibility detection MILP (6.2.5). If this problem
turns out to be infeasible or has optimal objective value zero, then we conclude that
the active transmission problem (6.1.1) is infeasible by Theorem 6.2.2.
In Chapter 4 we computationally showed that the domain relaxation (4.2.1), which
is based on bound relaxations, has lower computation times than the relaxation (4.3.1),
which relaxes the flow conservation constraint. Let us proceed similarly for a
comparison of the two non-convex problems (6.1.2) and (6.1.5). We implemented both
relaxations in C and use the computational setup described in Section 3.5. A scatter
plot comparing the run times of both relaxations is shown in Figure 6.3 for different
scenarios and the network net6. It turns out that the domain relaxation (6.1.2) has
better solving performance. Thus we use this relaxation in our solution approach
and will not consider the flow conservation relaxation (6.1.5) any longer.
The new solution approach that we follow for solving the topology optimization
problem (3.2.1) is as follows: At the time during the solving process when a node of
the branching tree is considered and where all binary decisions x are fixed, we consider
the corresponding active transmission problem and solve the domain relaxation (6.1.2).
If we obtain a solution with zero slack, then this yields a feasible solution for the
active transmission problem. In this sense, solving the domain relaxation (6.1.2) is
a primal heuristic for the active transmission problem. Otherwise, if we obtain a
solution with positive slack, then the solution is infeasible for the active transmission
problem and violates at most the constraints (6.1.1d) and (6.1.1f)-(6.1.1i). Now we
apply Theorem 6.2.2 and solve the infeasibility detection MILP (6.2.5). If (6.2.5) is
infeasible or has optimal objective value zero, then the active transmission problem
is infeasible by Theorem 6.2.2. In this case we prune the corresponding node of
the branch-and-bound tree. Otherwise, if we cannot decide that the current active
transmission problem is infeasible, then we continue with branching.
We implemented the algorithm above in C, i.e., solving the domain relaxation
(6.1.2) and the infeasibility detection MILP (6.2.5). The MILP is expressed using
176 Chapter 6 Sufficient Conditions for Infeasibility

101

Run time domain relaxation (sec)

100

10−1

10−2

10−2 10−1 100 101

Run time flow conservation relaxation (sec)

Figure 6.3: Run time comparison for the domain relaxation (6.1.2) and the flow conservation relax-
ation (6.1.5) on instances of network net6.

indicator constraints as described in Remark 6.2.3. For the computational studies

we use the setup described in Section 3.5. We compare three strategies for solving
the topology optimization problem (3.2.1).

1. The first strategy is to use SCIP without any adaptations on the solver settings.
All branching decisions are up to the solver, and the topology optimization
problem (3.2.1) is basically solved by branch-and-bound, separation and spatial
branching as described in Section 2.2.

2. The second strategy is to enforce a certain branching priority rule, so that SCIP
first branches on binary and discrete decision variables. Only after all discrete
variables are fixed it is allowed to perform spatial branching on continuous
variables.

3. The third strategy implements the domain relaxation strategy presented in

this chapter for solving the active transmission problem (6.1.1). We use the
nonlinear solver IPOpt for solving the domain relaxation (6.1.2) and SCIP
for solving the infeasibility detection MILP (6.2.5) in advance. For solving
this MILP we set a time limit of 15 s. Additionally we set branching priorities
according to the second strategy.
6.4 Integration and Computational Results 177

100

Number of solved instances (%) 75

50

25
strategy 3
strategy 2
strategy 1
0
100 101 102 103
Run time (sec)

Figure 6.4: Performance plot for different nominations on the network net6 and a time limit of 3600 s.
The different strategies are described in Section 6.4. Strategy 1 and 2 mainly consist of SCIP. Strategy
3 also corresponds to SCIP together with our elaborated solution method presented in this chapter. The
underlying data are available in Table A.10.

Computational Results

We consider the real-world network net6 shown in Figure 3.10. This network contains
active elements, i.e., valves, compressors and control valves and is in industrial use.
We solve the nomination validation problem which is a feasibility problem. The 43
instances are nominations from our industrial partner. Figure 6.4 shows a performance
plot of the instances that could be solved within a given time limit of 3600 s. From
Figure 6.4 we conclude that the first strategy clearly performs worse than the second
and third one. The third strategy keeps up with the second one but performs slightly
worse.
Now we turn to the networks net3, net4, and net5. The instances are topology
expansion problems, i.e., we optimize over a set of extensions and the objective
function is nonzero. For all these instances we set a time limit of 39 600 s. We
consider an additional fourth strategy:

4. The fourth strategy implements the domain relaxation (6.1.2) for computing a
primal feasible solution. In comparison to the third strategy we do not check
infeasibility conditions of the active transmission problem by solving (6.2.5).
178 Chapter 6 Sufficient Conditions for Infeasibility

strategy 1 2 3 4 all
solved instances 43 48 63 50 63

Table 6.5: Summary of the Tables A.11-A.13 showing the globally solved instances out of 76 nominations
in total. The third strategy globally solves all instances which are solved to global optimality by the
second and the first strategy.

(A,B) = (2,3) (A,B) = (2,4)

solved(48) incomp.(3) solved(46) incomp.(15)
time [s] nodes gap [%] time [s] nodes gap [%]
strategy A 50.4 2,565 28 42.4 2,215 31
strategy B 62.3 2,021 17 59.4 2,280 40
shifted geom. mean 23 % −21 % −40 % 40 % 3% 29 %

Table 6.6: Run time, number of branch-and-bound nodes and gap comparison for the strategies 2 and
3 and additionally 2 and 4 (aggregated results). The columns “solved” contain mean values for those
instances globally solved by both strategies A and B. The columns “incomplete” show mean values for
those instances having a primal feasible solution available but were not globally solved by both strategies
A and B. The underlying data are available in Tables A.11-A.13.

So we do not detect infeasibility of the active transmission problem and thus

do not manually prune any node of the branch-and-bound tree.

The computational results are shown in Tables A.11-A.13 and a summary is

available in Table 6.5 and 6.6. We use the geometric mean of run time, number of
branch-and-bound nodes and gap as described in Section 3.5. The third strategy
clearly outperforms the other strategies in terms of number of solved instances. All
instances globally solved by strategies 1 or 2 or 4 are also solved by the third strategy.
Approximately 20 % more instances of the test set (15 out of 76) are solved by the
third strategy compared to the second one. The second strategy performs better than
the first one because it solves 5 more instances within the time limit. We conclude
that branching priorities as set by the second strategy are a first step to improve
the solving performance of SCIP. This is consistent with our observations from
Chapter 4.
Approximately 63 % of the instances of the test set (48 out of 76) are globally
solved by the second strategy. Here the run time increases by approximately 23 % on
average following strategy 3 while the number of nodes is decreased by approximately
21 %. Clearly the reduction does not pay off due to the increasement in run time. We
conclude that the domain relaxation (6.1.2) together with the infeasibility detection
MILP (6.2.5) consume more time than spatial branching on those instances globally
solved by SCIP in combination with branching priorities.
6.4 Integration and Computational Results 179

100
strategy 3
strategy 4
strategy 2
Number of solved instances (%) 75
strategy 1

50

25

0
10−1 100 101 102 103 104
Run time (sec)

Figure 6.5: Performance plot for different nominations on the networks net3, net4, and net5 and a time
limit of 39 600 s. The different strategies are described in Section 6.4. Strategy 1 and 2 mainly consist
of SCIP. Strategy 3 also corresponds to SCIP together with our elaborated solution method presented
in this chapter. Strategy 4 is a weaker version of strategy 3 and only necessary for evaluation of the
computational results. The underlying data are available in Tables A.11-A.13.

The fourth strategy allows to solve 65 % of the instances of the test set (50
out of 76) and hence performs nearly similar as the second strategy in terms of
number of globally solved instances. But the run time increases by approximately
40 % and the number of nodes by approximately 3 %. We conclude that the primal
heuristic (solving the domain relaxation (6.1.2)) as well as the verification of the
infeasibility conditions (represented by the infeasibility detection MILP (6.2.5)) are
both important for the performance of the third strategy.

A performance plot of all instances is shown in Figure 6.5. We conclude that the
third strategy yields the best results and use it in our practical application. Here we
accept an increase in run time because more instances are solved by this strategy.

In a last step we focused on the network net7. Here the results do not differ
much in comparison to the initial computational study in Section 1.4. Thus we do
not report them. Instead we present a primal heuristic for these instances in the
next chapter. We also apply this heuristic to the network net5 because very few
feasible solutions are computed by the afore presented approach.
180 Chapter 6 Sufficient Conditions for Infeasibility

105

104
Run time strategy 3 (sec)
103

102

101

100

10−1

10−2

10−2 10−1 100 101 102 103 104 105

Run time strategy 2 (sec)
Figure 6.6: Run time comparison on different nominations on the networks net3, net4, and net5. Each
cross (×) corresponds to a single instance of the test set. Note that multiple crosses are drawn in the
upper right corner of the plots and cannot be differed. They represent those instances that ran into the
time limit by both strategies. The underlying data are available in Tables A.11-A.13.

Summary

We presented a new method for proving infeasibility of the active transmission prob-
lem (6.1.1) when given a feasible solution for the relaxations (6.1.2) or (6.1.5). This
allows the following solution approach for the topology optimization problem (3.2.1):
We use SCIP and therein solve the active transmission problem (6.1.1) by computing
a local optimal solution of the domain relaxation (6.1.2), which is a non-convex opti-
mization problem. In advance we analyze this solution by the infeasibility detection
MILP (6.2.5) and possibly conclude that the active transmission problem is infeasible.
This clearly outperforms SCIP used without any adaptations on the implementation.
Approximately 20 % more instances of the test set are solved to global optimality
compared to the solver SCIP. The test set is formed by topology expansion instances
on the third type of network considered in this thesis, i.e., networks that contain
control valves and compressors.
Chapter 7

A Primal Heuristic based on Dual

Information

In this chapter we present a primal heuristic for the topology optimization prob-
lem (3.2.1). In contrast to the previous Chapters 4-6 we do not focus on a special
class of gas networks, i.e., we do not restrict to networks that contain only certain
elements.
Recall that the subproblem which results from fixing all binary variables x in
(3.2.1) yields the active transmission problem (6.1.1). It is a continuous nonlinear
optimization problem and the constraints are continuously differentiable functions.
The outline of the heuristic is as follows: We consider a feasible relaxation of the
active transmission problem and compute a KKT point. If we obtain a primal
solution with zero slack, then we derive a feasible solution for (3.2.1) by neglecting
the slack variables. Otherwise we make use of the dual solution of the KKT point
and apply parametric sensitivity analysis to identify constraints of the subproblem
which we adapt in a next step. For these constraints we consider the associated
binary variables of the MINLP and switch them to different values. The new binary
values lead to another subproblem of the original MINLP which contains continuous
variables only. As before, we consider a relaxation of the new active transmission
problem and compute a KKT point. On the basis of the choice of the switching
variables we have good chances to obtain a solution that has less slack than the
solution that we computed before. We repeat this process for a predefined number
of iterations. Furthermore we integrate this method in a branch-and-prune search
in order to avoid cycles. Our aim is to come up with a certain active transmission
problem together with a feasible solution, i.e., a feasible solution with zero slack for
the corresponding relaxation.
Recall that the topology optimization problem (3.2.1) is solved by branch-and-
bound, separation, and spatial branching. Within this approach the heuristic is

181
182 Chapter 7 A Primal Heuristic based on Dual Information

invoked at those nodes in the branch-and-bound tree where all binary decision
variables have binary values, either by branching or by the solution of the LP
relaxation. Fixing all binary variables of the topology optimization problem yields
the active transmission problem (6.1.1). By Lemma 6.1.1 a feasible relaxation is
given by the domain relaxation (6.1.2). By using the dual solution of a KKT point
of this relaxation we identify binary variables that we change in each iteration. This
selection will be improved further. We make use of the problem structure of the
topology optimization problem by switching selectively between the different modes
of a valve, a control valve or a compressor.
We apply the heuristic to the real-world network net7. Note that net7 remains
with very few primal feasible solutions available compared to the computational
results of the previous chapters. The associated problem for net7 is the nomination
validation problem and so we solve it only for feasibility. Our initial computational
study showed that the MINLP solvers Antigone, Baron and SCIP were not
able to compute a feasible solution for more than one instance of the nomination
validation problems on the final version of network net7 that we were given from
our cooperation partner, see Section 1.4. Using the heuristic we are able to solve 18
out of 30 instances which corresponds to approximately 60 % of all these instances.
When taking all the different versions of the network into account, the heuristic
allows to solve globally approximately 61 % more instances of the test set compared
to SCIP without any adaptations. We also apply the heuristic to the network net5.
Similar to net7, this network remains with very few primal feasible solutions available
compared to the results of the previous chapters. Here we combine the heuristic with
the solution approach presented in the Chapter 6.
The outline of this chapter is as follows: In Section 7.1 we describe the MINLP
that we consider in this chapter. The heuristic is presented in Section 7.2. In
Section 7.3 we show an adaptation of the heuristic to the topology optimization
problem (3.2.1). Afterwards we give some information of the implementation details
in Section 7.4. Computational results are given in Section 7.5.

7.1 A Relaxation of the MINLP

In this chapter we consider a mixed-integer nonlinear optimization problem with
indicator constraints and variables x, z of the following form:

min f (x, z) (7.1.1a)

s. t. g(x, z) ≤ 0, (7.1.1b)
7.1 A Relaxation of the MINLP 183

hi (z) − ∆i ≤ 0 ∀ i ∈ I, (7.1.1c)
x i = 1 ⇒ ∆i = 0 ∀ i ∈ I, (7.1.1d)
z ≤ z ≤ z, (7.1.1e)
z∈R , n
(7.1.1f)
∆ ∈ RI≥0 , (7.1.1g)
x ∈ {0, 1} . I
(7.1.1h)

Here n ∈ N are the dimensions of the real variables. Furthermore I := {1, 2, . . .}

with cardinality |I| is a nonempty index set. The functions f, g : {0, 1}I × Rn → R
and hi : Rn → R (for i ∈ I) are C 2 (twice continuously differentiable). The binary
variable xi represents the decision whether the constraint hi (z) ≤ 0 is active (for
xi = 1) or inactive (for xi = 0). A constraint hi is inactive, if and only if a slack
variable ∆i is positive, where ∆ = (∆1 , . . . , ∆|I| ) denotes the vector of continuous
slack variables. The constraints (7.1.1d) are the indicator constraints of the problem.
For the ease of notation we first focus on MINLP (7.1.1) in a general form before
turning to the problem of optimizing the topology of gas networks in Section 7.3.
Now we turn to a relaxation of a subproblem of (7.1.1) which forms the basis for
our heuristic and is obtained by fixing all binary variables. Assume we are given
binary values x∗ for the binary variables x. Then the subproblem is obtained by
fixing x = x∗ . Note that it is a nonlinear optimization problem which might be
infeasible. To circumvent its infeasibility we consider a relaxation of this subproblem
of (7.1.1). This relaxation is obtained as follows: Let I 0 ⊆ I contain those indices
such that the binary variables x∗i are 1 for all i ∈ I 0 . We define I 00 ⊆ I 0 and relax all
constraints hi (z) ≤ 0 with i ∈ I 00 . Furthermore we introduce a subset J ⊆ {1, . . . , n}
and relax the variable bounds of xj for all j ∈ J. Therefor we introduce two vectors
of slack variables s+ , s− ∈ RJ≥0 . Then the relaxation writes as follows for p = 0 ∈ RI :
0

min σf (x∗ , z) + (1 − σ) (s+

j + sj ) + (1 − σ) ∆i (7.1.2a)
X X
−

j∈J i∈I 00

s. t. g(x∗ , z) ≤ 0, (7.1.2b)
hi (z) ≤ pi ∀ i ∈ I 0 \ I 00 , (7.1.2c)
hi (z) − ∆i ≤ pi ∀i ∈ I , 00
(7.1.2d)
zj − s+
j ≤ zj ∀ j ∈ J, (7.1.2e)
zj + s−
j ≥ zj ∀ j ∈ J, (7.1.2f)
z ∈ Rn , (7.1.2g)
184 Chapter 7 A Primal Heuristic based on Dual Information

∆ ∈ RI+ , (7.1.2h)
00

±
s ∈ RJ+ . (7.1.2i)

We assume a definition of the index sets I 00 and J in such a way that the relaxation
(7.1.2) is feasible. The parameter σ ∈ [0, 1] in the objective function (7.1.2a) controls
the compromise between improving feasibility and optimality: For σ = 0 we aim to
compute a primal feasible solution for MINLP (7.1.1) while for σ = 1 our goal is
to compute a globally optimal solution for (7.1.1). Hence a lower value of σ puts a
higher emphasis on feasibility. The term

(1 − σ) (s+
j + sj ) + (1 − σ) ∆i
X X
−

j∈J i∈I 0

is called the slack part of the objective function.

7.2 A Primal Heuristic for MINLP with Indicator

Constraints
In the following we describe the primal heuristic for the MINLP with indicator
constraints (7.1.1). It is invoked for a given vector of binary values x∗ . We regard
(7.1.2) as an optimization problem with parameters (x∗ , p). The heuristic aims to
compute a vector x∗ and a feasible solution for (7.1.2) with parameters (x∗ , 0) so that
the objective value of this solution equals zero. When neglecting the slack variables
we obtain a feasible solution for (7.1.1).
The outline of this section is as follows: First we give a motivation of the
mathematical background in Section 7.2.1. Then we present the procedure of our
heuristic in Section 7.2.2. In Section 7.2.3 we describe how to embed the heuristic in
a branch-and-prune search.

7.2.1 Theoretical Motivation

Let us denote the objective function in (7.1.2a) by f˜(x∗ , z, s, ∆). We consider a KKT
ˆ and dual solution (λ̂, η̂, µ̂, ξ,
point of (7.1.2) with primal solution (ẑ, ŝ, ∆) ˆ ζ̂). Here
the dual variables λ correspond to inequalities (7.1.2b), η to (7.1.2c) and (7.1.2d),
µ± to (7.1.2e) and (7.1.2f), ξ to (7.1.2h), and ζ ± to (7.1.2i), respectively. Our aim is
to make use of equation (5.4) of Fiacco and Ishizuka (1990), which states a relation
between a modification of a constraint and its impact on the objective function value
7.2 A Primal Heuristic for MINLP with Indicator Constraints 185

via Lagrange multipliers from a corresponding KKT point. The main result that we
derive from their equation is

ˆ p∗ ) = −η̂i
∂pi f˜(x∗ , ẑ, ŝ, ∆, ∀ i ∈ I 0, (7.2.1)

where the dual value η̂i is the Lagrange multiplier of the constraints hi (z)−∆i −p∗i ≤ 0
or hi (z) − p∗i ≤ 0 of (7.1.2).

In the following, we briefly outline the main steps of the derivation of Fiacco
and Ishizuka (1990). Therefor we consider problem (7.1.2) as a parametric nonlinear
optimization problem of the general form

min F (r, p), (7.2.2a)

r

Gi (r, p) ≥ 0, ∀ i ∈ I, (7.2.2b)

where r ∈ RNn are variables and p ∈ RI are parameters. We note that r = (z, s, ∆) for
the relaxation (7.1.2). Furthermore we assume that F and G are twice continuously
differentiable, real valued functions.

A KKT point (r̂, η̂) of (7.2.2) with Lagrange multipliers η ∈ RI≥0 depends on the
chosen parameter p, hence we denote it by (r̂(p), η̂(p)), in order to emphasize this
dependency. We select a parameter p0 and obtain a solution r̂ = r̂(p0 ) by solving
(7.2.2). In order to estimate the change in the objective function when altering the
parameter p in a small neighborhood of p0 , we differentiate the objective function
F (r̂(p), p) with respect to p and obtain:

dF dr ∂F
(r̂, p0 ) = ∇r F (r̂, p0 ) (p0 ) + (r̂, p0 ). (7.2.3)
dp dp ∂p

Writing the Lagrange function of (7.1.2) as L(r, η, p) = F (r, p) − η G(r, p), it holds
∇r L(r̂, η̂, p0 ) = 0 by (2.4.2) for the KKT point (r̂, η̂, p0 ). Hence we can rewrite the
first summand on the right-hand side of (7.2.3) as

∇r F (r̂, p0 ) = η0 ∇r G(r̂, p0 ),

where Lagrange multipliers η0 are the values η̂ for the parameter choice p = p0 . We
use this to rewrite (7.2.3) as

dF dr ∂F
(r̂, p0 ) = η0 ∇r G(r̂, p0 ) (p0 ) + (r̂, p0 ). (7.2.4)
dp dp ∂p
186 Chapter 7 A Primal Heuristic based on Dual Information

We denote by I 0 ⊆ I the subset of active constraints from I for (r̂, p0 ), i.e., Gi (r̂, p0 ) =
0 for i ∈ I 0 . Further define G̃ := (Gi (r̂, p0 ))i∈I 0 . By Fiacco and Ishizuka (1990)
there exists a neighborhood of p0 such that G̃(r̂(p), p) = 0 for all p within this p0 -
neighborhood. This means, the function G̃(r̂(·), ·) is locally constant, which implies
that dp (r̂(p), p)
dG̃
= 0 for all p in the neighborhood of p0 . We compute this derivation
in p0 and obtain

dG̃ dr
0= (r̂, p0 ) = ∇r G̃(r̂, p0 ) (p0 ) + ∇p G̃(r̂, p0 ). (7.2.5)
dp dp

We put (7.2.5) into (7.2.4), make use of η0 i = 0, i ∈ I \ I 0 for the inactive constraints
and obtain
dF ∂F
(r̂, p0 ) = −η0 ∇p G(r̂, p0 ) + (r̂, p0 ). (7.2.6)
dp ∂p
Now we consider the derivation of the Lagrange function L with respect to p:

dL dr dη ∂L
(r̂, η0 , p0 ) = ∇r L(r̂, η0 , p0 ) (p0 ) + ∇η L(r̂, η0 , p0 ) (p0 ) + (r̂, η0 , p0 ).
dp | {z } dp } dp ∂p
=0,by KKT
| {z
=G(r̂,η0 ,p0 )=0

From
∂L ∂F
(r̂, η0 , p0 ) = −η0 ∇p G(r̂, p0 ) + (r̂, p0 )
∂p ∂p
and (7.2.6) we conclude that

dF dL
(r̂, p0 ) = (r̂, η0 , p0 ). (7.2.7)
dp dp

In order to make use of this equation, we first write down the Lagrange function
L(x∗ , z, s, ∆, p∗ , µ, η, λ, ξ, ζ) of (7.1.2), which is defined as follows:
 
n
L(x∗ , z, s, ∆, p∗ , µ, η, λ, ξ, ζ) = σf (x∗ , z) + (1 − σ)  (s+
j + sj ) + ∆i 
X X
−

j=1 i∈I 00

+ λ g(x∗ , z)
+ ηi (hi (z) − p∗i ) − η i ∆i
X X

i∈I 0 i∈I 00
n 
µ+ +

+ j (zj − sj − z j ) + µj (z j − zj − sj )
X
− −

j=1
X
− ξi
i∈I 0
7.2 A Primal Heuristic for MINLP with Indicator Constraints 187

n
(ζj+ + ζj− ).
X
−
j=1

Making use of equation (7.2.7) we obtain from this Lagrange function:

ˆ p∗ ) = −η̂i
∂pi f˜(x∗ , ẑ, ŝ, ∆, ∀ i ∈ I 0. (7.2.8)

This equation expresses the change of the objective function f˜, when parameter pi is
changed.

7.2.2 The Basic Dual Value MINLP Heuristic

In the previous section we derived relation (7.2.8) between the dual value of a
constraint of (7.1.2) and the impact on its objective by changing this constraint.
Making use of this relation our heuristic works as follows: An iterative process
is started at problem (7.1.2) based on the binary decisions x∗ . In each step the
heuristic selects one fixed binary variable and assigns a different binary value to it
(this corresponds to a flip, from upper to lower bound, or vice versa). This is done
as follows: First a KKT point of the relaxation (7.1.2) with parameters (x∗ , 0) is
computed. Then all the constraints (7.1.2c) and (7.1.2d) are ranked according to
the current values of the dual solutions, i.e., the right-hand sides in (7.2.8). This
ranking yields a ranking of those binary variables x which are associated by (7.1.1c)
and (7.1.1d). The variable corresponding to the constraint with the highest rank is
then the most promising candidate for an assignment of a different value. That is, if
η̂i ≥ max{η̂j : j ∈ I 0 } for some i ∈ I 0 , then the corresponding variable xi (which is
currently fixed to x∗i = 1) will be fixed to 0 in the next iteration. By most promising
we mean that for the new parameter x̃∗ for (7.1.2), which is obtained from x∗ by
changing a specific value, the optimal objective value of (7.1.2) is decreased. Recall
that a feasible solution for (7.1.2) with optimal objective value zero also yields a
feasible solution for (7.1.1) when neglecting the slack variables. When changing x∗
to x̃∗ we also ensure that (7.1.2) is feasible for the new parameters (x̃∗ , 0).
We also keep track of the variable which is changed, so that we do not repeat
the same decision again in later steps, in order to prevent the heuristic from cycling.
Note that cycles occur if (7.1.1) contains the relation x1 = 1 − x2 for two binary
variables x1 and x2 . In this case it might occur that both binary variables x1 and x2
are iteratively changed.
In the next iteration, we solve the slack model (7.1.2) again. Note that the set
I changes from one iteration to the next. The objective function values f˜ from
0
188 Chapter 7 A Primal Heuristic based on Dual Information

the previous iteration is compared with the objective function value of the current
iteration, denoted by f˜+ . We allow a slight increase (worsening) of at most 20 % (or
any other user-defined parameter) to accept this move. If the increase is more than
that, the heuristic terminates without any result. Otherwise the process is iteratively
continued.
Summing up, the heuristic selects one fixed binary variable in each step, and
assigns a different binary value to it. In the long run, the objective function value
of the slack model (7.1.2) typically decreases. Two cases can occur: First, at some
iteration, we reach a point where the slack part of the objective function value is 0.
In this case we obtain a feasible solution for (7.1.1) by neglecting the slack variables
and the heuristic terminates successfully. The second case that might occur is that
the slack part of the objective function does not converge to 0. If after a user-defined
number of rounds the slack part is still nonzero, then the heuristic terminates. Thus
it failed to construct a feasible solution.

7.2.3 Embedding the Heuristic in a Branch-and-Prune Search

The heuristic outlined above can be embedded in a tree search. It can run into a
dead end if the slack part of the objective does not converge to 0 after a certain
number of iterations. The idea of the tree search is to restart the heuristic at an
earlier stage and thus to cover a wider range of potential changes of binary variables.
To this end, we do not only consider the single best variable with respect to the
dual values as before. Instead, we use a small pool (of a user-defined size, typically 5
to 10 variables), and take the best dual variable until the pool is filled. Each iteration
of the heuristic is considered as a node in the tree (and the start as the root node),
and for each variable in the pool a child node is created and inserted in the tree.
We then traverse the tree in a depth-first search. If the heuristic fails, the node is
pruned and a back-tracking to the previous unpruned nodes takes place. An example
of this tree search is discussed in Example 7.3.2.

7.3 A Specialization to the Topology Optimization

Problem
We are going to apply the heuristic method to the natural gas network topology
optimization problem (3.2.1). We do not insist on the assumptions because our aim
is to develop a heuristic method. Recall that we assumed above that all constraints of
MINLP (7.1.1) are twice continuously differentiable. This is not the case for (3.2.1)
7.3 A Specialization to the Topology Optimization Problem 189

as the function q 7→ q|q|, for instance, is not C 2 . Nevertheless, we are going to apply
the heuristic to (3.2.1). In Section 7.3.1 we present the relaxation (7.1.2) in the
special case of solving (3.2.1). This relaxation is slightly different from the domain
relaxation (6.1.2). In Section 7.3.2 and Section 7.3.3 we present special adaptations
of the heuristic to the topology optimization problem. We note that the variables
(z, x) of (7.1.1) change their roles to (q, π, p, y, x) with z = (q, π, p, y) when turning
to (3.2.1).

7.3.1 The Relaxation

Let us write down the relaxation (7.1.2) for the topology optimization problem (3.2.1).
We define the set of arcs that are selected at the current node by A0 := {(a, i) ∈ AX |
xa,i = 1, i > 0}, i.e., A0 contains all arcs where the flow is not fixed to zero by
(3.2.1d). Furthermore let the set A00 := {(a, 0) ∈ AX | xa,0 = 1} contain all active
arcs that are in closed state. We define the index set J 0 appropriately such that the
flow and node potential variable bounds are relaxed. Furthermore we choose I 00 such
that all constraints (3.2.1c), that are active for the choice of x, are relaxed, too. We
introduce ∆v ∈ R≥0 for all v ∈ V and ∆a ∈ R≥0 and ∆ ˜ a ∈ Rνa for all a ∈ A0 as
≥0
slack variables. Then, for σ = 0, the relaxation (7.1.2) writes as follows:

νa
min ∆v + ∆a + ˜ a)
(∆ (7.3.1)
X X XX
k
v∈V a∈A0 a∈A0 k=1

s. t. αa qa |qa |ka − βa ya − (πv − γa πw ) = 0 ∀ a = (v, w) ∈ A0 ,

qa = 0 ∀ a = (v, w) ∈ A00 ,
T ˜ a ≤ ba
Aa (qa , pv , pw ) − ∆ ∀ a = (v, w) ∈ A0 ,

qa = dv
X X
qa − ∀ v ∈ V,
a∈δ +0 0 (v) a∈δ −0 0 (v)
A ∪A A ∪A
0 0

|pv |pv − πv = 0 ∀ v ∈ V,

π v − ∆v ≤ π v ∀ v ∈ V,

πv + ∆ v ≥ π v ∀ v ∈ V,

q a − ∆a ≤ q a ∀ a ∈ A0 ,

qa + ∆ a ≥ q a ∀ a ∈ A0 ,
190 Chapter 7 A Primal Heuristic based on Dual Information

ya ≤ y a ∀ a ∈ A0 ,

ya ≥ y a ∀ a ∈ A0 ,

πv ∈ R ∀ v ∈ V,

qa ∈ R ∀ a ∈ A0 ∪ A00 ,

ya ∈ R ∀ a ∈ A0 ,

∆v ∈ R≥0 ∀ v ∈ V,

∆a ∈ R≥0 ∀ a ∈ A0 ,
˜ a ∈ Rν a
∆ ∀ a ∈ A0 .
≥0

We note that this relaxation equals the domain relaxation of the active transmission
problem (6.1.2) except the additional flow variables qa for a ∈ A00 and the constraints
qa = 0 for these arcs. So this relaxation is always feasible by Lemma 6.1.1 which
allows to proceed as already described in Section 7.2.2.

Remark 7.3.1:
Note that the domain relaxation (6.1.2) of the active transmission problem (6.1.1)
turned out to be more efficient than the flow conservation relaxation (6.1.5). Hence
we do not investigate further whether a similar heuristic process based on the flow
conservation relaxation (6.1.5) could be derived.

7.3.2 Handling Different Modes of Active Devices

When applying the heuristic described in Section 7.2 we identify binary variables that
are switched to different values. For the topology optimization problem (3.2.1) some
of these binary variables correspond to different modes of the active elements in the
network. Recall that a valve for instance can be open or closed. A typical situation
in the network is a compressor (or a control valve) that can be active (compressing
or regulating), in bypass, or closed. Let us briefly recall each state.

• If the compressor is active, then the gas can flow through the compressor. In
this case, the amount of flow is restricted by lower and upper bounds. The
output pressure at the exit side of the compressor is higher than the input
pressure, and the compression ratio between output and input pressure has
to satisfy certain bounds. The operating range of a compressor a ∈ A is thus
described by νa linear inequalities in the system (Aa , ba ), see (3.2.1c). They
7.3 A Specialization to the Topology Optimization Problem 191

pw
pv

2.2, 1.2

Active
Closed (qa = 0)

2.1, 3.2
1.1, 3.1
qa
Bypass (pw /pv = 1) pv

Figure 7.1: Shown are the different operation modes of a compressor. The arrows indicate the changes to
binary variables for a compressor, indicated by the heuristic, labeled by cases, see Section 7.3.2. The gray
shaded ellipse marks those parts of the states bypass and closed in which changes between the modes
bypass and closed are possibly performed by the heuristic.

define a set of feasible points illustrated by the shaded area in the upper-right
of Figure 7.1.

• If the compressor is in bypass, then again, flow through the compressor is

allowed. Furthermore the pressure at the input side is identical to the pressure
at the output side, and the flow can vary arbitrarily (see the horizontal line in
Figure 7.1).

• Finally, if the compressor is in closed mode, then the flow is zero, and the
pressures at input and output side can take arbitrary values (see the vertical
line in Figure 7.1).

The situation for a control valve is similar, see Figure 7.2. The only difference is that
the output pressure is lower than the input pressure in ’active’ mode.
We consider a compressor or a control valve a = (v, w) ∈ A. Recall that xa,1 , xa,2
are binary variables associated with the compressor (or control valve) denoting the
bypass and active state respectively. The binary variable xa,0 is used to represent the
case that the active element is closed. The crucial part of the topology optimization
192 Chapter 7 A Primal Heuristic based on Dual Information

pv − pw

2.2, 1.2

Active
Closed (qa = 0)

2.1, 3.2
1.1, 3.1

qa
Bypass (pv = pw )

Figure 7.2: Shown are the different operation modes of a control valve. The arrows indicate the changes
to binary variables for a compressor, indicated by the heuristic, labeled by cases, see Section 7.3.2. The
gray shaded ellipse marks those parts of the states bypass and closed in which changes between the modes
bypass and closed are possibly performed by the heuristic.

model that we will exploit in the sequel are the following three families of indicator
constraints:

qa,1 = 0 (7.3.2a)
(
xa,0 = 1 ⇒
qa,2 = 0 (7.3.2b)

πv − πw = 0 (7.3.3a)
(
xa,1 = 1 ⇒
qa,2 = 0 (7.3.3b)

 Aa (qa,2 , pv , pw )T ≤ ba (7.3.4a)
xa,2 =1⇒
 qa,1 = 0 (7.3.4b)

Recall that exactly one of these three cases applies because of (3.2.1e).
The heuristic creates its own branching tree T as described in Section 7.2.3. Its
root node consists of some fixation of the discrete decision variables to binary values.
We solve the slack variant (7.3.1) of model (3.2.1). We obtain a ranking among the
binary variables (according to the dual solution values) as described in Section 7.2.2.
This ranking indicates which binary variable should be flipped from one to zero.
Changing the state of a compressor or a control valve goes along with changing the
7.3 A Specialization to the Topology Optimization Problem 193

values of two binary variables. Hence another binary variable must then be flipped
from zero to one in order to preserve the feasibility of (7.3.1).
Denote by ρf a certain user-specified threshold value for the flow (default value
is 20), and by ρp a threshold value for the pressure difference (default value is 10).
Using these values new child nodes for tree T are created where we distinguish the
following cases.

Case 1: xa,0 from 1 to 0. The subnet is currently “closed”. The ranking indicates
that either the variable xa,1 or xa,2 , having the present value 0, should be set
to 1.
Case 1.1: “From closed to bypass.” If the pressure difference at both end nodes
|pw − pv | is below ρp , then a new child node is created where xa,1 is set
from 0 to 1 and xa,2 remains at 0.
Case 1.2: “From closed to active.” If the dual value of at least one of the
constraints (7.3.2a) or (7.3.2b) is positive, then in the new child we flip
xa,2 from 0 to 1 and xa,1 remains at 0.

Case 2: xa,2 from 1 to 0. The subnet is currently “active”. We collect all constraints
of (7.3.4a) that are either fulfilled with equality or where the associated value
of ∆ is greater than zero. (Recall that we solve a slack model, which ensures
feasibility.) Denote by I ⊆ {1, . . . , νa } the corresponding index set. For each
i ∈ I we consider the i-th constraint in (7.3.4a) and denote by (a1 , a2 , a3 ) the
coefficients from the i-th row of Aa . We fix the input pressure pv = 1 and check
in which quadrant the remaining two-dimensional vector (a2 , a3 ) lies.
Case 2.1: “From active to bypass.” If a2 < 0, then we create a child node in
which we set xa,2 to 0 and xa,1 to 1.
Case 2.2: “From active to closed.” If a1 < 0, then we create a child node in
which we set xa,2 to 0 and also xa,0 to 1.

Case 3: xa,1 from 1 to 0. The subnet is currently in “bypass”.

Case 3.1: “From bypass to closed.” If the absolute value of flow is below ρf ,
then a new child node is created with the decision to set xa,1 to 0, to leave
xa,2 at 0, and thus set xa,0 to 1. Hence at this child node, the subnetwork
is ’closed’.
Case 3.2: “From bypass to active.” If the dual value of the constraint (7.3.3a)
is negative, then a new child node is created, in which we set xa,2 from 0
to 1 and xa,1 from 1 to 0.
194 Chapter 7 A Primal Heuristic based on Dual Information

Otherwise: Consider the next best variable according to the ranking. If no more
variable exists, no other child node is created.

Note that multiple cases can occur simultaneously, hence it is possible that multiple
child nodes are created.
The “otherwise” case can occur for example if the ranking favors an increase of
compression for an active compressor, while the compression ratio is already at its
upper limit.
Regarding control valves we create new child nodes in the very same way as we
do for compressors. For simple (non-control) valves, the state “active” does not exist;
they can only be in “bypass” or “closed” mode. So we apply the same child node
creation routine as for compressors but we skip those parts that are concerned with
“active” states.
After inserting the new child nodes to the tree T we perform the following node
selection strategy in order to decide where to continue the search. We always proceed
according to the rule: “active” first, “bypass” second, “closed” third. We always pass
along from a parent node to one of its child nodes according to this rule (depth-first
search). When entering the child node again the slack model (7.3.1) is solved first.
This yields a new ranking of the decision variables, which leads to a new decision
about which of the variables should be flipped in the next round.
The branching process terminates if at one node the slack model (7.3.1) has a
zero objective function value or a certain user-defined branching depth (i.e., number
of nodes in the path to the root node) is reached. Then the corresponding child
node is pruned and the search procedure continues at another unfinished child node.
Here we apply a simple backtracking rule, going back to the previous parent node
where we came from before. An example for the branching tree T is given in the
next section.

7.3.3 Handling Loop Extensions

Our topology optimization model (3.2.1) allows a single original arc a = (v, w) ∈ A
together with additional parallel arcs (a, i) ∈ AX , i > 1. For these arcs the important
part of the model consists of the following constraints:

xa,i = 1 ⇒ αa,i qa,i |qa,i |ka − βa,i ya,i − (πv − γa πw ) = 0 ∀ (a, i) ∈ AX , i 6= 0,

xa,i = 0 ⇒ qa,i = 0 ∀ (a, i) ∈ AX , i 6= 0,

xa,i = 1.
X

i:(a,i)∈AX
7.3 A Specialization to the Topology Optimization Problem 195

Recall that the extended arc set AX does not contain (a, 0) as we consider a passive
network element here. Thus, if xa,i is chosen to be switched from 1 to 0 by our
heuristic, then the constraint αa,i qa,i |qa,i |ka − βa,i ya,i − (πv − γa πw ) = 0, which is
active in the current relaxation (7.3.1) becomes inactive in the next problem. When
writing the relaxation (7.3.1) this constraint is expressed as

αa,i qa,i |qa,i |ka − βa,i ya,i − (πv − γa πw ) ≤ 0, (7.3.5a)

−αa,i qa,i |qa,i | ka
+ βa,i ya,i + (πv − γa πw ) ≤ 0. (7.3.5b)

Now we distinguish two cases:

Case 1: The dual value of constraint (7.3.5a) is positive. This means that the
difference of node potential values is too large for the solution flow value. If the
constant αa,i is not at its minimum, i.e., xa,j = 0 for a maximal index j < i
with (a, j) ∈ AX , and the compression ya,i is maximal, i.e., ya,i = y a,i , then
we flip xa,i from 1 to 0 and xa,j from 0 to 1 in the new child.

Case 2: The dual value of constraint (7.3.5b) is positive. This reflects that the
difference of potential values is too small for the solution flow value. If the
constant αa,i is not at its maximum, i.e., xa,j = 0 for a minimal index j > i
with (a, j) ∈ AX , and the compression ya,i is minimal, i.e., ya,i = y a,i , we flip
xa,i from 1 to 0 and xa,j from 0 to 1 in the new child node.

Otherwise: The dual values of (7.3.5a) and (7.3.5b) are zero. Consider the next best
variable according to the ranking. If no more variables exist, a child node is
not created.

Now let us explain the branching tree T as an example.

Example 7.3.2:
An example for the branching process and the corresponding branching tree T is
shown in Figure 7.3. Here we consider a network that consists of one control valve
(CV), one compressor station (CS), and a pipe, which can be extended by a parallel
pipe (loop). Assume that either from the LP relaxation or by branching at the current
node the CV is set to active mode and the CS is set to bypass mode and no extension
for the pipe is selected. Assume as well that this selection is not feasible. Now the
heuristic is invoked. It creates a ranking of the corresponding decision variables,
which means, a ranking of the network elements. We assume that the CV is ranked
higher than the CS while the pipe has the lowest rank. Hence the heuristic creates a
branch (left son), where it changes the stage of the CV from active to bypass first
196 Chapter 7 A Primal Heuristic based on Dual Information

CV (active) | CS (bypass) | loop 1x

CV (bypass) | CS (bypass) | 1x CV (closed) | CS (bypass) | 1x

CV (bypass) | CS (active) | 1x CV (bypass) | CS (closed) | 1x

CV (bypass) | CS (active) | 2x CV (bypass) | CS (closed) | 2x

Figure 7.3: Example of the branching tree created by the heuristic for a network that contains 1 control
valve (CV), 1 compressor station (CS) and a pipe that can be extended by a parallel pipe (loop). The
tree is discussed in Example 7.3.2.

(according to Case 2.1 in Section 7.3.2). Another branch (right son) is created where
it changes the stage of the CV from active to closed. Now we focus on the left son and
solve the NLP (7.1.2) again and obtain a new ranking. Among all elements that have
not been changed before, the CS has the highest ranking now. Hence it is changed,
once from bypass to active (left son) and once from bypass to closed (right son).
For the left son, we solve the NLP relaxation (7.1.2) and assume that it has still a
positive slack value. Now the loop is selected in a new child of the tree (according to
Case 1 in this section) and the NLP relaxation is solved. We assume that it still has
a positive slack value. Since we cannot alter more elements in this small network,
this part of the tree is finished, and we track back to the parent node and once more
to the next parent node. For the right son, we solve the NLP relaxation and again
get an infeasible subproblem, so we also consider the child in the next step. Now
we assume the corresponding NLP has a positive slack value again. Then we prune
this node and afterwards are back at the root node of the search tree created by the
heuristic. The sub tree where the CV was changed from active to bypass is finished,
hence we now change the CV from active to closed (shown on the right). Then we
solve the NLP (7.1.2) and assume, that we obtain a solution with slack zero. At this
point, when neglecting the slack variables, a new primal feasible solution is found for
(7.3.1).

7.4 Implementation Details

Our heuristic is implemented as a plug-in for the solver SCIP. The following settings
are offered to the user:
7.4 Implementation Details 197

Maxcalls. Specifies the maximal number of nodes for the branching tree within the
heuristic. (Default value is 30.)

Maxequals. Abort the heuristic without a solution, if the highest-ranked “Maxequals”

fraction of binary variables all have equal ranks. (Default value is 1/3, i.e.,
33.33 % of all binary variables.)

Mingap. Do not call the heuristic, if the gap is below the threshold value “Mingap”.
The gap is defined as: (best primal − best dual) / (best dual). (Default value
is 0.05, i.e., a SCIP gap of 5 %.)

Sigma. Sets the value for σ in objective (7.1.2a). (Default value is 0, i.e., emphasis
only on feasibility.)

Leaves only. Call the heuristic only at nodes of the branching tree where all binary
variables are fixed. (Default value is “No”.)

We apply the heuristic to the topology optimization problem (3.2.1). The heuris-
tic is invoked at those nodes in the branch-and-bound tree where all binary decision
variables have binary values (either by branching or by the solution of the LP relax-
ation). In order to make this happen as early as possible, we introduce a branching
priority rule that first branches on integral variables, before spatial branching on
continuous variables is applied. Furthermore, when changing x∗ in our heuristic to
x0 in the next iteration, we ensure that x0 is feasible for the constraints L x ≤ 0, see
(3.2.1h). This is done as follows: Recall that the left-hand side of (3.2.1h) is a vector
of linear functions. We assume that xa0 ,i has to be changed from zero to one. Then
we solve the following MIP problem to obtain the next vector x0 :

min ||x0 − x∗ ||1

s. t. L x0 ≤ t,
(7.4.1)
x0a0 ,i = 1,
x0 ∈ {0, 1}AX .

If this problem is infeasible, then we prune the current node of the branching tree
of our heuristic. Otherwise we obtain an optimal solution x0 which has minimal
deviations from x∗ together with the change xa0 ,i = 1.
198 Chapter 7 A Primal Heuristic based on Dual Information

7.5 Computational Results

We implemented the heuristic described above in C and use the computational setup
described in Section 3.5. For our computational study we considered the network
net7. The initial network net7a consists of 4165 nodes and 4079 arcs, among them
3638 passive pipelines, 12 compressor stations, 121 control valves, and 308 valves.
The industry partner defined a test set of 30 nominations for net7a that differ among
each other in the pressure and flow bounds on the entry and exit nodes. The problem
is a feasibility problem only (i.e., does there exist a feasible flow in the network).
Hence it is considered as “solved” as soon as a feasible solution is constructed; no
objective function values for opening valves (i.e., adding further topology extensions)
are set. We imposed a time limit of 4 hours for all instances which is still more
than the cooperation partner suggests. We compare three strategies for solving the
topology optimization problem (3.2.1).

1. “SCIP/default”: The first strategy is to use SCIP without any adaptations on

the solver settings. All branching decisions are up to the solver. We enforce a
certain branching priority rule, so that SCIP first branches on binary decision
variables x. Only after all discrete variables are fixed it is allowed to perform
spatial branching on continuous variables.

2. “SCIP/dualval(basic)”: The second strategy is to use the heuristic as described

in Section 7.2. We establish the branching tree T as described in Section 7.3.2
and Section 7.3.3, but neglect the special analysis which sorts out some changes
as described in the case discussion in Section 7.3.2.

3. “SCIP/dualval(adapted)”: The third strategy is to use the heuristic as de-

scribed in Section 7.2. Additionally we use the special adaptations described in
Section 7.3.2 and Section 7.3.3.

These strategies are used to demonstrate the impact of our heuristic on the solving
performance of SCIP. Additionally, we use Baron and Antigone to solve the
instances.
A brief summary of the computational results is shown in Table 7.1. A detailed
presentation of the results is available in Tables A.14-A.18. The third and fourth
column of Table 7.1 show that neither Baron nor Antigone solved any of the
test instances (see also Table A.1 and Table A.3). The fifth column of Table 7.1
shows the number of instances that were solved with SCIP/default, strategy 1. The
sixth column of Table 7.1 shows the number of instances that were solved by the
heuristic (SCIP/dualval (basic), strategy 2). The last column of Table 7.1 shows
7.5 Computational Results 199

network
SCIP
Baron Antigone
strategy 1 strategy 2 strategy 3
net7a - - 4 23 30
net7b - - 4 12 18
net7c - - 1 9 18
net7d - - - 5 17
net7e - - 1 8 18

Table 7.1: Globally solved instances out of 30 nominations for each of the networks net7a-net7e. The
underlying data are available in Tables A.14–A.18.

the number of instances that were solved by the heuristic exploiting the problem
structure (SCIP/dualval (adapted), strategy 3). Thus, for example SCIP/dualval
(basic) was able to find a feasible solution within 4 hours in 23 of the net7a-type
instances.
The modified version of the heuristic is able to solve all problem instances for
net7a. In later phases of our collaboration (net7b–net7e), the industrial project
partner made minor alterations to the networks’ topologies, and, more important,
also made the instances more and more difficult by imposing tighter bounds on the
minimum and maximum pressure levels at the nodes. Hence it became more difficult
for any heuristic to construct feasible solutions. However, the adapted heuristic
was still able to solve globally approximately 67 % of the instances (101 out of 150),
whereas the standard MINLP heuristics of SCIP found less feasible solutions the
more difficult the instances became (6 % of the instances of the test set are globally
solved, 10 out of 150). The basic heuristic (strategy 2) that is not adapted to the
problem’s special structure is also pretty good in finding feasible solutions (when
compared to SCIP/default, strategy 1). Here approximately 38 % of the instances of
the test set are globally solved (57 out of 150).
Figure 7.4 shows a comparison of SCIP/default with default settings (in particular,
with all standard heuristics) and SCIP/dualval(adapted) with the additional adapted
heuristic on all instances from Table 7.1. It turns out that for those instances which
SCIP/default is already able to solve, the additional time spent in the dual value
heuristic does not pay off, but slows down the overall solution process instead. For
many of those instances that SCIP/default could not solve, SCIP/dualval(adapted)
was able to compute a feasible solution within the time limit.
We also applied the heuristic in its adapted version to network net5. Here we
combined it with the solution approach indicated by strategy 3 presented in the
previous chapter. The corresponding instances are topology expansion problems.
Our computational results show that a primal feasible solution is available for every
200 Chapter 7 A Primal Heuristic based on Dual Information

104

Run time strategy 3 (sec)

103

102

102 103 104

Run time strategy 1 (sec)

Figure 7.4: Run time comparison between SCIP/default (strategy 1) and SCIP/dualval(adapted) (strat-
egy 3). Each cross (×) marks the run time of these two solvers. A cross above the diagonal line
indicates that the strategy SCIP/default is faster, and a cross below the diagonal line indicates that
SCIP/dualval(adapted) is faster. Note that there are instances, where both strategies did not find any
solution. Each of these instances correspond to the same cross (×) mark in the upper right corner of
Figure 7.4. The time limit was set to 4 hours.

instance. The results are available in Table A.19. Without our heuristic a feasible
solution was available for approximately 45 % of the instances only, see Table A.13.

Summary

We presented a new heuristic algorithm that exploits dual information coming

from KKT points in order to find feasible solutions for the mixed-integer nonlinear
optimization problem (7.1.1) with indicator constraints. We applied this heuristic
to the topology optimization problem (3.2.1). Using the heuristic in the standard
fashion, it is better than the existing heuristics that are already available in the solver
SCIP. After exploiting the problem structure, the heuristic is able to identify many
more feasible solutions than any other heuristic included in SCIP that we are aware
of. In total approximately 61 % more instances of the test set consisting of instances
of the network net7 are globally solved in total. We also applied the heuristic to the
network net5 and combined it with the solution approach presented in the previous
chapter. The computational results showed that the number of instances with a
primal feasible solution available was increased by 55 %. Recall that both networks
7.5 Computational Results 201

out of those considered in this thesis remained with very few feasible solutions after
the solution approaches of the previous Chapters 4-6.
Chapter 8

Conclusions

This thesis deals with gas network optimization, that is nomination validation and
topology expansion problems arising from gas transport as introduced in Chapter 1.
We presented a mixed-integer nonlinear program (3.2.1) that models both problems at
the same time and considered it as the topology optimization problem in gas networks.
Our industrial cooperation partner provided us with data of different networks in
combination with corresponding nominations. Facing the fact that state-of-the-art
solvers like Baron, Antigone or SCIP are slow on real-world instances or unable
to compute any primal feasible solution, we set our aim to improve the performance
of SCIP for solving our model. Exploiting specific knowledge of our model was
the key to success. Our strategy shows that an MINLP solver in combination with
special-tailored adaptations offers the potential to solve globally large-scale non-
convex MINLP problems of real-world size. We believe that our results presented in
this thesis provide evidence for the suitability of the approach and constitute a step
in the right direction of solving the topology optimization problem.

Future studies aim at solving the topology optimization problem for transient
gas flows. In the focus of the study there is the identification of more general
conditions for the passive transmission problem. Assuming that the convex domain
relaxation of the passive transmission problem is performed, an interpretation of
the dual solution of the relaxation would lead to a primal heuristic for identifying
feasible solutions. Discretization of the time component in the model for transient
flows allows the inclusion of the results obtained. Apart from this, an extension of
the pc-regularization towards time dependency should be investigated in order to
derive cuts by an augmented Benders argument. Finally, this approach used for gas
transmission networks could also be extended to water or electricity networks, under
some additional assumptions.

203
204 Chapter 8 Conclusions

Benefits for a TSO

Let us briefly discuss different aspects which allow a TSO to improve the operation
of its network. Utilizing the algorithms developed in this thesis allows the company
to compute operation modes for the active network elements, arc flows, and node
pressures for a given nomination. This computation requires that the flow specified
by the nomination is transported through the network and all technical and physical
constraints are fulfilled. Here the software can be used to assist a manual approach of
expert knowledge in combination with simulation software. Given a set of extensions
the model can also be used for the topology expansion problem.
Other aspects are also useful for a TSO. The company is able to get a visualization
of the pressure distribution, see Figure 1.3. Given the case that the simulation software
is not delivering a feasible solution the possibility of proving global infeasibility of an
active transmission problem as described in Section 6.2 improves this situation. A
failing simulation software does not generally imply that the current transportation
situation is indeed infeasible. Especially a visualization as shown in Figure 4.2
provides assistance in comprehending the reasons of an infeasibility. The picture
emphasizes those parts of the network that cause the infeasibility of a passive
transmission problem as described in Section 4.2.3. Furthermore considering an
infeasible passive transmission problem and solving the flow conservation relaxation
yields a set of specific slack values. These values allow to circumvent infeasibilities.
More precisely, an adaptation of the considered nomination can be performed such
that the corresponding gas quantities are transported through the network, i.e., the
passive transmission problem is feasible for the adapted nomination.
Appendix A

Tables

nom net7a net7b net7c net7d net7e

4 limit limit limit 647 limit

6 limit limit limit 912 limit
8 limit limit 725 706 limit
19 limit limit limit limit 663
20 limit limit 564 limit limit
23 limit limit limit 3,029 limit
24 limit limit limit limit 724
25 limit limit 571 limit 996
26 limit limit limit 574 limit
27 limit limit limit 589 limit
29 limit limit limit 730 limit

Table A.1: Run time results in seconds using Baron to solve the topology optimization problem on 30
nominations on the network net7. The time limit was set to 4 hours and the results are discussed in
Chapter 7. All nominations not depicted here ran into the time limit without a feasible solution. Those
instances with a finite time limit were detected to be infeasible. No primal solution was available for all
the other instances.

207
208 Appendix A Tables

nom net7a net7b net7c net7d net7e

3 limit 103 limit limit limit
4 limit 110 limit limit limit
5 78 limit limit limit limit
11 limit limit limit limit 132
15 limit limit limit limit limit
16 966 limit limit limit limit
17 109 limit limit limit limit
22 limit 297 161 limit limit
27 779 4,171 limit limit limit

Table A.2: Run time results in seconds using SCIP to solve the topology optimization problem on 30
nominations on the network net7. The time limit was set to 4 hours and the results are discussed in
Chapter 7. All nominations not depicted here ran into the time limit without a feasible solution. Those
instances with a finite time limit were feasible. No primal solution was available for all the other instances.
 209

nom net7a net7b net7c net7d net7e

1 37 37 1,808 limit 375
2 27 27 6,665 13,825 9,221
3 62 93 711 limit 3,064
4 28 28 50 4,859 52
5 28 28 50 1,874 52
6 28 29 50 49 52
7 27 28 50 585 52
8 28 28 51 4,009 55
9 28 28 49 50 53
10 28 38 5,971 52 limit
11 37 27 131 351 304
12 27 28 49 50 51
13 28 27 51 457 52
14 28 28 50 49 51
15 28 27 50 50 52
16 28 28 49 49 52
17 28 27 50 52 52
18 7,985 38 limit limit 51
19 37 28 49 limit 1,542
20 28 28 50 60 52
21 28 28 49 53 52
22 27 27 49 49 52
23 27 27 49 55 51
24 27 28 49 49 51
25 28 28 51 limit 52
26 28 27 51 209 52
27 28 28 50 50 53
28 27 28 49 50 51
29 33 14,205 52 51 53
30 27 28 48 52 limit

Table A.3: Run time results in seconds using Antigone to solve the topology optimization problem on
30 nominations on the network net7. The time limit was set to 4 hours and the results are discussed in
Chapter 7. Those instances with a finite time limit were detected to be infeasible. No primal solution
was available for all the other instances.
 210

domain relaxation (strategy 3) flow cons. relaxation (strategy 4) SCIP prio (strategy 2) SCIP (strategy 1)
nom gap primal dual time nodes gap primal dual time nodes gap primal dual time nodes gap primal dual time nodes
1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
2 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
3 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
4 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
5 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
6 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
7 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
8 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
9 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
10 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
11 - 95 95 6 105 - 95 95 6 105 - 95 95 7 780 - 95 95 6 1,087
12 - 95 95 1 1 - 95 95 1 1 - 95 95 1 1 - 95 95 1 1
13 - 95 95 1 1 - 95 95 1 1 - 95 95 1 1 - 95 95 1 1
14 - 95 95 1 1 - 95 95 1 1 - 95 95 1 1 - 95 95 1 1
15 - 95 95 1 1 - 95 95 1 1 - 95 95 1 1 - 95 95 1 1
16 - 249 249 7 303 - 249 249 8 303 - 249 249 1,483 763,842 - 249 249 33,270 19,155,684
17 - 320 320 9 312 - 320 320 12 312 - 320 320 68 24,322 - 320 320 12,773 7,225,869
18 - 475 475 11 445 - 475 475 32 445 - 475 475 69 25,140 636 700 95 limit 14,090,518
19 - 525 525 65 3,750 - 525 525 252 3,594 475 limit 22,919,263 160 743 285 limit 16,516,334
20 - 525 525 36 2,957 - 525 525 100 2,957 - 525 525 64 19,410 - 525 525 76 19,323
21 - 525 525 3 173 - 525 525 5 173 - 525 525 2 173 - 525 525 2 173
22 - 525 525 1 12 - 525 525 1 12 - 525 525 1 12 - 525 525 1 12
23 - 525 525 1 6 - 525 525 1 6 - 525 525 1 6 - 525 525 2 6
24 - 525 525 1 6 - 525 525 1 6 - 525 525 1 6 - 525 525 1 6
25 - 525 525 1 6 - 525 525 1 6 - 525 525 1 6 - 525 525 1 6
26 - 584 584 15 1,534 - 584 584 27 1,526 - 584 584 76 23,830 - 584 584 24,658 9,692,200
27 - 633 633 2 23 - 633 633 3 23 - 633 633 2,999 1,082,717 29 679 525 limit 13,668,016
28 - 633 633 47 3,122 - 633 633 143 3,216 - 633 633 46 11,870 72 905 525 limit 15,964,903
29 - 693 693 33 1,428 - 693 693 135 1,764 - 693 693 371 105,119 114 1,125 525 limit 15,632,260
30 - 788 788 28 893 - 788 788 154 2,382 - 788 788 6,589 2,403,190 78 938 525 limit 16,565,148
31 - 997 997 38 1,263 - 997 997 355 5,016 37 1,085 788 limit 21,938,417 143 1,279 525 limit 16,769,864
32 - 1,092 1,092 118 6,456 - 1,092 1,092 419 6,456 57 1,092 693 limit 13,526,935 180 1,476 525 limit 11,628,716
33 - 1,171 1,171 379 16,938 - 1,171 1,171 1,268 18,758 67 1,319 788 limit 7,929,185 125 1,318 584 limit 6,976,475
34 - 1,174 1,174 257 12,370 - 1,174 1,174 871 12,580 104 1,614 788 limit 9,530,654 175 1,612 584 limit 11,104,065
35 - 1,182 1,182 64 4,072 - 1,182 1,182 849 12,888 28 1,254 978 limit 16,014,817 199 1,896 633 limit 11,797,821
36 - 1,234 1,234 200 14,073 - 1,234 1,234 445 10,100 23 1,234 997 limit 8,524,150 98 1,259 633 limit 15,467,469
37 - 1,241 1,241 80 7,412 - 1,241 1,241 511 12,466 65 1,812 1,092 limit 4,921,651 95 1,241 633 limit 11,612,658
38 - 1,318 1,318 64 7,666 - 1,318 1,318 228 7,666 10 1,332 1,202 limit 26,954,607 110 1,332 633 limit 19,373,729
39 - 1,336 1,336 194 19,978 - 1,336 1,336 404 16,302 7 1,336 1,237 limit 11,395,890 241 2,164 633 limit 10,694,870
40 - 1,459 1,459 158 20,362 - 1,459 1,459 826 29,252 22 1,620 1,325 limit 14,668,683 105 1,620 788 limit 13,737,907
41 15 1,545 1,332 limit 3,223,996 - 1,545 1,545 1,269 45,195 15 1,545 1,332 limit 8,085,049 86 1,863 997 limit 15,256,088

Table A.4: Results on network net4 and 41 nominations. A time limit of 11 h was imposed. The different strategies for solving the topology optimization problem
are presented in Chapter 4.
Appendix A Tables
 domain relaxation (strategy 3) flow cons. relaxation (strategy 4) SCIP prio (strategy 2) SCIP (strategy 1)
nom gap primal dual time nodes gap primal dual time nodes gap primal dual time nodes gap primal dual time nodes
1 - 555 555 152 690 - 555 555 249 690 194 1,314 446 limit 3,275,942 116 729 337 limit 4,273,098
2 - 701 701 1,090 26,834 - 701 701 3,186 27,968 - 701 701 1,531 183,041 148 836 337 limit 2,417,953
3 - 1,429 1,429 4,375 32,660 - 1,429 1,429 9,499 34,675 952 limit 5,261,200 103 1,429 701 limit 2,236,848
4 - 1,495 1,495 2,303 38,971 - 1,495 1,495 6,699 38,971 1,038 limit 4,800,505 73 1,799 1,038 limit 2,206,174
5 - 1,731 1,731 11,698 80,143 - 1,731 1,731 27,220 80,143 1,038 limit 3,953,611 1,038 limit 1,410,299
6 98 2,575 1,300 limit 298,530 107 2,575 1,242 limit 133,550 1,257 limit 4,282,280 1,038 limit 1,914,502
7 82 2,634 1,443 limit 324,788 90 2,634 1,384 limit 152,222 1,604 limit 4,225,594 1,148 limit 1,323,814
8 118 3,575 1,635 limit 664,458 166 4,130 1,548 limit 270,658 1,713 limit 5,028,474 1,384 limit 3,144,894
9 236 6,553 1,948 limit 387,517 250 6,553 1,871 limit 221,056 1,995 limit 4,244,062 1,666 limit 847,847
10 179 6,154 2,199 limit 734,553 2,147 limit 514,498 1,995 limit 3,951,594 1,995 limit 2,059,568
11 2,401 limit 2,049,433 2,337 limit 1,195,274 2,343 limit 5,293,619 2,130 limit 2,739,786

Table A.5: Results on network net5 and 11 nominations. A time limit of 11 h was imposed. The different strategies for solving the topology optimization problem
are presented in Chapter 4.
211
 212

SCIP (strategy 1) cutoff (strategy 2)

scale loops gap primal dual time nodes gap primal dual time nodes
2.0 8 - 1,500 1,500 212 176,615 - 1,500 1,500 255 212,747
2.0 9 - 1,500 1,500 113 83,832 - 1,500 1,500 157 114,824
2.0 10 - 1,500 1,500 140 100,641 - 1,500 1,500 133 94,086
2.0 11 - 1,500 1,500 130 91,718 - 1,500 1,500 139 94,076
2.0 12 - 1,500 1,500 132 91,469 - 1,500 1,500 141 88,662
2.1 8 - 1,800 1,800 2,127 1,034,782 - 1,800 1,800 1,955 1,090,390
2.1 9 - 1,800 1,800 859 625,947 - 1,800 1,800 771 524,701
2.1 10 - 1,700 1,700 719 444,753 - 1,700 1,700 807 480,961
2.1 11 - 1,700 1,700 774 504,599 - 1,700 1,700 926 552,831
2.1 12 - 1,700 1,700 728 471,606 - 1,700 1,700 912 533,415
2.2 8 - 2,200 2,200 7,069 4,521,738 - 2,200 2,200 7,267 4,211,609
2.2 9 - 2,000 2,000 3,874 2,274,304 - 2,000 2,000 4,148 2,282,214
2.2 10 - 2,000 2,000 2,594 1,441,304 - 2,000 2,000 2,339 1,339,507
2.2 11 - 2,000 2,000 2,919 1,713,374 - 2,000 2,000 2,590 1,580,747
2.2 12 - 2,000 2,000 2,699 1,494,467 - 2,000 2,000 3,562 1,866,970
2.3 8 63 3,100 1,900 limit 10,801,193 61 3,000 1,858 limit 10,007,510
2.3 9 18 2,400 2,028 limit 13,912,926 - 2,400 2,400 38,627 14,172,581
2.3 10 - 2,300 2,300 16,496 6,038,279 - 2,300 2,300 11,947 5,797,020
2.3 11 - 2,200 2,200 5,380 3,104,723 - 2,200 2,200 6,658 3,525,269
2.3 12 - 2,200 2,200 9,881 4,006,313 - 2,200 2,200 10,991 4,547,700
2.4 8 133 5,900 2,523 limit 4,451,666 155 6,200 2,425 limit 4,308,121
2.4 9 67 3,200 1,914 limit 5,744,040 60 3,100 1,933 limit 8,051,340
2.4 10 39 2,700 1,929 limit 10,522,647 40 2,700 1,916 limit 8,037,016
2.4 11 33 2,600 1,948 limit 12,099,342 33 2,600 1,952 limit 13,062,343
2.4 12 - 2,500 2,500 38,011 13,944,805 22 2,500 2,044 limit 12,970,687
2.5 9 131 5,200 2,245 limit 7,162,185 97 4,400 2,231 limit 7,122,703
2.5 10 86 4,000 2,140 limit 10,425,831 58 3,400 2,150 limit 13,198,840
2.5 11 61 3,100 1,922 limit 3,805,350 66 3,200 1,922 limit 3,860,573
2.5 12 47 3,000 2,030 limit 10,600,042 48 3,000 2,025 limit 9,488,796

Table A.6: Results on network net1 and 30 instances. A time limit of 11 h was imposed. The different strategies for solving the topology optimization problem
are presented in Chapter 5. Those instances not depicted are infeasible. The column “scale” shows the scaling of the nomination and the columns “loops” the
number of available loops. The column “cuts” shows the number of generated cuts.
Appendix A Tables
 no good cuts + restarts (strategy 5) with cuts (no restart) (strategy 3) with cuts (restart) (strategy 4)
scale loops gap primal dual time nodes gap primal dual time nodes cuts gap primal dual time nodes cuts
2.0 8 - 1,500 1,500 579 381,940 - 1,500 1,500 109 84,549 193 - 1,500 1,500 226 49,316 31
2.0 9 - 1,500 1,500 371 246,107 - 1,500 1,500 138 97,257 247 - 1,500 1,500 342 11,834 40
2.0 10 - 1,500 1,500 585 206,881 - 1,500 1,500 156 93,624 339 - 1,500 1,500 425 23,428 61
2.0 11 - 1,500 1,500 736 272,253 - 1,500 1,500 137 76,879 337 - 1,500 1,500 777 22,293 45
2.0 12 - 1,500 1,500 287 93,061 - 1,500 1,500 80 42,624 161 - 1,500 1,500 557 12,269 41
2.1 8 - 1,800 1,800 3,020 1,864,188 - 1,800 1,800 1,923 952,577 1,345 - 1,800 1,800 288 100,391 130
2.1 9 - 1,800 1,800 2,955 1,708,259 - 1,800 1,800 724 537,401 612 - 1,800 1,800 478 72,130 119
2.1 10 - 1,700 1,700 2,928 1,742,387 - 1,700 1,700 720 495,061 449 - 1,700 1,700 674 72,263 91
2.1 11 - 1,700 1,700 2,157 1,251,067 - 1,700 1,700 997 643,326 713 - 1,700 1,700 590 34,296 91
2.1 12 - 1,700 1,700 2,385 1,286,824 - 1,700 1,700 739 478,849 427 - 2,100 2,100 961 22,659 31
2.2 8 - 2,200 2,200 38.724 15,439,053 - 2,200 2,200 17,063 7,498,640 1,548 - 2,200 2,200 392 127,909 91
2.2 9 - 2,000 2,000 12,033 7,257,944 - 2,000 2,000 4,109 2,731,815 1,003 - 2,000 2,000 287 80,895 91
2.2 10 - 2,000 2,000 5,055 2,774,703 - 2,000 2,000 3,204 2,059,187 1,207 - 2,000 2,000 704 175,816 91
2.2 11 - 2,000 2,000 8,502 4,051,322 - 2,000 2,000 3,619 2,344,021 1,466 - 2,000 2,000 574 107,757 127
2.2 12 - 2,000 2,000 5,713 3,012,721 - 2,000 2,000 3,225 1,912,371 1,217 - 2,000 2,000 924 166,555 61
2.3 8 92 3,100 1,609 limit 11,695,102 70 3,100 1,815 limit 8,467,054 1,497 13 3,000 2,644 limit 2,622,733 5,224
2.3 9 40 2,500 1,775 limit 16,540,306 - 2,400 2,400 38,746 14,393,567 1,430 - 2,400 2,400 474 194,864 91
2.3 10 - 2,300 2,300 28,348 14,080,352 - 2,300 2,300 14,596 7,812,121 2,026 - 2,300 2,300 568 217,980 66
2.3 11 - 2,200 2,200 14,635 7,731,582 - 2,200 2,200 4,502 2,438,022 999 - 2,200 2,200 925 211,591 136
2.3 12 - 2,200 2,200 17,920 7,763,367 - 2,200 2,200 11,676 4,476,996 2,080 - 2,200 2,200 1,168 272,979 153
2.4 8 144 5,700 2,332 limit 14,554,022 120 5,800 2,628 limit 7,970,320 4,862 78 5,600 3,139 limit 9,595,085 1,032
2.4 9 73 3,200 1,843 limit 10,190,515 53 3,100 2,025 limit 11,958,426 1,891 - 3,100 3,100 620 281,801 204
2.4 10 55 2,800 1,806 limit 15,469,392 44 2,800 1,937 limit 11,329,023 2,802 - 2,700 2,700 1,054 421,319 166
2.4 11 40 2,600 1,854 limit 17,939,546 32 2,600 1,958 limit 16,327,371 1,790 - 2,600 2,600 1,122 457,510 136
2.4 12 35 2,500 1,845 limit 14,579,946 37 2,600 1,887 limit 11,569,901 2,340 - 2,500 2,500 1,697 439,561 204
2.5 9 126 4,600 2,035 limit 12,010,467 120 4,700 2,133 limit 7,206,449 1,317 39 4,500 3,221 limit 4,502,599 3,483
2.5 10 66 3,300 1,981 limit 14,218,724 61 3,400 2,107 limit 9,107,376 1,419 11 3,300 2,964 limit 3,978,200 3,753
2.5 11 68 3,100 1,842 limit 14,247,921 44 3,000 2,081 limit 14,940,544 1,729 - 3,000 3,000 2,524 1,017,901 354
2.5 12 57 3,000 1,900 limit 11,001,843 42 2,900 2,037 limit 12,776,311 1,900 - 2,900 2,900 1,967 921,701 136

Table A.7: Results (continued) on network net1 and 30 instances. A time limit of 11 h was imposed. The different strategies for solving the topology optimization
problem are presented in Chapter 5. Those instances not depicted are infeasible. The column “scale” shows the scaling of the nomination and the columns “loops”
the number of available loops. The column “cuts” shows the number of generated cuts.
213
 SCIP (strategy 1) cutoff (strategy 2)
214

scale loops gap primal dual time nodes gap primal dual time nodes
2.0 3 - 100 100 2 469 - 100 100 2 127
2.0 4 - 100 100 2 412 - 100 100 3 141
2.0 5 - 100 100 3 527 - 100 100 5 220
2.1 3 - 100 100 1 297 - 100 100 2 139
2.1 4 - 100 100 2 394 - 100 100 2 188
2.1 5 - 100 100 2 342 - 100 100 3 223
2.2 3 - 200 200 2 421 - 200 200 2 131
2.2 4 - 200 200 2 558 - 200 200 6 453
2.2 5 - 200 200 5 993 - 200 200 7 532
2.3 3 - 200 200 1 293 - 200 200 2 261
2.3 4 - 200 200 4 1,013 - 200 200 10 1,089
2.3 5 - 200 200 2 313 - 200 200 4 280
2.4 3 - 300 300 2 1,108 - 300 300 3 872
2.4 4 - 300 300 3 966 - 300 300 4 283
2.4 5 - 300 300 5 1,155 - 300 300 4 731
2.5 3 - 400 400 4 1,620 - 400 400 10 3,848
2.5 4 - 400 400 4 1,280 - 400 400 9 2,044
2.5 5 - 400 400 6 1,751 - 400 400 6 1,209
2.6 3 - 500 500 11 6,584 - 500 500 4 1,629
2.6 4 - 500 500 9 5,154 - 500 500 23 9,447
2.6 5 - 500 500 5 2,919 - 500 500 7 3,057
2.7 3 - 700 700 26 21,115 - 700 700 38 26,640
2.7 4 - 600 600 38 28,762 - 600 600 15 9,740
2.7 5 - 600 600 52 34,265 - 600 600 21 12,649
2.8 3 - 800 800 164 140,745 - 800 800 210 143,142
2.8 4 - 800 800 252 193,589 - 800 800 233 173,792
2.8 5 - 800 800 141 118,166 - 800 800 167 120,963
2.9 3 - 1,000 1,000 895 690,477 - 1,000 1,000 1,183 789,484
2.9 4 - 900 900 1,190 805,692 - 900 900 1,947 950,030
2.9 5 - 900 900 1,427 869,203 - 900 900 1,778 981,023
3.0 3 - 1,300 1,300 5,243 4,327,789 - 1,300 1,300 7,870 5,773,387
3.0 4 - 1,100 1,100 6,380 2,557,433 - 1,100 1,100 4,830 2,141,380
3.0 5 - 1,100 1,100 7,994 5,176,313 - 1,100 1,100 9,591 5,208,192
3.1 3 53 1,700 1,107 limit 24,381,440 82 2,000 1,098 limit 25,098,997
3.1 4 57 1,300 827 limit 7,358,275 58 1,300 819 limit 6,957,830
3.1 5 - 1,200 1,200 30,775 10,316,726 - 1,200 1,200 23,027 8,665,593
3.2 3 125 3,000 1,330 limit 40,470,382 126 3,000 1,323 limit 38,022,524
3.2 4 108 1,700 816 limit 6,878,911 120 1,800 816 limit 6,646,229
3.2 5 39 1,400 1,006 limit 24,473,376 63 1,500 917 limit 18,459,280
3.3 3 400 7,400 1,479 limit 24,672,027 396 7,400 1,490 limit 26,207,732
3.3 4 125 2,100 931 limit 9,094,229 107 1,900 915 limit 7,202,437
3.3 5 105 1,800 877 limit 8,842,062 99 1,700 851 limit 7,189,779
3.4 4 136 2,600 1,098 limit 13,464,692 145 2,700 1,100 limit 14,353,164
3.4 5 116 2,000 924 limit 7,219,501 146 2,300 933 limit 7,122,962
3.5 4 153 3,000 1,182 limit 14,580,836 165 3,100 1,166 limit 13,819,689
3.5 5 198 3,000 1,006 limit 11,273,438 115 2,200 1,020 limit 13,411,095
3.6 4 149 3,200 1,280 limit 17,477,233 190 3,800 1,306 limit 26,036,497
3.6 5 164 2,800 1,057 limit 15,527,932 162 2,800 1,066 limit 15,283,010
3.7 4 284 5,200 1,352 limit 13,760,208 284 5,300 1,379 limit 14,050,957
3.7 5 206 3,600 1,174 limit 17,115,145 158 3,100 1,197 limit 18,264,790
3.8 4 409 7,500 1,471 limit 10,677,370 417 7,500 1,449 limit 9,927,261
3.8 5 237 4,100 1,213 limit 14,844,993 219 3,900 1,222 limit 15,409,491
3.9 5 237 4,200 1,243 limit 12,183,208 330 5,400 1,254 limit 13,395,223

Table A.8: Results on network net2 and 60 instances. A time limit of 11 h was imposed. The different strategies for solving the topology optimization problem
are presented in Chapter 5. Those instances not depicted are infeasible. The column “scale” shows the scaling of the nomination and the columns “loops” the
number of available loops. The column “cuts” shows the number of generated cuts.
Appendix A Tables
 no good cuts + restarts (strategy 5) with cuts (no restart) (strategy 3) with cuts (restart) (strategy 4)
scale loops gap primal dual time nodes gap primal dual time nodes cuts gap primal dual time nodes cuts
2.0 3 - 100 100 6 183 - 100 100 1 16 2 - 100 100 1 2 1
2.0 4 - 100 100 3 29 - 100 100 2 41 12 - 100 100 1 11 2
2.0 5 - 100 100 1 10 - 100 100 2 51 15 - 100 100 3 40 4
2.1 3 - 100 100 1 1 - 100 100 2 101 17 - 100 100 1 2 1
2.1 4 - 100 100 10 252 - 100 100 1 62 5 - 100 100 2 50 3
2.1 5 - 100 100 19 404 - 100 100 3 158 11 - 100 100 13 222 14
2.2 3 - 200 200 5 238 - 200 200 2 94 20 - 200 200 7 284 23
2.2 4 - 200 200 24 1,013 - 200 200 5 240 37 - 200 200 32 893 12
2.2 5 - 200 200 24 701 - 200 200 8 445 57 - 200 200 37 1,021 16
2.3 3 - 200 200 18 1,039 - 200 200 9 469 91 - 200 200 5 96 19
2.3 4 - 200 200 32 1,181 - 200 200 5 578 24 - 200 200 38 1,242 22
2.3 5 - 200 200 55 2,013 - 200 200 9 610 47 - 200 200 72 1,821 16
2.4 3 - 300 300 19 1,186 - 300 300 3 199 13 - 300 300 24 1,138 16
2.4 4 - 300 300 45 2,347 - 300 300 9 927 62 - 300 300 64 2,723 19
2.4 5 - 300 300 68 2,191 - 300 300 13 1,079 84 - 300 300 91 2,699 30
2.5 3 - 400 400 17 1,591 - 400 400 17 3,518 131 - 400 400 3 152 3
2.5 4 - 400 400 44 3,605 - 400 400 8 1,299 32 - 400 400 56 2,470 22
2.5 5 - 400 400 63 3,441 - 400 400 9 1,381 40 - 400 400 71 3,546 7
2.6 3 - 500 500 20 2,639 - 500 500 8 1,729 60 - 500 500 6 726 32
2.6 4 - 500 500 47 4,432 - 500 500 8 2,201 38 - 500 500 58 2,844 41
2.6 5 - 500 500 82 6,839 - 500 500 9 2,291 26 - 500 500 83 4,490 18
2.7 3 - 700 700 101 47,908 - 700 700 80 25,554 572 - 700 700 36 2,842 67
2.7 4 - 600 600 58 16,759 - 600 600 19 6,465 101 - 600 600 47 3,288 41
2.7 5 - 600 600 88 21,317 - 600 600 21 9,188 79 - 600 600 77 8,535 41
2.8 3 - 800 800 321 259,584 - 800 800 444 103,865 3,893 - 800 800 101 20,504 306
2.8 4 - 800 800 641 402,363 - 800 800 224 106,306 1,230 - 800 800 86 8,690 91
2.8 5 - 800 800 511 340,700 - 800 800 145 94,302 490 - 800 800 82 11,344 28
2.9 3 - 1,000 1,000 2,229 1,310,819 - 1,000 1,000 958 344,129 5,182 - 1,000 1,000 149 28,936 417
2.9 4 - 900 900 8,465 3,901,049 - 900 900 1,514 638,805 5,600 - 900 900 129 38,188 204
2.9 5 - 900 900 3,271 2,361,298 - 900 900 1,381 858,811 3,446 - 900 900 118 24,799 91
3.0 3 - 1,300 1,300 33,215 15,303,215 - 1,300 1,300 7,993 4,268,476 19,253 - 1,300 1,300 394 95,624 1,002
3.0 4 - 1,100 1,100 11,235 5,408,097 - 1,100 1,100 4,127 1,819,166 9,411 - 1,100 1,100 205 83,185 204
3.0 5 - 1,100 1,100 14,619 9,351,222 - 1,100 1,100 2,872 1,957,876 5,347 - 1,100 1,100 519 139,101 566
3.1 3 125 1,900 840 limit 17,659,666 69 1,700 1,005 limit 16,055,176 124,498 16 1,700 1,460 limit 1,320,841 13,121
3.1 4 90 1,400 735 limit 15,299,955 - 1,300 1,300 35,008 7,668,150 23,436 - 1,300 1,300 1,353 222,636 1,548
3.1 5 46 1,200 818 limit 23,515,936 - 1,200 1,200 36,281 14,962,333 23,035 - 1,200 1,200 460 148,351 459
3.2 3 152 2,300 912 limit 25,980,239 147 3,000 1,211 limit 22,952,077 142,038 26 2,200 1,739 limit 1,537,141 11,754
3.2 4 100 1,600 798 limit 17,818,411 65 1,500 906 limit 11,585,021 22,924 - 1,500 1,500 2,464 487,931 1,730
3.2 5 96 1,500 762 limit 23,043,580 44 1,400 971 limit 18,874,203 13,320 - 1,400 1,400 1,792 520,520 1,032
3.3 3 225 4,000 1,229 limit 38,453,552 268 5,300 1,438 limit 30,121,605 67,404 47 3,300 2,235 limit 1,553,748 17,631
3.3 4 146 2,200 893 limit 16,903,163 101 1,900 944 limit 10,384,128 15,039 22 1,800 1,468 limit 1,396,263 11,754
3.3 5 99 1,700 852 limit 25,467,262 83 1,700 924 limit 21,068,388 11,460 - 1,600 1,600 3,962 748,767 2,033
3.4 4 168 2,600 969 limit 21,328,194 112 2,300 1,083 limit 21,156,982 40,985 43 2,400 1,676 limit 1,404,166 11,754
3.4 5 148 1,900 763 limit 25,743,706 99 1,900 952 limit 18,087,948 3,929 - 1,800 1,800 18,578 1,805,176 3,495
3.5 4 205 3,100 1,014 limit 24,477,965 159 3,000 1,155 limit 21,401,578 30,400 37 2,700 1,960 limit 1,414,853 11,754
3.5 5 152 2,400 950 limit 22,388,336 133 2,400 1,027 limit 17,266,232 8,457 23 2,100 1,701 limit 2,628,007 7,836
3.6 4 233 3,800 1,137 limit 19,267,344 194 3,700 1,256 limit 18,175,173 35,035 28 3,100 2,420 limit 1,709,756 11,754
3.6 5 219 3,100 969 limit 23,991,442 152 2,700 1,067 limit 15,352,407 7,940 36 2,600 1,906 limit 2,807,075 7,836
3.7 4 382 5,400 1,118 limit 17,811,354 272 4,900 1,313 limit 9,690,771 18,186 31 3,800 2,900 limit 1,782,097 11,754
3.7 5 193 3,200 1,088 limit 23,594,610 145 2,900 1,183 limit 16,222,287 8,113 37 2,900 2,105 limit 2,180,909 7,836
3.8 4 351 5,600 1,239 limit 22,488,013 345 6,400 1,437 limit 12,589,804 14,127 40 4,900 3,480 limit 1,696,707 11,754
3.8 5 245 4,100 1,187 limit 21,967,929 188 3,600 1,249 limit 16,366,510 8,054 50 3,600 2,392 limit 2,305,573 7,836
3.9 5 299 4,900 1,227 limit 25,521,263 199 3,900 1,301 limit 14,552,546 4,267 40 3,900 2,774 limit 2,554,368 7,787

Table A.9: Results (continued) on network net2 and 60 instances. A time limit of 11 h was imposed. The different strategies for solving the topology optimization
problem are presented in Chapter 5. Those instances not depicted are infeasible. The column “scale” shows the scaling of the nomination and the columns “loops”
the number of available loops. The column “cuts” shows the number of generated cuts.
215
216 Appendix A Tables

strategy 1 strategy 2 strategy 3

nom time nodes time nodes time nodes
1 7 131 2 112 2 112
2 3 39 1 26 1 26
3 1 44 1 34 1 34
4 4 199 1 30 1 30
5 3 162 2 35 3 35
6 327 7,022 3 235 3 235
7 1 26 10 205 93 205
8 2 72 1 50 1 50
9 1 1 1 8 1 8
10 2 18 1 17 1 17
11 2 17 1 20 1 20
12 2 55 1 28 1 28
13 3 236 3 53 3 53
14 12 788 1 26 1 26
15 1 44 1 71 1 71
16 limit 31,562 35 4,033 35 4,033
17 limit 53,277 56 8,012 55 8,012
18 16 575 5 327 4 327
19 1 19 1 12 1 12
20 2 27 1 19 1 19
21 1 39 1 20 1 20
22 251 22,587 127 3,042 36 798
23 2 35 7 129 57 129
24 2 46 4 41 7 41
25 1 1 0 1 1 1
26 1 14 1 54 1 54
27 2 41 1 33 1 33
28 4 304 2 24 2 24
29 limit 24,619 5 192 6 192
30 1 52 10 153 77 153
31 1 16 2 119 3 119
32 1 1 1 82 1 82
33 1 35 9 971 9 971
34 5 80 9 131 59 131
35 2 61 1 22 2 22
36 limit 28,498 1 40 1 40
37 1 25 2 232 2 232
38 3 226 13 2,651 12 2,651
39 3 132 6 161 5 161
40 2 134 9 140 28 140
41 58 1,183 1 8 1 8
42 1 15 2 121 2 121
43 2 22 3 202 3 202

Table A.10: Results on network net6 and 43 instances. A time limit of 1 h was imposed. The different
strategies for solving the topology optimization problem are presented in Chapter 6.
 SCIP (strat. 1) SCIP with priorities (strat. 2) domain relaxation heuristic (strat. 4) domain relaxation and check (strat. 3)
nom gap primal dual time nodes gap primal dual time nodes gap primal dual time nodes gap primal dual time nodes
1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
2 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
3 - 34 34 131 41,225 - 34 34 49 4,538 - 34 34 63 3,343 - 34 34 53 3,120
4 - 34 34 128 41,225 - 34 34 49 4,538 - 34 34 63 3,343 - 34 34 53 3,120
5 - 125 125 9,529 5,445,216 - 125 125 425 239,810 - 125 125 632 223,460 - 125 125 959 218,147
6 - 125 125 9,523 5,445,216 - 125 125 427 239,810 - 125 125 630 223,460 - 125 125 945 218,147
7 - 150 150 217 159,121 - 150 150 144 46,594 - 150 150 263 115,702 - 150 150 166 43,836
8 - 150 150 216 159,121 - 150 150 144 46,594 - 150 150 263 115,702 - 150 150 707 225,457
9 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
10 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
11 - 19 19 14 3,403 - 19 19 34 2,473 - 21 21 14 1,429 - 19 19 215 2,247
12 - 19 19 14 3,403 - 19 19 34 2,473 - 21 21 14 1,429 - 19 19 215 2,247
13 44 143 99 limit 9,636,729 - 130 130 657 338,707 - 130 130 2,441 465,459 - 130 130 1,269 352,263
14 44 143 99 limit 9,630,988 - 130 130 648 338,707 - 130 130 2,437 465,459 - 130 130 1,282 352,263
15 - 181 181 1,800 1,340,255 - 181 181 330 169,996 - 181 181 371 145,275 - 181 181 406 182,279
16 - 181 181 1,793 1,340,255 - 181 181 330 169,996 - 181 181 368 145,275 - 181 181 403 182,279
17 - 146 146 161 114,789 - 146 146 74 41,384 - 146 146 99 43,720 - 146 146 133 43,025
18 - 146 146 161 114,789 - 146 146 74 41,384 - 146 146 99 43,720 - 146 146 132 43,025
19 - 193 193 185 84,772 - 193 193 556 103,197 - 193 193 418 55,527 - 193 193 476 59,532
20 - 193 193 186 84,772 - 193 193 557 103,197 - 193 193 419 55,527 - 193 193 477 59,532
21 9 345 316 limit 12,424,524 - 345 345 6,009 1,911,912 - 345 345 8,490 2,333,807 - 345 345 6,739 2,189,269
22 9 345 316 limit 12,424,117 - 345 345 5,983 1,911,912 - 345 345 8,439 2,333,807 - 345 345 6,736 2,189,269
23 23 467 378 limit 14,846,550 20 467 388 limit 2,487,515 - 467 467 8,945 4,969,890 - 467 467 10,177 5,521,093
24 23 467 378 limit 14,882,249 20 467 388 limit 2,489,140 - 467 467 8,920 4,969,890 - 467 467 10,213 5,521,093

Table A.11: Results on network net3 and 24 instances. A time limit of 11 h was imposed. The different strategies for solving the topology optimization problem
are presented in Chapter 6.
217
 218

SCIP (strat. 1) SCIP with priorities (strat. 2) domain relaxation heuristic (strat. 4) domain relaxation and check (strat. 3)
nom gap primal dual time nodes gap primal dual time nodes gap primal dual time nodes gap primal dual time nodes
1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
2 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
3 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
4 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
5 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
6 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
7 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
8 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
9 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
10 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1 - 0 0 1 1
11 - 95 95 166 7,798 - 95 95 39 1,834 - 95 95 104 1,587 - 95 95 116 1,042
12 - 95 95 13 872 - 95 95 17 796 - 95 95 97 1,458 - 95 95 33 665
13 - 95 95 17 1,207 - 95 95 65 3,291 - 95 95 202 2,787 - 95 95 22 1,247
14 - 95 95 12 1,290 - 95 95 13 1,189 - 95 95 66 1,179 - 95 95 17 807
15 - 95 95 1 1 - 95 95 1 1 - 95 95 1 1 - 95 95 1 1
16 - 249 249 6,315 246,873 - 249 249 97 5,009 61 249 154 limit 298,007 - 249 249 173 3,318
17 12 320 285 limit 1,691,917 12 320 285 limit 1,354,643 - 320 320 26,805 200,944 - 320 320 330 3,185
18 1 475 466 limit 1,496,516 - 475 475 23,777 916,290 8 475 439 limit 508,811 - 475 475 986 3,120
19 10 525 475 limit 1,479,491 - 525 525 634 23,754 - 525 525 3,893 70,054 - 525 525 661 6,377
20 - 525 525 8 701 - 525 525 8 719 - 525 525 8 719 - 525 525 8 725
21 - 525 525 3 246 - 525 525 3 246 - 525 525 3 246 - 525 525 3 246
22 - 525 525 1 28 - 525 525 1 28 - 525 525 1 28 - 525 525 1 28
23 - 525 525 2 6 - 525 525 2 6 - 525 525 1 6 - 525 525 2 6
24 - 525 525 1 5 - 525 525 1 5 - 525 525 1 5 - 525 525 1 5
25 - 525 525 1 5 - 525 525 2 5 - 525 525 1 5 - 525 525 1 5
26 - 584 584 6 135 11 584 525 limit 1,928,200 11 584 525 limit 369,106 - 584 584 4 13
27 - 633 633 10,076 376,520 8 633 584 limit 1,472,909 - 633 633 896 5,907 - 633 633 69 536
28 - 633 633 9 185 - 633 633 9 185 - 633 633 27 185 - 633 633 8 21
29 78 938 525 limit 1,648,593 - 693 693 199 5,681 - 693 693 942 7,944 - 693 693 717 2,842
30 78 938 525 limit 1,549,956 24 788 633 limit 1,175,948 15 788 679 limit 218,167 - 788 788 1,710 9,182
31 89 997 525 limit 1,667,251 26 997 788 limit 659,384 57 997 633 limit 256,190 - 997 997 1,618 15,045
32 131 1,218 525 limit 4,325,792 38 1,092 788 limit 476,112 38 1,092 788 limit 195,827 - 1,092 1,092 2,018 7,051
33 133 1,227 525 limit 1,410,512 49 1,174 788 limit 278,890 27 1,171 918 limit 197,865 - 1,171 1,171 5,129 25,100
34 153 1,479 584 limit 1,389,987 25 1,174 938 limit 762,033 49 1,174 788 limit 187,185 - 1,174 1,174 5,133 26,092
35 86 1,182 633 limit 1,043,948 50 1,182 788 limit 340,425 50 1,182 788 limit 212,598 - 1,182 1,182 4,580 25,073
36 94 1,234 633 limit 1,175,254 26 1,237 978 limit 337,358 26 1,237 978 limit 320,287 - 1,234 1,234 4,463 30,296
37 103 1,286 633 limit 1,305,666 24 1,241 997 limit 513,348 22 1,241 1,015 limit 201,148 - 1,241 1,241 4,856 40,554
38 133 1,479 633 limit 908,286 11 1,318 1,182 limit 594,167 9 1,318 1,202 limit 252,838 - 1,318 1,318 5,013 50,056
39 92 1,336 693 limit 707,675 8 1,336 1,234 limit 407,863 8 1,336 1,230 limit 286,698 - 1,336 1,336 3,149 35,690
40 85 1,465 788 limit 537,315 17 1,459 1,245 limit 647,954 14 1,459 1,270 limit 294,554 5 1,459 1,381 limit 430,424
41 66 1,545 926 limit 1,326,633 20 1,545 1,286 limit 837,664 19 1,545 1,294 limit 310,438 4 1,545 1,481 limit 547,743

Table A.12: Results on network net4 and 41 nominations. A time limit of 11 h was imposed. The different strategies for solving the topology optimization problem
are presented in Chapter 6.
Appendix A Tables
 SCIP (strat. 1) SCIP with priorities (strat. 2) domain relaxation heuristic (strat. 4) domain relaxation and check (strat. 3)
nom gap primal dual time nodes gap primal dual time nodes gap primal dual time nodes gap primal dual time nodes
1 996 0 limit 332,054 1,999 0 limit 251,088 1,556 0 limit 142,814 1,239 0 limit 116,281
2 5,499 2,270 40 limit 739,421 40 limit 100,414 40 limit 102,524 14,755 6,024 40 limit 110,307
3 1,972 1,908 92 limit 1,918,145 1,181 1,180 92 limit 888,972 4,725 4,444 92 limit 1,568,040 92 limit 1,871,730
4 63 1,148 701 limit 342,856 63 1,148 701 limit 376,552 873 6,832 701 limit 145,366 228 2,306 701 limit 134,085
5 726 limit 2,646,703 728 limit 3,192,607 727 limit 2,843,775 726 limit 2,423,993
6 772 limit 2,955,237 759 limit 3,787,138 1,070 8,890 759 limit 2,158,999 833 7,071 757 limit 2,422,856
7 1,038 limit 3,004,645 987 limit 2,446,627 1,038 limit 3,059,269 970 limit 3,024,349
8 1,019 limit 3,700,692 994 limit 3,242,856 959 12,660* 995,717 994 limit 3,243,034
9 1,099 limit 3,520,812 1,099 limit 3,522,377 1,099 limit 3,519,282 1,099 limit 3,530,143
10 1,252 limit 2,552,061 1,250 limit 2,887,158 1,257 limit 2,598,481 1,266 limit 2,774,184
11 1,187 limit 3,153,578 1,187 limit 3,160,494 1,187 limit 3,142,580 1,187 limit 3,162,582

Table A.13: Results on network net5 and 11 nominations. A time limit of 11 h was imposed. The different strategies for solving the topology optimization problem
are presented in Chapter 6. The instance marked with * had numerical troubles due to a CPLEX LP error.
219
220 Appendix A Tables

strategy 1 strategy 2 strategy 3

no priorities with priorities dualval (basic) dualval (adapted)
nom time nodes time nodes time nodes calls time nodes calls
1 limit 241,297 limit 144,404 159 79 1 172 79 1
2 limit 109,869 limit 611,865 limit 744 118 134 308 1
3 limit 122,915 limit 112,174 limit 731 118 177 295 1
4 limit 138,123 limit 27,594 2,784 249 18 315 88 1
5 78 58 101 701 limit 195 50 207 64 1
6 limit 119,661 limit 592,352 272 64 1 254 64 1
7 limit 87,125 limit 162,034 193 296 1 114 296 1
8 limit 134,228 limit 503,061 90 134 1 84 134 1
9 limit 481,537 limit 128,105 1,036 328 10 1,758 328 10
10 limit 231,002 limit 335,555 limit 1,205 79 145 62 1
11 limit 113,277 limit 110,301 647 206 6 117 180 1
12 limit 140,155 limit 711,685 214 51 1 385 51 1
13 limit 114,814 limit 129,237 limit 714 108 5,759 356 45
14 limit 103,021 limit 388,074 limit 648 99 222 189 1
15 limit 141,208 limit 134,031 207 63 2 110 61 1
16 966 7,999 limit 87,825 84 268 0 87 268 0
17 109 183 limit 618,599 515 95 4 198 76 1
18 limit 105,106 limit 82,548 6,173 234 36 99 65 1
19 limit 134,483 limit 193,098 4,186 634 27 117 90 1
20 limit 133,759 limit 116,794 limit 704 141 2,036 88 6
21 limit 248,095 limit 399,486 198 177 2 120 172 1
22 limit 651,493 limit 771,975 603 313 4 269 85 1
23 limit 277,121 101 437 202 37 1 183 37 1
24 limit 637,225 limit 1,108,129 487 117 3 161 62 1
25 limit 133,181 limit 111,325 3,572 202 40 89 68 1
26 limit 105,742 limit 121,007 1,385 82 12 1,638 85 14
27 779 7,314 limit 130,146 222 72 2 259 72 2
28 limit 119,299 109 459 70 39 0 71 39 0
29 limit 110,801 limit 112,277 130 70 1 99 70 1
30 limit 112,109 limit 178,885 1,406 282 15 154 201 1

Table A.14: Results on network net7a and 30 nominations. A time limit of 4 h was imposed. The different
strategies for solving the topology optimization problem are presented in Chapter 7. Those instances with
a finite time limit were feasible. The column “calls” shows the number of calls of the primal heuristic
based on dual information.
 221

strategy 1 strategy 2 strategy 3

no priorities with priorities dualval (basic) dualval (adapted)
nom time nodes time nodes time nodes calls time nodes calls
1 limit 140,940 limit 68,362 limit 2,812 130 limit 2,665 102
2 limit 90,836 limit 381,513 143 261 1 142 261 1
3 103 442 limit 96,709 limit 36,705 118 limit 36,249 112
4 110 903 limit 59,288 limit 464 83 173 188 1
5 limit 140,474 86 274 482 69 5 151 54 1
6 limit 412,625 limit 639,087 limit 12,284 115 573 11,786 1
7 limit 117,310 limit 98,116 limit 3,451 159 limit 3,233 131
8 limit 312,566 limit 129,571 limit 932 157 limit 752 121
9 limit 152,721 limit 113,130 563 606 4 158 593 1
10 limit 105,796 limit 126,752 limit 63,450 80 7,546 66,510 21
11 limit 156,664 limit 68,237 218 121 1 545 273 3
12 limit 234,332 limit 450,738 limit 1,649 145 limit 1,179 94
13 limit 771,320 limit 1,220,105 2,964 112 12 limit 481 64
14 limit 491,487 limit 128,450 1,865 326 24 116 56 1
15 limit 144,534 limit 144,477 limit 1,093 156 8,260 639 61
16 limit 613,588 limit 116,138 limit 1,876 208 limit 1,265 124
17 limit 412,334 limit 392,347 5,019 559 34 811 262 5
18 limit 148,791 limit 94,420 limit 9,055 138 125 138 1
19 limit 337,272 limit 131,345 limit 1,304 137 limit 1,145 99
20 limit 490,877 limit 125,256 4,186 429 44 97 115 1
21 limit 249,526 limit 236,225 1,854 194 10 232 56 1
22 297 10,988 limit 158,298 limit 610 102 300 169 1
23 limit 103,000 limit 611,098 279 1,236 1 254 1,236 1
24 limit 137,949 limit 123,919 limit 1,381 139 limit 881 74
25 limit 106,484 limit 124,315 limit 2,572 108 limit 2,473 98
26 limit 127,864 limit 126,024 limit 1,784 152 limit 1,384 110
27 4,171 175,096 limit 613,955 limit 919 123 149 68 1
28 limit 147,134 limit 125,502 limit 1,561 157 limit 883 94
29 limit 118,015 limit 179,521 2,653 160 24 2,305 129 18
30 limit 210,212 limit 124,893 1,561 562 12 406 471 3

Table A.15: Results on network net7b and 30 nominations. A time limit of 4 h was imposed. The different
strategies for solving the topology optimization problem are presented in Chapter 7. Those instances with
a finite time limit were feasible. The column “calls” shows the number of calls of the primal heuristic
based on dual information.
222 Appendix A Tables

strategy 1 strategy 2 strategy 3

no priorities with priorities dualval (basic) dualval (adapted)
nom time nodes time nodes time nodes calls time nodes calls
1 limit 209,792 limit 175,127 5,081 1,554 28 7,574 1,558 30
2 limit 235,184 limit 293,344 544 1,233 3 202 1,225 1
3 limit 95,608 limit 194,961 limit 1,674 59 1,359 673 4
4 limit 725,232 limit 455,667 limit 4,942 86 limit 5,047 101
5 limit 453,246 limit 409,942 limit 741 69 7,398 661 45
6 limit 730,765 limit 681,565 1,725 7,254 9 442 5,201 1
7 limit 410,226 limit 544,259 269 699 1 317 699 1
8 limit 115,252 limit 592,581 limit 594 88 7,498 399 49
9 limit 116,342 14,088 104,447 limit 104,776 33 5,513 361 33
10 limit 225,347 limit 490,606 limit 1,060 90 limit 911 64
11 limit 96,301 limit 113,693 limit 2,254 86 190 174 1
12 limit 549,066 limit 789,363 419 384 2 228 375 1
13 limit 126,058 limit 119,756 limit 610 101 limit 548 85
14 limit 569,523 limit 375,609 limit 2,178 118 limit 2,190 121
15 limit 252,135 limit 237,202 limit 5,624 104 limit 5,116 78
16 limit 553,331 limit 119,828 3,864 349 33 2,806 204 20
17 limit 236,160 limit 145,606 limit 786 78 limit 753 68
18 limit 333,974 limit 329,769 limit 1,241 68 limit 1,214 58
19 limit 373,279 limit 213,921 limit 773 69 5,136 299 17
20 limit 165,352 limit 146,906 limit 4,872 77 limit 2,972 64
21 limit 133,832 limit 174,777 6,991 6,962 47 142 263 1
22 161 1,193 limit 672,644 1,328 217 10 645 190 4
23 limit 479,245 limit 367,011 limit 2,031 65 761 512 3
24 limit 700,383 limit 535,530 limit 13,133 30 2,234 5,525 4
25 limit 137,073 limit 116,393 limit 3,925 73 limit 339 47
26 limit 425,587 limit 92,517 limit 870 90 200 58 1
27 limit 104,986 limit 315,932 limit 764 95 limit 697 77
28 limit 128,744 limit 909,649 321 4,709 1 337 4,709 1
29 limit 101,648 limit 110,769 limit 663 59 limit 680 64
30 limit 256,820 limit 83,425 limit 979 82 limit 933 70

Table A.16: Results on network net7c and 30 nominations. A time limit of 4 h was imposed. The different
strategies for solving the topology optimization problem are presented in Chapter 7. Those instances with
a finite time limit were feasible. The column “calls” shows the number of calls of the primal heuristic
based on dual information.
 223

strategy 1 strategy 2 strategy 3

no priorities with priorities dualval (basic) dualval (adapted)
nom time nodes time nodes time nodes calls time nodes calls
1 limit 92,986 limit 53,134 limit 1,911 75 limit 1,832 64
2 limit 107,559 limit 221,183 limit 1,525 81 limit 1,462 61
3 limit 294,798 limit 213,185 limit 16,201 97 2,305 10,172 3
4 limit 264,519 limit 407,828 limit 575 87 error 5
5 limit 481,910 limit 119,520 limit 6,422 95 1,716 3,110 10
6 limit 686,619 limit 695,384 limit 533 59 1,291 303 5
7 limit 495,048 limit 471,546 1,727 1,011 8 8,721 1,246 43
8 limit 125,606 limit 327,295 limit 92,402 0 limit 94,489 0
9 limit 52,482 limit 452,027 limit 1,789 99 limit 1,454 63
10 limit 194,612 limit 296,502 limit 681 91 2,981 436 19
11 limit 117,377 limit 98,475 limit 133,894 10 limit 18,324 52
12 limit 482,468 limit 465,670 6,544 975 25 1,470 852 6
13 limit 114,820 limit 140,953 limit 593 67 limit 439 37
14 limit 560,563 limit 378,138 limit 1,350 60 7,847 1,247 29
15 limit 258,260 limit 268,361 limit 358 89 237 81 1
16 limit 536,908 170 1,511 7,616 524 61 5,044 276 36
17 limit 341,725 limit 210,353 limit 1,502 117 397 372 1
18 limit 175,509 limit 102,087 limit 359,541 14 limit 374,860 14
19 limit 197,670 limit 288,764 9,091 1,114 41 11,581 1,112 43
20 limit 136,959 limit 129,193 limit 226 36 limit 669 60
21 limit 114,578 limit 466,233 8,580 12,477 56 3,476 11,595 16
22 limit 392,281 limit 402,646 limit 388 94 361 57 2
23 limit 545,123 limit 480,047 limit 7,061 50 limit 7,765 75
24 limit 764,936 limit 464,159 limit 577 78 2,949 350 20
25 limit 161,312 limit 629,419 limit 1,721 100 7,751 318 40
26 limit 119,817 limit 166,047 limit 5,484 76 limit 3,542 36
27 limit 91,989 limit 242,744 limit 2,525 91 limit 2,500 71
28 limit 423,135 limit 682,647 limit 5,590 81 785 98 4
29 limit 131,923 limit 230,005 limit 12,891 85 642 239 4
30 limit 257,926 limit 162,310 limit 3,119 54 limit 2,972 42

Table A.17: Results on network net7d and 30 nominations. A time limit of 4 h was imposed. The different
strategies for solving the topology optimization problem are presented in Chapter 7. The error originates
from numerical troubles in the LP solver. Those instances with a finite time limit were feasible. The
column “calls” shows the number of calls of the primal heuristic based on dual information.
224 Appendix A Tables

strategy 1 strategy 2 strategy 3

no priorities with priorities dualval (basic) dualval (adapted)
nom time nodes time nodes time nodes calls time nodes calls
1 limit 47,244 limit 195,673 limit 1,623 49 limit 1,640 53
2 limit 265,266 limit 167,224 limit 27,063 46 1,869 17,775 3
3 limit 193,970 limit 154,650 limit 1,422 57 limit 1,513 77
4 limit 425,537 limit 388,587 6,013 691 38 2,438 296 7
5 limit 154,950 limit 362,629 limit 616 53 259 381 1
6 limit 311,869 limit 148,075 limit 48,493 62 7,196 48,133 28
7 limit 306,996 limit 335,788 limit 2,139 55 limit 2,160 59
8 limit 311,992 limit 162,684 limit 459 76 114 275 1
9 limit 55,740 limit 33,938 limit 824 67 limit 828 65
10 limit 397,544 limit 305,236 limit 776 49 limit 849 70
11 132 466 limit 291,708 limit 489 60 1,292 168 6
12 limit 52,270 limit 526,302 558 2,133 2 779 2,188 4
13 limit 59,516 limit 244,825 limit 513 73 10,266 449 53
14 limit 295,171 limit 197,110 limit 22,572 77 limit 4,895 64
15 limit 356,782 limit 254,441 954 258 6 220 239 1
16 limit 85,951 limit 119,517 895 428 6 1,161 428 6
17 limit 362,794 limit 382,453 limit 48,787 67 limit 45,209 31
18 limit 269,555 limit 79,645 limit 3,752 54 452 3,228 1
19 limit 335,511 limit 263,183 limit 3,981 47 2,905 3,625 11
20 limit 138,307 limit 46,787 limit 929 86 129 133 1
21 limit 100,438 limit 59,936 1,697 2,462 10 522 1,449 3
22 limit 256,607 limit 344,673 8,336 1,193 41 limit 1,386 59
23 limit 415,409 limit 512,692 limit 879 64 limit 760 47
24 limit 363,684 limit 536,135 451 6,806 1 1,698 10,269 6
25 limit 77,653 limit 127,088 limit 254 67 limit 169 42
26 limit 58,964 limit 26,786 limit 915 58 limit 998 82
27 limit 58,944 limit 235,868 limit 18,841 81 limit 17,303 63
28 limit 422,310 limit 462,954 998 19,059 1 1,138 19,081 2
29 limit 117,490 limit 51,026 limit 554 93 98 43 1
30 limit 62,883 limit 309,384 limit 18,383 56 598 469 3

Table A.18: Results on network net7e and 30 nominations. A time limit of 4 h was imposed. The different
strategies for solving the topology optimization problem are presented in Chapter 7. Those instances with
a finite time limit were feasible. The column “calls” shows the number of calls of the primal heuristic
based on dual information.
 225

heuristic dualval (adapted)

& domain relaxation and check
nom gap primal dual time nodes
1 1,007 0 limit 84,984
2 3,759 1,565 40 limit 90,292
3 2,954 2,813 92 limit 155,623
4 1,867 13,801 701 limit 164,731
5 174 1,997 728 limit 248,268
6 886 7,513 761 limit 109,671
7 630 6,673 913 limit 125,628
8 1,089 9,915 833 limit 469,076
9 883 10,994 1,117 limit 158,182
10 619 12,450 1,731 limit 190,015
11 674 13,199 1,703 limit 137,384

Table A.19: Results on network net5 and 11 nominations. A time limit of 11 h was imposed. The strategy
for solving the topology optimization problem is presented in Chapter 7.
Bibliography
K. Abhishek, S. Leyffer, and J. T. Linderoth. FilMINT: An Outer-Approximation-
Based Solver for Nonlinear Mixed Integer Programs. In: INFORMS Journal On
Computing 22, No. 4, 2010, pp. 555–567 (cit. on p. 17).

T. Achterberg. SCIP - a framework to integrate Constraint and Mixed Integer

Programming. Tech. rep. ZR-04-19. Zuse Institute Berlin, 2004 (cit. on p. 14).

T. Achterberg. SCIP: Solving Constraint Integer Programs. In: Mathematical Pro-

gramming Computation 1, No. 1, 2009, pp. 1–41 (cit. on pp. 10, 13, 17).

T. Ahadi-Oskui, S. Vigerske, I. Nowak, and G. Tsatsaronis. Optimizing the design of

complex energy conversion systems by Branch and Cut. In: Computers & Chemical
Engineering 34, No. 8, 2010, pp. 1226–1236 (cit. on p. 14).

J. André, J. F. Bonnans, and L Cornibert. “Planning reinforcement of gas trans-

portation networks with optimization methods.” In: Proceedings 19th Mini EURO
Conference on the Energy Sector. 2006 (cit. on p. 7).

F. Babonneau, Y. Nesterov, and J.-P. Vial. Design and Operations of Gas Trans-
mission Networks. In: Operations Research 60, No. 1, 2012, pp. 34–47 (cit. on
p. 8).

P. Belotti, C. Kirches, S. Leyffer, J. Linderoth, J. Luedtke, and A. Mahajan. Mixed-

Integer Nonlinear Optimization. Preprint ANL/MCS-P3060-1121. Argonne Na-
tional Laboratory, 2012 (cit. on p. 13).

T. Berthold. “Heuristic algorithms in global MINLP solvers.” PhD thesis. Technische

Universität Berlin, 2014 (cit. on p. 17).

P. Bonami and J. Gonçalves. Heuristics for convex mixed integer nonlinear programs.
In: Computational Optimization and Applications 51 (2), 2012, pp. 729–747 (cit. on
p. 17).

227
228 Bibliography

P. Bonami, L. T. Biegler, A. R. Conn, G. Cornuéjols, I. E. Grossmann, C. D. Laird,

J. Lee, A. Lodi, F. Margot, N. Sawaya, and A. Wächter. An algorithmic framework
for convex mixed integer nonlinear programs. In: Discrete Optimization 5, No. 2,
2008, pp. 186–204 (cit. on p. 17).

J. Bonnans, G. Spiers, and J.-L. Vie. Global optimization of pipe networks by the
interval analysis approach: the Belgium network case. Rapport de Recherche RR
7796. INRIA, 2011 (cit. on pp. 7 sq.).

C. Borraz-Sánchez and R. Z. Ríos-Mercado. “A Tabu Search Approach for Minimizing

Fuel Consumption on Cyclic Natural Gas Pipeline Systems.” In: Proceedings of
12th CLAIO Conference. 2004 (cit. on p. 7).

C. Borráz-Sánchez and R. Ríos-Mercado. A Non-Sequential Dynamic Programming

Approach for Natural Gas Network Optimization. In: WSEAS Transactions on
Systems 3, No. 4, 2004, pp. 1384–1389 (cit. on p. 7).

I. D. Boyd, P. D. Surry, and N. J. Radcliffe. Constrained Gas Network Pipe Sizing with
Genetic Algorithms. Tech. rep. EPCC-TR94-11. Edinburgh Parallel Computing
Centre, 1994 (cit. on p. 7).

S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press,

2004 (cit. on p. 17).

C. Bragalli, C. D’Ambrosio, J. Lee, A. Lodi, and P. Toth. On the optimal design of

water distribution networks: a practical MINLP approach. In: Optimization and
Engineering 13 (2), 2012, pp. 219–246 (cit. on p. 35).

BGR 2013: Bundesanstalt für Geowissenschaften und Rohstoffe. Energiestudie 2013 -

Reserven, Ressourcen und Verfügbarkeit von Energierohstoffen. 2013. url: http:
/ / www . bgr . bund . de / DE / Themen / Energie / Produkte / energiestudie2013 _
Zusammenfassung.html. Downloaded on August 26, 2014 (cit. on p. 1).

GasNZV 2005: Bundesministerium der Justiz. Verordnung über den Zugang von
Gasversorgungnetzen (Gasnetzzugangsverordnung – GasNZV). 2005. url: http:
//www.bmwi.de/BMWi/Redaktion/PDF/Gesetz/energiewirtschaftsgesetz-
verordnung-gasnetz-gasnzv-april-2005,property=pdf,bereich=bmwi2012,
sprache=de,rwb=true.pdf. Version released Apr. 14, 2005 (cit. on p. 1).
Bibliography 229

J. Burgschweiger, B. Gnädig, and M. C. Steinbach. Optimization Models for Operative

Planning in Drinking Water Networks. In: Optimization and Engineering 10, No.
1, 2009, pp. 43–73 (cit. on p. 35).

M. R. Bussieck and S. Vigerske. MINLP Solver Software. In: Wiley Encyclopedia of

Operations Research and Management Science. Ed. by J. J. C. et.al. Wiley & Sons,
Inc., 2010 (cit. on p. 13).

R. G. Carter. Compressor Station Optimization: Computational Accuracy and Speed.

Tech. rep. PSIG 9605. Pipeline Simulation Interest Group, 1996 (cit. on p. 28).

R. G. Carter. “Pipeline Optimization: Dynamic Programming after 30 Years.” In:

Proceedings of the 30th PSIG Annual Meeting. 1998 (cit. on p. 7).

R. G. Carter, D. Schroeder, and T. Harbick. Some Causes and Effects of Discontinu-

ities in Modeling and Optimizing Gas Transmission Networks. Tech. rep. PSIG
9308. Pipeline Simulation Interest Group, 1993 (cit. on p. 28).

L. Castillo and A. Gonzáleza. Distribution network optimization: Finding the most

economic solution by using genetic algorithms. In: European Journal of Operational
Research 108, No. 3, 1998, pp. 527–537 (cit. on p. 7).

G. Cerbe. Grundlagen der Gastechnik: Gasbeschaffung – Gasverteilung – Gasverwen-

dung. Hanser Verlag, Leipzig, Germany, 2008 (cit. on pp. 1, 30).

A. Chebouba, F. Yalaoui, A. Smati, L. Amodeo, K. Younsi, and A. Tairi. Optimiza-

tion of Natural Gas Pipeline Transportation Using Ant Colony Optimization. In:
Computers & Operations Research 36, No. 6, 2009, pp. 1916–1923 (cit. on p. 7).

M. Collins, L. Cooper, R. Helgason, J. Kennington, and L. LeBlanc. Solving the pipe

network analysis problem using optimization techniques. In: Management Science
24, No. 7, 1978, pp. 747–760 (cit. on pp. 54 sq., 73).

A. Conn, N. Gould, and P. Toint. Trust-Region Methods. MPS-SIAM Series on

Optimization. Society for Industrial and Applied Mathematics, 2000 (cit. on p. 19).

C. D’Ambrosio and A. Lodi. Mixed integer nonlinear programming tools: a practical

overview. In: 4OR 9, No. 4, 2011, pp. 329–349 (cit. on p. 13).
230 Bibliography

C. D’Ambrosio and A. Lodi. Mixed integer nonlinear programming tools: an updated

practical overview. In: Annals of Operations Research 204, No. 1, 2013, pp. 301–320
(cit. on p. 13).

E. Danna, E. Rothberg, and C. L. Pape. Exploring relaxation induced neighborhoods

to improve MIP solutions. In: Mathematical Programming 102, No. 1, 2004, pp. 71–
90 (cit. on p. 17).

G. B. Dantzig. “Maximization of a linear function of variables subject to linear inequal-

ities.” In: Activity Analysis of Production and Allocation. Ed. by T. Koopmanns.
John Wiley & Sons, New York, 1951, pp. 339–347 (cit. on p. 15).

D. De Wolf and B. Bakhouya. Optimal dimensioning of pipe networks: the new

situation when the distribution and the transportation functions are disconnected.
Tech. rep. 07/02. Ieseg, Université catholique de Lille, HEC Ecole de Gestion de
l’ULG, 2008 (cit. on p. 45).

D. De Wolf and Y. Smeers. Optimal dimensioning of pipe networks with application

to gas transmission networks. In: Operations Research 44, No. 4, 1996, pp. 596–608
(cit. on p. 8).

D. De Wolf and Y. Smeers. The Gas Transmission Problem Solved by an Extension of

the Simplex Algorithm. In: Management Science 46, No. 11, 2000, pp. 1454–1465
(cit. on pp. 7, 28, 45).

K. Ehrhardt and M. C. Steinbach. “Nonlinear Optimization in Gas Networks.” In:

Modeling, Simulation and Optimization of Complex Processes. Ed. by H. G. Bock,
E. Kostina, H. X. Phu, and R. Ranacher. Springer Berlin Heidelberg, 2005, pp. 139–
148 (cit. on p. 7).

K. Ehrhardt and M. C. Steinbach. KKT Systems in Operative Planning for Gas

Distribution Networks. In: Proceedings in Applied Mathematics and Mechanics 4,
No. 1, 2004, pp. 606–607 (cit. on p. 7).

M. Feistauer. Mathematical Methods in Fluid Dynamics. Vol. 67. Pitman Monographs

and Surveys in Pure and Applied Mathematics Series. Longman Scientific &
Technical, Harlow, 1993 (cit. on p. 24).
Bibliography 231

A. V. Fiacco and Y. Ishizuka. Sensitivity and Stability Analysis for Nonlinear

Programming. In: Annals of Operations Research 27, No. 1, 1990, pp. 215–236
(cit. on pp. 184 sqq.).

FICO. Xpress-Optimizer Reference manual. 2009. url: http://www.fico.com/

xpress (cit. on p. 17).

C. A. Floudas. Nonlinear and Mixed Integer Optimization: Fundamentals and Appli-

cations. Oxford University Press, New York, 1995 (cit. on p. 14).

A. Fügenschuh, B. Hiller, J. Humpola, T. Koch, T. Lehman, R. Schwarz, J. Schweiger,

and J. Szabó. Gas Network Topology Optimization for Upcoming Market Require-
ments. In: IEEE Proceedings of the 8th International Conference on the European
Energy Market (EEM), 2011, pp. 346–351 (cit. on pp. iv, viii).

A. Fügenschuh, H. Homfeld, H. Schülldorf, and S. Vigerske. “Mixed-Integer Nonlinear

Problems in Transportation Applications.” In: Proceedings of the 2nd International
Conference on Engineering Optimization (CD-ROM). Ed. by H. Rodrigues et al.
2010 (cit. on p. 15).

A. Fügenschuh, B. Geißler, R. Gollmer, C. Hayn, R. Henrion, B. Hiller, J. Humpola,

T. Koch, T. Lehmann, A. Martin, R. Mirkov, A. Morsi, J. Rövekamp, L. Schewe,
M. Schmidt, R. Schultz, R. Schwarz, J. Schweiger, C. Stangl, M. C. Steinbach,
and B. M. Willert. Mathematical Optimization for Challenging Network Planning
Problems in Unbundled Liberalized Gas Markets. In: Energy Systems, 2013 (cit. on
pp. iv, viii).

A. Fügenschuh, B. Geißler, R. Gollmer, A. Morsi, M. E. Pfetsch, J. Rövekamp, M.

Schmidt, K. Spreckelsen, and M. C. Steinbach. Physical and technical fundamentals
of gas networks. In: Evaluating Gas Network Capacities. Ed. by T. Koch, B. Hiller,
M. E. Pfetsch, and L. Schewe. SIAM-MOS series on Optimization. SIAM, 2015.
Chap. 2, pp. 17–44 (cit. on p. 24).

B. Furey. A Sequential Quadratic Programming-Based Algorithm for Optimization

of Gas Networks. In: Automatica 29, No. 6, 1993, pp. 1439–1450 (cit. on p. 7).

GAMS Model Library. General Algebraic Modeling System (GAMS) Model Library.
url: http://www.gams.com/modlib/modlib.htm (cit. on p. 45).
232 Bibliography

B. Geißler. “Towards Globally Optimal Solutions for MINLPs by Discretization

Techniques with Applications in Gas Network Optimization.” PhD thesis. Friedrich-
Alexander-Universität Erlangen-Nürnberg, 2011 (cit. on pp. 9, 24).

I. Gentilini, F. Margot, and K. Shimada. The Traveling Salesman Problem with

Neighborhoods: MINLP Solution. In: Optimization Methods and Software 28, 2013,
pp. 364–378 (cit. on p. 51).

A. M. Geoffrion. Generalized Benders Decomposition. In: Journal of Optimization

Theory and Applications 10, No. 4, 1972, pp. 237–260 (cit. on pp. 12, 103).

B. Gilmour, C. Luongo, and D. Schroeder. Optimization in Natural Gas Transmission

Networks: A Tool to Improve Operational Efficiency. Tech. rep. Stoner Associates
Inc., 1989 (cit. on p. 7).

A. M. Gleixner, H. Held, W. Huang, and S. Vigerske. Towards globally optimal oper-

ation of water supply networks. In: Numerical Algebra, Control and Optimization
2, No. 4, 2012, pp. 695–711 (cit. on p. 35).

I. E. Grossmann and Z. Kravanja. Mixed-Integer nonlinear programming: A survey of

algorithms and applications. In: Large-Scale Optimization with Applications, Part
II: Optimal Design and Control. Ed. by A. R. Conn, L. T. Biegler, T. F. Coleman,
and F. N. Santosa. Vol. 93. The IMA Volumes in Mathematics and its Applications.
Springer, New York, 1997, pp. 73–100 (cit. on p. 14).

Y. Hamam and A. Brameller. Hybrid Method for the Solution of Piping Networks. In:
Proceedings of the Institution of Electrical Engineers 118, No. 11, 1971, pp. 1607–
1612 (cit. on p. 7).

C. T. Hansen, K. Madsen, and H. B. Nielsen. Optimization of pipe networks. In:

Mathematical Programming 52, No. 1-3, 1991, pp. 45–58 (cit. on p. 8).

M. Hübner. “Druckebenenübergreifende Grundsatzplanung von Gasverteilungsnetzen.”

PhD thesis. RWTH Aachen, 2009 (cit. on p. 8).

J. Humpola and A. Fügenschuh. Convex reformulations for solving a nonlinear

network design problem. In: Computational Optimization and Applications 62, No.
3, 2015, pp. 717–759 (cit. on pp. iv, viii).
Bibliography 233

J. Humpola and F. Serrano. Sufficient pruning conditions for MINLP in gas network
design. In: EURO Journal on Computational Optimization 5, No. 1, 2017, pp. 239–
261 (cit. on pp. iv, viii).

J. Humpola, A. Fügenschuh, and T. Lehmann. A primal heuristic for optimizing

the topology of gas networks based on dual information. In: EURO Journal on
Computational Optimization 3, No. 1, 2015, pp. 53–78 (cit. on pp. iv, viii).

J. Humpola, A. Fügenschuh, B. Hiller, T. Koch, T. Lehmann, R. Lenz, R. Schwarz,

and J. Schweiger. The specialized MINLP approach. In: Evaluating Gas Network
Capacities. Ed. by T. Koch, B. Hiller, M. E. Pfetsch, and L. Schewe. SIAM-MOS
series on Optimization. SIAM, 2015. Chap. 7, pp. 123–144 (cit. on pp. iv, viii, 101).

J. Humpola, A. Fügenschuh, and T. Koch. Valid inequalities for the topology opti-
mization problem in gas network design. In: OR Spectrum 38, No. 3, 2016, pp. 597–
631 (cit. on pp. iv, viii).

IBM. CPLEX. url: http://www.cplex.com (cit. on pp. 10, 17, 44).

T. Jeníček. “Steady-State Optimization of Gas Transport.” In: Proceedings of 2nd

International Workshop SIMONE on Innovative Approaches to Modeling and
Optimal Control of Large Scale Pipeline Networks. Prague, 1993, pp. 26–38 (cit. on
p. 7).

W. Karush. Minima of Functions of Several Variables with Inequalities as Side

Conditions. In: Traces and Emergence of Nonlinear Programming. Ed. by G. Giorgi
and T. H. Kjeldsen. Springer Basel, 2014, pp. 217–245 (cit. on p. 20).

S. Kim, R. Z. Ríos-Mercado, and E. A. Boyd. Heuristics for Minimum Cost Steady-

State Gas Transmission Networks. In: Computing Tools for Modeling, Optimization
and Simulation. Ed. by M. Laguna and J. L. González-Velarde. Vol. 12. Interfaces in
Computer Science and Operations Research. Kluwer, Boston, USA, 2000. Chap. 11,
pp. 203–213 (cit. on p. 7).

T. Koch, B. Hiller, M. E. Pfetsch, and L. Schewe, eds. Evaluating Gas Network

Capacities. SIAM-MOS series on Optimization. SIAM, 2015 (cit. on pp. 9, 24).

B. Korte and J. Vygen. Combinatorial Optimization: Theory and Algorithms. Springer

Verlag, Berlin, 2007 (cit. on pp. 2, 23, 35, 58).
234 Bibliography

H. W. Kuhn and A. W. Tucker. Nonlinear programming. In: Proceedings of the 2nd

Berkley Symposium on Mathematical Statistics and Probability. Ed. by J. Neyman.
Uinversity Press, Berkley, California, 1951, pp. 481–493 (cit. on p. 20).

H. Lall and P. Percell. A dynamic programming based Gas Pipeline Optimizer. In:
Analysis and Optimization of Systems. Ed. by A. Bensoussan and J. Lions. Vol. 144.
Lecture Notes in Control and Information Sciences. Springer Berlin Heidelberg,
1990, pp. 123–132 (cit. on p. 7).

Lamatto++. A Framework for Modeling and Solving Mixed-Integer Nonlinear Pro-

gramming Problems on Networks. url: http://www.mso.math.fau.de/edom/
projects/lamatto.html (cit. on p. 44).

C. Li, W. Jia, Y. Yang, and X. Wu. Adaptive Genetic Algorithm for Steady-State
Operation Optimization in Natural Gas Networks. In: Journal of Software 6, No.
3, 2011, pp. 452–459 (cit. on p. 7).

LIWACOM. Benutzerhandbuch. Version 5.5. LIWACOM Informations GmbH and

SIMONE Research Group s.r.o. 2005 (cit. on p. 8).

M. V. Lurie. Modeling of Oil Product and Gas Pipeline Transportation. Wiley-VCH,

Weinheim, 2008 (cit. on pp. 24, 27).

D. Mahlke, A. Martin, and S. Moritz. A Simulated Annealing Algorithm for Transient

Optimization in Gas Networks. In: Mathematical Methods of Operations Research
66, No. 1, 2007, pp. 99–116 (cit. on p. 7).

J. Mallinson, A. Fincham, S. Bull, J. Rollet, and M. Wong. Methods for optimizing

gas transmission networks. In: Annals of Operations Research 43, 1993, pp. 443–454
(cit. on p. 7).

O. Mariani, F. Ancillai, and E. Donati. Design of a Gas Pipeline: Optimal Configu-

ration. Tech. rep. PSIG 9706. Pipeline Simulation Interest Group, 1997 (cit. on
p. 8).

A. Martin, M. Möller, and S. Moritz. Mixed Integer Models for the Stationary Case
of Gas Network Optimization. In: Mathematical Programming 105, No. 2, 2006,
pp. 563–582 (cit. on p. 7).
Bibliography 235

A. Martin, B. Geißler, C. Hayn, J. Humpola, T. Koch, T. Lehman, A. Morsi,

M. Pfetsch, L. Schewe, M. Schmidt, R. Schultz, R. Schwarz, J. Schweiger, M.
Steinbach, and B. Willert. Optimierung Technischer Kapazitäten in Gasnetzen.
In: VDI-Berichte: Optimierung in der Energiewirtschaft 2157, 2011, pp. 105–115
(cit. on pp. iv, viii).

J. J. Maugis. Etude de réseaux de transport et de distribution de fluide. In: RAIRO

Operations Research 11, No. 2, 1977, pp. 243–248 (cit. on pp. 54 sq., 58, 73).

J. Mischner. Notices about hydraulic calculations of gas pipelines. In: GWF–

Gas/Erdgas 4, 2012, pp. 158–273 (cit. on p. 25).

R. Misener and C. Floudas. A Framework for Globally Optimizing Mixed-Integer

Signomial Programs. In: Journal of Optimization Theory and Applications 161, No.
3, 2014, pp. 905–932 (cit. on p. 10).

M. Möller. “Mixed Integer Models for the Optimisation of Gas Networks in the
Stationary Case.” PhD thesis. Technische Universität Darmstadt, Fachbereich
Mathematik, 2004 (cit. on p. 7).

A. Morsi. “Solving MINLPs on Loosely-Coupled Networks with Applications in Water

and Gas Network Optimization.” PhD thesis. Friedrich-Alexander-Universität
Erlangen-Nürnberg, 2013 (cit. on p. 9).

G. L. Nemhauser and L. A. Wolsey. Integer Programming. In: Optimization. Ed. by

G. L. Nemhauser, A. H. G. Rinnooy Kan, and M. J. Todd. Elsevier, 1989. Chap. 6,
pp. 447–527 (cit. on pp. 14 sqq.).

J. Nikuradse. Strömungsgesetze in rauhen Rohren. Forschungsheft auf dem Gebiete

des Ingenieurwesens. VDI-Verlag, Düsseldorf, 1933 (cit. on p. 25).

J. Nikuradse. Laws of Flow in Rough Pipes. Vol. Technical Memorandum 1292.

National Advisory Committee for Aeronautics Washington, 1950 (cit. on p. 25).

V. S. Nørstebø, F. Rømo, and L. Hellemo. Using operations research to optimise

operation of the Norwegian natural gas system. In: Journal of Natural Gas Science
and Engineering 2, 2010, pp. 153–162 (cit. on p. 7).
236 Bibliography

J. Oldham. “Combinatorial approximation algorithms for generalized flow problems.”

In: Proceedings of the tenth annual ACM-SIAM symposium on Discrete algorithms
SODA’99. 1999, pp. 704–714 (cit. on p. 66).

A. Osiadacz and M. Górecki. Optimization of Pipe Sizes for Distribution Gas Network
Design. Tech. rep. PSIG 9511. Pipeline Simulation Interest Group, 1995 (cit. on
p. 8).

A. J. Osiadacz. Simulation and analysis of gas networks. E. & F.N. Spon Ltd, London,
1987 (cit. on p. 4).

P. B. Percell and M. J. Ryan. Steady-State Optimization of Gas Pipeline Network

Operation. Tech. rep. PSIG 8703. Pipeline Simulation Interest Group, 1987 (cit. on
p. 7).

M. E. Pfetsch, A. Fügenschuh, B. Geißler, N. Geißler, R. Gollmer, B. Hiller, J.

Humpola, T. Koch, T. Lehmann, A. Martin, A. Morsi, J. Rövekamp, L. Schewe,
M. Schmidt, R. Schultz, R. Schwarz, J. Schweiger, C. Stangl, M. C. Steinbach,
S. Vigerske, and B. M. Willert. Validation of Nominations in Gas Network Opti-
mization: Models, Methods, and Solutions. In: Optimization Methods and Software,
2014 (cit. on pp. iv, viii, 24).

J. D. Pintér, ed. Global Optimization: Scientific and Engineering Case Studies. Vol. 85.
Nonconvex Optimization and Its Applications. Springer, 2006 (cit. on p. 14).

A. U. Raghunathan. Global Optimization of Nonlinear Network Design. In: SIAM

Journal on Optimization 23, No. 1, 2013, pp. 268–295 (cit. on pp. 51, 55).

R. Z. Ríos-Mercado and C. Borraz-Sánchez. Optimization Problems in Natural Gas

Transmission Systems: A State-of-the-Art Survey. Submitted. 2012 (cit. on p. 4).

R. Z. Ríos-Mercado, S. Wu, L. R. Scott, and E. A. Boyd. A Reduction Technique

for Natural Gas Transmission Network Optimization Problems. In: Annals of
Operations Research 117, No. 1, 2002, pp. 217–234 (cit. on p. 7).

R. Z. Ríos-Mercado, S. Kim, and E. A. Boyd. Efficient Operation of Natural Gas

Transmission Systems: A Network-Based Heuristic for Cyclic Structures. In: Com-
puters & Operations Research 33, No. 8, 2006, pp. 2323–2351 (cit. on p. 7).
Bibliography 237

D. Scheibe and A. Weimann. Dynamische Gasnetzsimulation mit GANESI. In: GWF–

Gas/Erdgas 9, 1999, pp. 610–616 (cit. on p. 8).

M. Schmidt, M. C. Steinbach, and B. M. Willert. High detail stationary optimization

models for gas networks. In: Optimization and Engineering, 2014, pp. 1–34 (cit. on
p. 9).

D. Shaw. Pipeline System Optimization: A Tutorial. Tech. rep. PSIG 9405. Pipeline
Simulation Interest Group, 1994 (cit. on p. 4).

C. Stangl. “Modelle, Strukturen und Algorithmen für stationäre Flüsse in Gasnetzen.”

PhD thesis. Universität Duisburg-Essen, 2014 (cit. on p. 9).

M. C. Steinbach. On PDE Solution in Transient Optimization of Gas Networks. In:

Journal of Computational and Applied Mathematics 203, No. 2, 2007, pp. 345–361
(cit. on p. 7).

C. K. Sun, V. Uraikul, C. W. Chan, and P. Tontiwachwuthikul. An integrated expert

system/operations research approach for the optimization of natural gas pipeline
operations. In: Engineering Applications of Artificial Intelligence 13, No. 4, 2000,
pp. 465–475 (cit. on p. 7).

M. Tawarmalani and N. Sahinidis. A polyhedral branch-and-cut approach to global

optimization. In: Mathematical Programming 103, No. 2, 2005, pp. 225–249 (cit. on
p. 10).

M. Tawarmalani and N. V. Sahinidis. Convexification and Global Optimization

in Continuous and Mixed-Integer Nonlinear Programming: Theory, Algorithms,
Software, and Applications. Vol. 65. Nonconvex Optimization and Its Applications.
Kluwer Academic Publishers, 2002 (cit. on p. 14).

A. Tomasgard, F. Rømo, M. Fodstad, and K. Midthun. Optimization Models for

the Natural Gas Value Chain. In: Geometric Modelling, Numerical Simulation,
and Optimization. Ed. by G. Hasle, K.-A. Lie, and E. Quak. Springer Verlag, New
York, 2007, pp. 521–558 (cit. on p. 7).

FNBGas 2013: Vereinigung der Fernleitungsnetzbetreiber Gas e. V. Netzentwicklungs-

plan Gas 2013. Konsultationsdokument der deutschen Fernleitungsnetzbetreiber.
238 Bibliography

url: http://www.netzentwicklungsplan-gas.de/. Downloaded on August 26,

2014 (cit. on p. 1).

S. Vigerske. “Decomposition in Multistage Stochastic Programming and a Constraint

Integer Programming Approach to Mixed-Integer Nonlinear Programming.” PhD
thesis. Humboldt-Universität zu Berlin, 2012 (cit. on pp. 10, 13 sq., 17 sqq.).

Z. Vostrý. “Transient Optimization of Gas Transport and Distribution.” In: Pro-

ceedings of 2nd International Workshop SIMONE on Innovative Approaches to
Modeling and Optimal Control of Large Scale Pipeline Networks. Prague, 1993,
pp. 53–62 (cit. on p. 7).

A. Wächter and L. T. Biegler. On the Implementation of a Primal-Dual Interior

Point Filter Line Search Algorithm for Large-Scale Nonlinear Programming. In:
Mathematical Programming 106, No. 1, 2006, pp. 25–57 (cit. on pp. 10, 20, 44).

J. F. Wilkinson, D. V. Holliday, E. H. Batey, and K. W. Hannah. Transient Flow in

Natural Gas Transmission Systems. American Gas Association, New York, 1964
(cit. on p. 25).

P. J. Wong and R. E. Larson. Optimization of Natural-Gas Pipeline Systems via

Dynamic Programming. In: IEEE Transactions Automatic Control 15, No. 5, 1968,
pp. 475–481 (cit. on p. 7).

P. J. Wong and R. E. Larson. Optimization of Tree-Structured Natural-Gas Trans-

mission Networks. In: Journal of Mathematical Analysis and Applications 24, 1968,
pp. 613–626 (cit. on p. 7).

S. Wright, M. Somani, and C. Ditzel. Compressor Station Optimization. Tech. rep.

PSIG 9805. Pipeline Simulation Interest Group, 1998 (cit. on p. 7).

S. Wu, R. Z. Ríos-Mercado, E. A. Boyd, and L. R. Scott. Model Relaxations for the

Fuel Cost Minimization of Steady-State Gas Pipeline Networks. In: Mathematical
and Computer Modelling 31, No. 2, 2000, pp. 197–220 (cit. on pp. 7, 28).

J. Zhang and D. Zhu. A Bilevel Programming Method for Pipe Network Optimization.
In: SIAM Journal on Optimization 6, No. 3, 1996, pp. 838–857 (cit. on p. 8).
Bibliography 239

Q. Zheng, S. Rebennack, N. Iliadis, and P. Pardalos. Optimization Models in the

Natural Gas Industry. In: Handbook of Power Systems I. Springer, 2010, pp. 121–
148 (cit. on p. 4).

H. Zimmer. Calculating Optimum Pipeline Operations. Tech. rep. SAND2009-5066C.

El Paso Natural Gas Company, 1975 (cit. on p. 7).
This thesis is about mathematical optimization for an efficient operation of gas
transmission networks. The challenging question is how to expand and operate
the network in order to facilitate the transportation of specified gas quantities at
minimum cost. This problem is a major challenge for gas network operators.
It is extremely hard to solve due to the combinatorial complexity of the active
network elements such as compressors, the nonlinear physical characteristic of
pipelines, and the immense sizes of the problem instances. Mathematical models
and optimization techniques can result in huge gains for the network operators in
terms of cost reductions and automated computations. We tackle this challenge
by developing novel mathematical theory and associated innovative optimization
algorithms for large scale instances. This allows us to produce solutions for a
real-world instance, i.e., the largest gas network in Germany.

Jesco Humpola received his diploma degree

in mathematics from the University of Bonn
in 2009 and in 2014 his Ph.D. in applied
mathematics from the Technical University
Berlin. He has participated in the development
of optimization software for gas transmission
networks, for airline crew scheduling, and for
VLSI chip design.

Logos Verlag Berlin

ISBN 978-3-8325-4505-5