ECONOMIC DISPATCH USING HYBRID

APPROACHES

By

Engr. Tahir Nadeem Malik

2K1-UET/PhD-EE-05

Supervisor
Prof. Dr. Azzam ul Asar

Department of Electrical Engineering

University of Engineering and Technology
Taxila (PAKISTAN)
March 2009
 ECONOMIC DISPATCH USING HYBRID APPROACHES

By
Engr. Tahir Nadeem Malik
2K1-UET/PhD-EE-05
A thesis submitted in partial fulfillment of the requirement for the degree of Doctor of

Philosophy.

Approved by
External Examiners

______________________________ ______________________________
Prof. Dr. Iftikhar Ahmed Khan Dr. Muhammad Ejaz Hassan
Foreign faculty Professor Department of Electrical Engineering
Department of Electrical Engineering NUST College of E&ME,
NWFP University of Engineering & Peshawar Road,
Technology, Peshawar. Rawalpindi.

Internal Examiner/ Thesis Supervisor

___________________________________
Prof. Dr. Azzam ul Asar
Dean Engineering, NWFP, UET, Peshawar.

Department of Electrical Engineering

University of Engineering and Technology
Taxila (PAKISTAN)
 DECLARATION
The substance of this thesis is original work of the author and due reference and
acknowledgement has been made, where necessary, to the work of others. No part of the
thesis has already accepted for any degree, and it is not being currently submitted in
candidature of any degree.

Engr. Tahir Nadeem Malik

(2K1-UET/Ph D-EE-05)
Thesis Scholar

.
DEDICATED TO

My Parents
Family
and
Dr. Aftab Ahmad
 v

ABSTRACT

Power Economic Dispatch (ED) is necessary and vital step in power system
operational planning. It is nonconvex constrained optimization problem defined as the
process of calculating the generation of the generating units for the minimum total
production cost in such a way that both equality and inequality constraints are satisfied. In
system operation studies generators are represented by input-output curves. These
characteristics curves are inherently nonlinear and non-smooth due to valve point effect,
multiple fuels and operational constraints such as prohibited operating zones. The
accurate economic dispatch depends mainly upon the accurate representation of these
curves and their handling in the optimization process.

Generally, economic dispatch is formulated as convex problem and has been

solved using mathematical programming techniques by approximating generator
input/output characteristic curves of monotonically increasing nature thus resulting in an
inaccurate dispatch. However, the nonconvex ED problem cannot be handled effectively
by such approaches. The Genetic algorithm is the potential solution methodology due to
its inherent ability to address the convex and nonconvex problems equally. This
dissertation presents the application of genetic algorithm (GA) for the solution of
economic dispatch problem independently as well as in hybrid form in conjunction with
the other techniques.

The problem is addressed first by developing a extensible and flexible

computational framework called “PED_Frame” as common environment which becomes
a platform for the computer implementation of different algorithms under consideration.
This framework has been used for implementation of economic dispatch algorithms for (i)
GA based models, (ii) Hybrid models.

Economic dispatch problem has been formulated in binary coded genetic

algorithm environment based on real power search and λ search methodologies. Two
biological mechanism “inversion” and “deletion-regeneration” has also been mapped as
an operator with crossover probability. Various GA based evolution models have been
constructed by adopting different initial population generation schemes, selection
 vi

methods, and crossover operators. Convex ED studies have been conducted using
standard test systems and results have been compared with λ iteration approach.

GA based hybrid approach for convex ED dispatch is proposed. This approach

initially run GA based ED with λ-search and passes the control to conventional λ iteration
technique. This approach gives another systematic method for selection of initial value of
λ. The results of the proposed approach on standard test system show that costs of
generation by this approach is almost the same as the λ iteration alone, however, it takes
less number of iterations.

The performance of GA based economic dispatch problem has been evaluated

with reference to different evolution models on the basis of empirical data available by
actually running the program for the nonconvex ED due to valve point effect.

National utility system has been reviewed with reference to its operation
problems. Four test systems close to original network have been developed and tested by
load flow analysis using Newton’s Raphson algorithm. Finally 12-Machine 32 bus test
circuit, 15, 25 and 34 Machines systems for economic dispatch studies have been
developed. ED studies have been conducted using test circuit

The Genetic algorithm has the inherent ability to bring the solution to the global
minimum region of search space in a short time and then takes longer time to converge to
the solution. This research work proposed hybrid approaches to fine tune the near optimal
results produced by GA. In this context, three hybrid approaches have been used for the
solution of nonconvex economic dispatch problem with valve point effect. These include
(i) A Synergy of GA and ED using Newton’s Second Order Approach, (ii) Neuro-Genetic
Hybrid Approach, and (iii) Hybrid of GA and Sequential Quadratic Programming. These
models have been tested on standard test systems and the results obtained from all the
three hybrid approaches offer significant improvement in the generation cost showing the
promise of the proposed approaches.
 vii

ACKNOWLEDGEMENTS
To begin with the name of ALMIGHTY ALLAH, who gave me the strength and

spirit to fulfill the mandatory requirements of this dissertation. All the respect to the

HOLY PROPHET (May Peace Be Upon Him), who enabled me to recognize our creator

and direct the human being to the right path. I started my research work with Dr. Aftab

Ahmad as my supervisor. But this journey was broken due to his sad and sudden demise.

Fate has cut his feathers and shortens his journey otherwise he would have flown to the

heights others have not yet imagined in his field. I wish to place on record my deep sense

of admiration to Dr. Aftab Ahmad for his encouragement, advice and help in my work.

May Almighty Allah rest the departed soul in eternal peace!

The journey of Ph.D research will remain unforgettable in my life. The sad demise
of my father, mother and father-in-law within the span of eight months brought the grief
and sorrow for me and my family. I express my sincere respect and honour to them as
they were major candles in my life. May Almighty Allah rest the departed souls in eternal
peace!

The work was restarted under the supervision of Dr. Azam ul Asar. His valuable
knowledge and vast experience of the subject removed the difficulties at all the critical
junctures. I particularly acknowledge the positive influence of many simulating
discussion I had had with him. Without his generous encouragement and patient guidance
the completion of this task was not possible. I will always remain grateful for his
technical and moral support.

I would like to thank the authorities of University of Engineering & Technology,

Taxila for providing the financial means for conducting this research. Special thanks to
the members of my Ph.D research monitoring committee for guidance & suggestions
regarding the research work.
 viii

Thanks are also due to many friends and colleagues Prof. Aftab Ahmad, Engr.
Methab Ahmad, Abdul Qudus Abbasi, Engr. Shahid Shafiq, Abrar Hashmi, Engr.
Zakaullah Sheikh….., for their helpful discussions during the period of research work.
Although only one name appears on the spine of this thesis, many others have contributed
in providing the social and moral support in completion of research.

Finally, I would like to give my special gratitude to my wife and children for their
patience and support.
 ix

CONTENTS
ABSTRACT
ACKNOWLEDGEMENT
LIST OF FIGURES
LIST OF TABLES

CHAPTER 1
Introduction
1.1 General 1
1.2 Problem Statement 1
1.3 Objectives 3
1.4 Scope of the Work 4
1.5 Thesis Outline 5

CHAPTER 2
Power Economic Dispatch --- A Brief Survey
2.1 Introduction 7
2.2 Power System Operational Planning 7
2.3 Economic dispatch --- Literature Survey 8
2.3.1 Thermal Machine’s Cost Curve in Economic Dispatch 10
2.3.2 Nonconvex Cost Curve 10
2.3.3 Economic Dispatch Based on Traditional Optimization--- Bottlenecks 12
2.3.4 Genetic Algorithm Based Economic Dispatch 12

CHAPTER 3
Optimization Methods
3.1 Introduction 22
3.2 Evolutionary Computation 22
3.3 Genetic Algorithm 23
3.3.1 Encoding & Decoding 24
3.3.2 Chromosome Length 25
3.3.3 Fitness Function 25
3.3.4 Genetic Algorithm Operators 26
 x

3.3.5 GA Parameters 28
3.3.6 Termination Criteria 29
3.4 Particle Swarm Optimization (PSO) 29
3.4.1 Guaranteed Convergence PSO (GCPSO) Approach 31
3.4.2 Parameters Settings in PSO Model 32
3.5 Artificial Neural Networks 32
3.5.1 Feedforward Neural Networks 33
3.6 Optimization Using Newton's Approach 33
3.6.1 Minimizing the Lagrangian Function 34
3.6.2 General Non-Linear Optimization Problem 35
3.7 Sequential Quadratic Programming (SQP) 38
3.7.1 MATLAB Sequential Quadratic Programming 40
3.8 Hybrid GA 41

CHAPTER 4
Computational Framework for Economic Dispatch Problem
4.1 Introduction 43
4.2 Framework 43
4.3 Computational Framework “PED_Frame” --- Need & Scope 44
4.4 Features and Characteristics of Proposed Framework 44
4.5 PED_Frame Design 45
4.5.1 MFC Document -View Architecture 45
4.5.2 Input/Output Using Grid 47
4.5.3 Analysis Status Visualization 47
4.5.4 Vector-Matrix Library 47
4.5.5 File Read Writes 47
4.6 Framework Implementation 48
4.7 Steps for Implementation of Customized Solutions Using PED_Frame 48
4.7.1 Economic Dispatch Using Mathematical Programming based Approach 48
4.7.2 Economic Dispatch Using GA 49
4.7.3 Economic Dispatch Using Hybrid Approach 49
4.8 DOS Mode Operation 49
4.9 PED_Frame --- Execution Mode 50
 xi

CHAPTER 5
Economic Dispatch Solution using Genetic Algorithm
5.1 Introduction 56
5.2 Economic Dispatch Problem --- Mathematical Formulation 56
5.2.1 Convex Economic Dispatch 56
5.2.2 Nonconvex Economic Dispatch 57
5.2.3 Transmission Loss in Economic Dispatch 57
5.2.4 Economic Dispatch Using GA--- Search Methodology 58
5.2.5 Objective Functions 59
5.3 Economic Dispatch Solution --- GA Based Modeling 59
5.3.1 Initial Population Generation 59
5.3.2 Reproduction Operators 60
5.3.3 Crossover Operators 60
5.3.4 Inversion (INV) 61
5.3.5 Deletion -Regeneration (DEL-REG) 61
5.3.6 Chromosome Length 62
5.3.7 Decoding 62
5.3.8 Fitness Function 62
5.3.9 Termination Criteria 62
5.4 Economic Dispatch Using GA --- (λ –Search) 63
5.5 Economic Dispatch Using GA --- Real Power –Search (RP-Search) 64
5.6 GA Based Hybrid Approach for Convex Economic Dispatch 64
5.7 Computer Implementation 65
5.8 Case Studies --- Convex Economic Dispatch 65
5.8.1 3-Machines Test Systems 65
5.8.2 IEEE Test Systems 66
5.8.3 GA λ-Search Metrology with Mismatch Tolerance 0.0001 69
5.8.4 Summary and Conclusion for the Case Studies 69
5.9 Case Studies --- Convex Economic Dispatch Using Hybrid Approach 69
5.10 Nonconvex Economic Dispatch 70
5.10.1 Intra Comparison --- Performance Indicators Variations (PIV) 73
5.10.2 Inter Comparison --- Performance Indicators Variations (PIV) 74
5.10.3 Convergence Profiles 83
5.10.4 Discussion 84
 xii

CHAPTER 6
Economic Dispatch Studies on NTDC System
6.1 Introduction 99
6.2 WAPDA System ---- Brief Overview 99
6.3 National Transmission & Distribution Company 99
6.4 Operational Constraints in NTDC System 100
6.4.1 Hydro-Electric Generation Constraints 100
6.4.2 Thermal Generation Constraints 102
6.4.3 Effects of Transmission Constraints 102
6.4.4 Seasonal Variations in Power Demand 102
6.5 Test Circuit for NTDC System 103
6.6 Case Studies 108

CHAPTER 7
Proposed Hybrid Models for Nonconvex Economic Dispatch
Problem
7.1 Introduction 110
7.2 GA Evolution Model for Hybrid Approaches 110
7.3 Hybrid Model-I: A Synergy of GA and Newton’s Second Order Approach
111
7.3.1 Economic Dispatch Using NSOA ---- Problem formulation 111
7.3.2 A Synergy of GA-NSOA --- Methodology 113
7.3.3 Test Results 115
7.4 Hybrid Model –II: Neuro-Genetic Hybrid Approach 124
7.4.1 Artificial Neural Network using SI Learning Rule 125
7.4.2 Test Results 126
7.5 Hybrid Model–III: A Hybrid of GA and Sequential Quadratic Programming
131
7.5.1 Sequential Quadratic Programming (SQP) 131
7.5.2 Genetic Algorithm – Sequential Quadratic Programming (GA-SQP) Hybrid
Methodology 132
7.5.3 Case Studies 132
 xiii

CHAPTER 8
Conclusions and Suggestions
References
APPENDIX A
Test Systems --- Description & Data
A.1 Three Machines Test Systems 154
A.1.1 Test System I --- 3-Machine 6-Bus System with PD = 210 MW 154
A.1.2 Test System II --- 3-Machine 6-Bus System with PD = 850 MW 155
A.1.3 Test System II --- 3-Machine 5-Bus System with PD = 150 MW 156
A.2 IEEE Test Systems 156
A.3 American Power Company (APC) Test System 158
A.4 20- Machines System 159
A.5 13-Machines System with Nonconvex Cost Curve 160
A.6 NTDC System 160
APPENDIX B
Quadratic Optimization
B.1 Minimizing Quadratic Functions 162
B.1.1 Minimizing Multivariable Functions 163
APPENDIX C
Output Results --- Chapter 5 & 6 165

APPENDIX D
Derived Publications 177
 xiv

LIST OF FIGURES
Figure 2-1: Power System Control Activities ..................................................................... 8
Figure 2-2: Value Point Effect .......................................................................................... 11
Figure 2-3: Hybrid Cost & Incremental Cost ................................................................... 11
Figure 3-1: Simple Genetic Algorithm Structure ............................................................. 24
Figure 4-1: Class diagram of PED_Frame ........................................................................ 45
Figure 4-2: PED_Frame in visual C++ Environment ....................................................... 50
Figure 4-3: File Pull Down Menu in PED_Frame ............................................................ 50
Figure 4-4: Cost curves of machines read from data file in PED_Frame Environment ... 51
Figure 4-5: B-coefficient read from data file in PED_Frame Environment ..................... 51
Figure 4-6: λ Iteration method in execution mode in PED_Frame ................................... 52
Figure 4-7: Parameter settings for λ Iteration method in PED_Frame ............................. 52
Figure 4-8: λ Iteration: Output results in PED_Frame ..................................................... 53
Figure 4-9: λ Iteration using GA in PED_Frame .............................................................. 53
Figure 4-10: GA parameter settings in PED_Frame ......................................................... 54
Figure 4-11: Progress indicator in PED_Frame ................................................................ 54
Figure 4-12: λ Iteration using GA: Output results in PED_Frame ................................... 55
Figure 4-13: Economic Dispatch using GA based on Real Power Search: Output results
in PED_Frame ............................................................................................ 55
Figure 5-1: Genetic Algorithm Working Steps................................................................. 58
Figure 5-2: Pseudo Code for λ –Search Methodology ..................................................... 64
Figure 5-3: Minimum Generation Cost and Minimum Execution Time against Crossover
Operators for Roulette Wheel Selection ........................................................ 76
Figure 5-4: Minimum Generation Cost and Minimum Execution Time against Crossover
Operators for Tournament Selection .............................................................. 77
Figure 5-5: Minimum Generation Cost and Minimum Execution Time against Crossover
Operators for Rank Selection ......................................................................... 78
Figure 5-6: Minimum Generation Cost and Minimum Execution Time against Crossover
Operators for Stochastic Remainder Selection .............................................. 79
Figure 5-7: Minimum Generation Cost and Minimum Execution Time against Crossover
Operators for Random Encoding Scheme...................................................... 80
Figure 5-8: Minimum Generation Cost and Minimum Execution Time against Crossover
Operators for Gray Encoding Scheme ........................................................... 81
 xv

Figure 5-9: Minimum Generation Cost and Minimum Execution Time against Crossover
Operators for Complementary Encoding Scheme ....................................... 82
Figure 5-10: Power Mismatch against No. of Generations with Roulette Wheel Selection,
Three Encoding Schemes and One Point Crossover.................................... 86
Figure 5-11: Power Mismatch against No. of Generations with Roulette Wheel Selection,
Three Encoding Schemes and Uniform Crossover ...................................... 87
Figure 5-12: Power Mismatch against No. of Generations with Roulette Wheel Selection,
Three Encoding Schemes and Inversion ...................................................... 88
Figure 5-13: Power Mismatch against No. of Generations with Roulette Wheel Selection,
Three Encoding Schemes and Deletion & Regeneration ............................. 89
Figure 5-14: Power Mismatch against No. of Generations with Tournament Selection,
Three Encoding Schemes, One Point & Uniform Crossover .................... 90
Figure 5-15: Power Mismatch against No. of Generations with Tournament Selection,
Three Encoding Schemes and Inversion .................................................... 91
Figure 5-16: Power Mismatch against No. of Generations with Tournament Selection,
Three Encoding Schemes and Deletion & Regeneration ........................... 92
Figure 5-17: Power Mismatch against No. of Generations with Rank Selection, Three
Encoding Schemes, One Point & Uniform Crossover ............................... 93
Figure 5-18: Power Mismatch against No. of Generations with Tournament Selection,
Three Encoding Schemes and Inversion .................................................... 94
Figure 5-19: Power Mismatch against No. of Generations with Tournament Selection,
Three Encoding Schemes and Deletion & Regeneration ........................... 95
Figure 5-20: Power Mismatch against No. of Generations with Stochastic Remainder
Selection, Three Encoding Schemes, One Point & Uniform Crossover.... 96
Figure 5-21: Power Mismatch against No. of Generations with Stochastic Remainder
Selection, Three Encoding Schemes and Inversion ................................... 97
Figure 5-22: Power Mismatch against No. of Generations with Stochastic Remainder
Selection, Three Encoding Schemes and Deletion & Regeneration .......... 98
Figure 6-1: WAPDA Hydroelectric Generation System ................................................ 103
Figure 6-2: 77-Bus Test System ..................................................................................... 104
Figure 6-3: 55-Bus Test System ..................................................................................... 104
Figure 6-4: 44-Bus Test System ..................................................................................... 105
Figure 6-5: 32-Bus Test System ..................................................................................... 105
Figure 6-6: Voltage Profile NTDC 32-Bus Test Circuit................................................. 106
 xvi

Figure 6-7: Voltage Profile NTDC 44-Bus Test Circuit................................................. 106

Figure 6-8: Voltage Profile NTDC 55-Bus Test Circuit................................................. 106
Figure 6-9: Voltage Profile NTDC Test Circuit ............................................................. 107
Figure 7-1: Flow Chart for Economic Dispatch Using NSOA ....................................... 114
Figure 7-2: Flow Chart for GA-NSOA Hybrid Approach .............................................. 115
Figure 7-3: Total Generation Cost Profiles: GA & GA-NSOA Approaches for 3–
Machines System ..................................................................................... 119
Figure 7-4: Total Generation Cost Profiles: GA & GA-NSOA Approaches for Case-I
IEEE 5-Machines 14 Bus System .............................................................. 120
Figure 7-5: Total Generation Cost Profiles: GA & GA-NSOA Approaches for for Case-II
IEEE 5-Machines 14 Bus System .............................................................. 121
Figure 7-6: Total Generation Cost Profiles: GA & GA-NSOA Approaches for for Case-I
IEEE 6-Machines 30 Bus System .............................................................. 122
Figure 7-7: Total Generation Cost Profiles: GA & GA-NSOA Approaches for for Case-II
IEEE 6-Machines 30 Bus System .............................................................. 123
Figure 7-8: Flowchart for Neuro-Genetic Approach ...................................................... 125
Figure 7-9: Steps for SI Learning ANN.......................................................................... 126
Figure 7-10: Test Case 1: Comparison of Percentage Error of the Generation Cost...... 130
Figure 7-11: Test Case 2: Comparison of Percentage Error of the Generation Cost...... 130
Figure 7-12: Test Case 3: Comparison of Percentage Error of the Generation Cost...... 131
Figure 7-13: Total Generation Cost Profiles: GA & GA-SQP Approaches for 3-Machines
Test System ................................................................................................ 134
Figure 7-14: Total Generation Cost Profiles: GA & GA-SQP Approaches for5-Machines
Test System ................................................................................................ 135
Figure 7-15: Total Generation Cost Profiles: GA and GA-SQP Approaches for 6-
Machines Test System ............................................................................. 136
Figure A-1: Generators Cost Curve with Value-Point Effect ......................................... 154
 xvii

LIST OF TABLES
Table 4-1: Classes in PED_Frame .................................................................................... 46
Table 5-1: Comparison of the Results for 3-Machines Test Systems I, II, and III ........... 66
Table 5-2: Comparison of the Results for IEEE 5-Machines Test System....................... 67
Table 5-3: Comparison of the Results for IEEE 6-Machines Test System....................... 67
Table 5-4: Comparison of the Results for 6-Machines 26 Bus System ............................ 68
Table 5-5: Comparison of the Results: Convex ED using Hybrid Approach ................... 70
Table 5-6: Generation Cost, Minimum Execution Time & No. of Convergences for
Roulette Wheel Selection and Tournament Selection against Crossover
Operators with Three Encoding Schemes ...................................................... 71
Table 5-7: Generation Cost, Minimum Execution Time & No. of Convergences for Rank
Selection and Stochastic Remainder Selection against Crossover Operators
with Three Encoding Schemes....................................................................... 72
Table 5-8: Minimum No. of Generations for Convergence After 20 Trials ..................... 72
Table 6-1: Summary of Load Flow Results .................................................................... 105
Table 6-2: Summary of Load Flow Result: NTDC 32-Bus Test Circuit ........................ 107
Table 6-3: Comparison of the Results for NTDC System .............................................. 109
Table 7-1: GA – Models & GA Parameters 3 – Machines System & IEEE Test Systems
Case - I ......................................................................................................... 116
Table 7-2: GA – Models & GA Parameters IEEE Test Systems – Case - II .................. 117
Table 7-3: Output Results GA-NSOA: 3-Machines 6-Bus Test System After 50 Runs 119
Table 7-4: Output Results GA-NSOA Case-I: IEEE 5-Machines 14-Bus System After 50
Runs with PD=259 MW ............................................................................... 120
Table 7-5: Output Results GA-NSOA Case-II: IEEE 5-Machines 14-Bus System After
50 Runs with PD=259 MW .......................................................................... 121
Table 7-6: Output Results GA-NSOA Case-I: IEEE 6-Machines 30-Bus System After 50
Runs with PD=283.4 MW ............................................................................ 122
Table 7-7: Output Results GA-NSOA Case-II: IEEE 6-Machines 30-Bus System After
50 Runs with PD=283.4 MW ....................................................................... 123
Table 7-8: Percentage Reduction in Total Generation Cost in GA-NSOA Hybrid
Approach .................................................................................................... 124
Table 7-9: GA Based Economic Dispatch --- Best Solution After 20 Runs ................... 128
Table 7-10: Best Results by Hybrid Approach amongst the Sixty Samples ................... 129
 xviii

Table 7-11: GA Input Corresponding to the Best Results by Hybrid Approach in Table
7.11............................................................................................................. 130
Table 7-12: GA–Models & GA Parameters for GA-SQP Hybrid Approach ................. 133
Table 7-13: Output Results GA-SQP Approach 3-Machines Test System after 50 Runs
for PD=850 MW ......................................................................................... 134
Table 7-14: Output Results GA-SQP Approach IEEE 5-Machines 14-Bus System After
50 Runs For PD=259 MW .......................................................................... 135
Table 7-15: Output Results GA-SQP Approach IEEE 6-Machines 30-Bus System after
50 Runs for PD=283.4 MW ........................................................................ 136
Table 7-16: Percentage Reduction in Total Generation Cost in GA-SQP Hybrid
Approach .................................................................................................. 137
Table A-1: Test System I: 3-Machines 6 Bus Test System Generator’s Cost Curves & B-
Coefficients .................................................................................................. 155
Table A-2: Test System II: 3-Machines 6 Bus Test System Generator’s Cost Curves & B-
Coefficients .................................................................................................. 155
Table A-3: Test System III: 3-Machines 5 Bus Test System Generator’s Cost Curves &
B-Coefficients .............................................................................................. 156
Table A-4: IEEE 5-Machines 14-Bus System Generator’s Cost Curves ........................ 157
Table A-5: IEEE 6-Machines 30-Bus System Generator’s Cost Curves ........................ 157
Table A-6: Transmission Loss Coefficients IEEE Test Systems .................................... 158
Table A-7: 6-Machine 26-Bus Test System Generator’s Cost Curves ........................... 158
Table A-8: Transmission Loss Coefficients APC 6-Machines 26 Bus Systems ............ 159
Table A-9: 20-Machine Test System Generator’s Cost Curves...................................... 159
Table A-10: 13-Machines Test System Generator’s Cost Curves .................................. 160
Table A-11: 12-Machines 32-Bus NTDC Test Systems Generator’s Cost Curves ........ 160
Table A-12: Transmission Loss Coefficients 12 Machines 32-Bus NTDC System ....... 161
Table C-1: Evolution Models & GA Parameters for Convex Economic Dispatch 3
Machines Test System Using λ – Search & RP Search ............................. 165
Table C-2: Output Results Convex ED Test System–I: 3-Machine System .................. 165
Table C-3: Output Results Convex ED Test System – II: 3-Machine System ............... 166
Table C-4: Output Results Convex ED Test System – III: 3-Machine System.............. 166
Table C-5: Evolution Models for Convex Economic Dispatch for IEEE Test Systems
using λ - Search .......................................................................................... 167
 xix

Table C-6: Evolution Models for Convex Economic Dispatch for IEEE Test Systems
using Real Power - Search ......................................................................... 167
Table C-7: Output Results Convex ED Case – I: IEEE 5-Machines 14 Bus Test Systems
PD = 259 MW ............................................................................................... 168
Table C-8: Output Results Convex ED Case – II: IEEE 5-Machines 14 Bus Test Systems
PD = 259 MW ............................................................................................... 168
Table C-9: Output Results Convex ED Case – I: IEEE 6-Machines 30 Bus Test Systems
PD = 283.3 MW ............................................................................................ 169
Table C-10: Output Results Convex ED Case – II: IEEE 6-Machines 30 Bus Test
Systems PD = 283.3 MW ......................................................................... 169
Table C-11: Evolution Models for Convex Economic Dispatch for American Power
Company 6 – Machines 26 Bus Test System .......................................... 170
Table C-12: Output Results Convex ED APC 6-Machines 26 Bus Test Systems ......... 170
Table C-13: Evolution Model & GA Parameters --- Convex ED Using Hybrid Approach
.......................................................................................................................................... 171
Table C-14: Output Results Convex ED: Hybrid Approach 3-Machine System ........... 171
Table C-15: Output Results Convex ED: Hybrid Approach 20-Machine System ......... 172
Table C-16: Evolution Models for Convex Economic Dispatch for NTDC Test System
Using λ - Search ......................................................................................... 172
Table C-17: Evolution Models for Convex Economic Dispatch for NTDC Test System
using Real Power – Search ........................................................................ 173
Table C-18: Output Results Convex ED 12–Machine 32–Bus NTDC System.............. 173
Table C-19: Output Results Convex ED 15 – Machine NTDC System PD = 2400 MW174
Table C-20: Output Results Convex ED 25–Machine NTDC System PD = 2800 MW . 175
Table C-21: Output Results Convex ED 34–Machine NTDC System PD = 3000 MW . 176
 CHAPTER 1

Introduction

1.1 General
Optimization forms an important place in the day to day life. Most of the peoples
unknowingly apply the optimization procedures in their common daily tasks.
Optimization has vital and dominant role in the field of engineering and sciences.

The electrical power systems the most critical infrastructure has the status of
oxygen for the survival and development of present day world. Economic indicators of
the country have direct relations with energy consumption per capita per annum. The
secure, reliable, and continuous operation of the power system without interruption
subject to quality indicators such as voltage and frequency is the challenging task for the
system operation engineer.

Power systems operation combines a highly non-linear and computationally

difficult environment with a need for optimality. The economic operation and planning of
power system are ranked high amongst the major tasks in the power generation. Power
economic dispatch (ED) is necessary and vital step in power system operational planning
and has always occupied an important position in the electric power industry. This is non-
linear constrained optimization problem and has been solved by various optimization
techniques. Genetic algorithm is the one of the most developed and important paradigm
of the evolutionary computation. This thesis investigates the application of the genetic
algorithm (GA) independently and in hybrid form for the solution of economic dispatch
problem.

1.2 Problem Statement

Economic dispatch, by definition is an on-line function, carried out after every 15-
30 minutes or on request in Power Control Centers. It is defined as the process of

1
calculating the power generation of the generating units in the system in such a way that
the total system demand is supplied most economically.

It is the standard industrial practice that the fuel cost of generator is represented by
polynomial for economic dispatch computation (EDC). The key issue is to determine the
degree and the coefficients such that the error between the polynomial and test data is
sufficiently low. Traditionally, in the EDC, the cost function for each generator is
approximately represented by a quadratic function which is convex in nature.

Present day large power generating units with multi-valves steam turbines exhibit
a large variation in the input-output characteristic functions; as a result non-convexity
appears in the characteristic curves. Major nonconvex economic dispatch problems may
be listed as:
1. Economic dispatch with piecewise quadratic cost function (EDPQ)
• Piecewise quadratic cost function due to valve point effect
• Piecewise quadratic cost function due to multiple fuel mix
2. Economic dispatch with prohibited operating zones (EDPO).

Thus the actual economic dispatch problem is nonconvex in nature requiring the
accurate, robust and fast solution methodology.

Various mathematical programming based optimization techniques have been

applied to solve ED problem. Most of these are calculus-based optimization algorithms
that are based on successive linearization and use the first and second order
differentiations of objective function and its constraints equations as the search direction.
They usually require heat input, power output characteristics of generators to be of
monotonically increasing nature or of piecewise linearity thus resulting in an inaccurate
dispatch. Further to handle the nonconvex economic dispatch problem in classical
optimization dynamic programming has been used, but due to curse of dimensionality it
has limitations. In a nutshell the economic dispatch problem has been solved for many of
years and still no technique has been accepted as yet the best approach.

2
 GA used by John Holland in early seventies basically emulates the optimization
philosophy adopted in nature. The features of the working philosophy of the Genetic
algorithm on the basis of which it carries out optimization may be listed as follows:
1. based on multiple searching points, i.e., population based search,
2. using operators inspired by biological evolution, such as crossover and mutation,
3. based on probabilistic transition rules,
4. fast convergence to near global optimum,
5. superior global searching capability in a complex searching surface using little
information of searching space, such as derivative, continuity.
The key disadvantage of the GA is that its convergence speed near the global optimum
becomes slow. To use GA effectively it is necessary to overcome this deficiency. Hybrid
models based on the integration of GA with another technique to counter balance its
weaknesses may enhance its performance.

GA is the potential solution methodology for nonconvex economic dispatch

problem due to the independence of the objective function from the auxiliary information
such as differentiability, continuity etc. In this thesis efforts have been made to develop
binary coded genetic algorithm to solve economic dispatch problem using GA
independently and GA based hybrid models.

1.3 Objectives
The primary objectives of this thesis are outlined below:
1. To develop flexible and extensible computational framework as general
environment for implementing the various algorithms for economic dispatch
solution.
2. To formulate the economic dispatch problem in GA environment, its
implementation in computational framework and finally testing.
3. To analyze the Performance of GA based evolution models for convex and
nonconvex economic dispatch problem formed with reference to different initial
population encoding schemes, selection methods, crossover operators, and
biological mechanisms as operators.
4. To develop efficient GA based hybrid models through exploiting the strengths of
GA and overcoming its weakness by combing the strengths of other techniques.

3
 5. To compare the results of the proposed hybrid models with results of GA based
approach.
6. To review the operational problem of the National utility system “NTDC” and to
develop its test circuit for economic dispatch studies.

1.4 Scope of the Work

The contributions made in this thesis are:
1. The computational framework “PED_Frame” targeting power economic
dispatch software application development is proposed, designed and
implemented. The version one of the “PED_Frame” targets the
implementation of traditional, GA based and hybrid methodologies for
economic dispatch used in the present research work.
2. Formulation of economic dispatch problem in GA environment based on the
followings:
• λ - Search for convex economic dispatch
• Real power search for convex economic dispatch and nonconvex
economic dispatch due to valve point effect.
3. Two biological methods “Inversion” and “Deletion & Regeneration” have
been mapped as operators with crossover probability.
4. Development of the evolution models for GA with reference to the following:
• Three encoding strategies for initial population
• Four selection methods
• Eight crossover operators
5. Computer implementation of developed models in PED_Frame and testing on
standard test systems and utilities test systems.
6. The development of hybrid methodology for convex ED and its testing.
7. Performance evaluation of different evolution models for nonconvex
economic dispatch based on empirical results obtained on standard 18-
machines system.
8. The development and testing of following three hybrid models for nonconvex
economic dispatch:
a. Hybrid Model-I: A Synergy of GA and Newton’s Second Order Approach
b. Hybrid Model –II: Neuro-Genetic Hybrid Approach.

4
 c. Hybrid Model–III: A Hybrid of GA and Sequential Quadratic
Programming (SQP)
9. Formulation of Economic dispatch using Newton’s Second Order Approach
using sparsity based storage for Hybrid Model-I
10. Development of swarm intelligence learning based feedforward neural
network for Hybrid Model –II.
11. For the national utility system “NTDC” the tasks achieved includes:
• Operational problems of the NTDC systems
• Development of four test systems
• Testing of the four systems for their performance using Newton’s
Raphson Algorithm.
• Development of 12-Machine 32-bus ED test circuit with data including
cost curves, calculation of B-coefficients and system data.
• Development of 15, 25 and 34 machines test systems for ED studies by
neglecting losses.
• ED studies.

1.5 Thesis Outline

Chapter 2 presents the literature survey on economic dispatch problem with focus
on GA based approaches with the view to develop some observations and potential
avenues for further investigations.

Chapter 3 discusses the optimization techniques used in this thesis. These include
the evolutionary computation with focus on genetic algorithm, particle swarm
optimization, artificial neural networks, and local search techniques such as Newton
second order approach and sequential quadratic programming.

Chapter 4 highlights the design and implementation of the computational

framework “PED_Frame” as common environment for implementing the algorithms for
economic dispatch.

Chapter 5 deals with the mathematical modeling of economic dispatch problem in

genetic algorithm environment and algorithm development for GA based economic
dispatch. A number of case studies using standard test systems and utility systems have

5
been included for validation of algorithms. Finally there is nonconvex economic dispatch
solution using various evolution models of GA with a view to evaluate the performance.

Chapter 6 gives the discussion on Pakistani utility system NTDC along with
proposed test circuit and ED studies.

Chapter 7 presents following three hybrid models for nonconvex economic

dispatch.
• Hybrid Model –I: combines the GA with Newton second order approach and
tested on 3-Machine and IEEE 5 and 6-Machine test systems.
• Hybrid Model-II: Neuro-Genetic approach integrates GA with feed forward neural
network with swarm intelligence learning rule and has been tested on standard 3-
Machine test system.
• Hybrid Model-III: combines GA with sequential quadratic programming and has
been tested on the standard test systems.

Chapter 8 discusses the conclusions and suggests the areas for future research.

Appendix A describes the test systems for numerical results. These include the
standard test systems and utility system.

Appendix B includes the discussion with reference to the optimization methods in

chapter 3.

Appendix C includes the detailed results of case studies conducted in chapter 5

and chapter 6.

Appendix D includes the derived publications.

6
 CHAPTER 2

Power Economic Dispatch --- A Brief Survey

2.1 Introduction
Power Economic Dispatch is vital and essential daily optimization procedure in
the system operation. It is a constrained nonlinear optimization problem. This chapter
presents brief overview and literature survey on economic dispatch problem with focus on
GA based approaches. Final section gives review on economic dispatch using GA
independently and in hybrid form with some observations and potential avenues for
further investigations.

2.2 Power System Operational Planning

The operational planning of the power system involves the best utilization of the
available energy resources subjected to various constraints to transfer electrical energy
from generating stations to the consumers with maximum safety of personal/equipment
without interruption of supply at minimum cost. In modern complex and highly
interconnected power systems, the operational planning involves steps such as load
forecasting, unit commitment, economic dispatch, maintenance of system frequency and
declared voltage levels as well as interchanges among the interconnected systems in
power pools etc.

Figure 2.1 illustrates the operation and data flow in a modern power system on the
assumption of a fully automated power system based on real-time digital control [1].

There are three stages in system control, namely generator scheduling or unit
commitment, security analysis and economic dispatch.
• Generator scheduling involves the hour-by-hour ordering of generator units on/off
in the system to match the anticipated load and to allow a safety margin.

7
 • With a given power system topology and number of generators on the bars,
security analysis assesses the system response to a set of contingencies and
provides a set of constraints that should not be violated if the system is to remain
in secure state.
• Economic dispatch orders the minute-to-minute loading of the connected
generating plant so that the cost of generation is a minimum with due respect to
the satisfaction of the security and other engineering constraints.

Load
forecasting

Unit commitment Security analysis Economic dispatch

System topology change

Data base commands

Generator startup
& shutdown State Generator P & Q
estimation loading commands
commands

Measurements, switch
positions

Power system

Figure 2-1 Power System Control Activities

2.3 Economic Dispatch --- Literature Survey

Economic dispatch is generation allocation problem and defined as the process of
calculating the generation of the generating units so that the system load is supplied
entirely and most economically subject to the satisfaction of the constraints [2].

8
 Historically economic dispatch is being carried out since 1920. It was the time
when engineers were concerned with the problem of economic allocation of generation or
the proper division of the load among the generating units available. The developments in
ED may be listed [2] as:
1. Prior to 1930, the methods in use includes: the base load method and best
point loading
2. It was recognized as early as 1930, that the incremental method, later known
as the equal incremental method, yielded the most economic results.
3. The analogue computer was developed to solve the coordination equations.
4. A transmission loss penalty factor computer was developed in 1954.
5. An electronic differential analyzer was developed and used in ED for both off-
line and on-line use by 1955.
6. The digital computer was investigated in 1954 for ED and is being used to
date.

H. H. Happ presented [2] comprehensive survey on economic dispatch. The

survey covers the following aspects:
• Developments in Economic dispatch since early 1920’s.
• Step by step developments taking into account the effect of transmission loss in
economic dispatch problem.
• Multi-area concepts in economic dispatch.
• Valve point loading
• Optimal load flow

Three part series [3-5] presented by IEEE working group 71-2 (W.G. on
Operating Economics) describes the major security –economy functions and bibliography
from 1959 to 1977. Part I [3] contains the descriptions of major economy-security
functions; Part II [4] & Part III [5] give the bibliography. The economic dispatch has been
addressed under “Real Time Operation” function with categories as: 1) Economic
Dispatch [RE], 2) Security Dispatch [RS], and 3) Environmental Dispatch [RV].
B .H. Chowdhury, et al. [6] presented a survey of papers and reports addressing
various aspects of economic dispatch during the period 1977-88. The survey has been
focused under the following four areas of economic dispatch:

9
 • Optimal power flow.
• Economic dispatch in relation to AGC.
• Dynamic dispatch.
• Economic dispatch with non-conventional generation sources.
Mathematical programming based approaches for ED may be classified as:
1. Merit Order approach (Old Method)
2. Equal Incremental Cost Criterion (widely used)
3. Linear Programming (Easy Constraint Handling)
4. Dynamic Programming (Curse of dimensionality)
5. Nonlinear Programming

2.3.1 Thermal Machine’s Cost Curve in Economic Dispatch

Generators in system operation studies are represented by input/output cost
curves. It is the standard industrial practice that the fuel cost of generator is represented
by polynomial [7] for Economic Dispatch Computation (EDC). The key issue is to
determine the degree and the coefficients such that the error between the polynomial and
test data is sufficiently low. Traditionally, in the EDC, the cost function for each
generator has been approximately represented by a single quadratic function. This curve
is convex in nature. Various mathematical and optimization techniques have been applied
to solve ED problem. Most of these are calculus-based optimization algorithms that are
based on successive linearization and use the first and second order differentiations of
objective function and its constraints equations as the search direction. They usually
require heat input, power output characteristics of generators to be of monotonically
increasing nature or of piecewise linearity.

2.3.2 Nonconvex Cost Curve

Present day large power generating units with multi-valves steam turbines exhibit
a large variation in the input-output characteristic functions, thus nonconvexity appears in
the characteristic curves. Two major nonconvex economic dispatch problems may be
listed as [8]:
1. Economic dispatch with piecewise quadratic cost function (EDPQ)
2. Economic dispatch with prohibited operating zones (EDPO).

10
2.3.2.1 Piecewise Quadratic Cost Curve
a. Valve Point Effect
The valve point effects [9], owing to wire drawing as each steam admission valve
starts to open, typically produce a ripple like heat rate curve, which is non-smooth
in nature as shown in Figure 2.2. This type of the characteristic must be used to
schedule steam units accurately, but it can not be used in traditional optimization
methods because it does not meet the convex condition. Smooth quadratic
function approximation of cost curves is used in classical economic dispatch.

Figure 2-2 Value Point Effect

b. Multiple Fuel Mix
Some generation units especially those units which are supplied with multiple fuel
sources (gas and oil) are faced with the problem of determining the most
economical fuel to burn. As fossil fuel costs increase, it becomes even more
important to have a good model for the production cost of each generator.
Therefore a more accurate formulation [10] is obtained for the ED problem by
using hybrid cost function and hybrid incremental cost function and expressing
these generation cost function as a piecewise quadratic function and linear
function respectively as shown in Figure 2.3.

Figure 2-3 Hybrid Cost & Incremental Cost

11
2.3.2.2 Prohibited Operating Zone
Due to the physical restrictions of power plant components (faults in the machines
themselves or the associated auxiliaries), the whole of the unit operating range is not
always available for load allocation; and generating units may have prohibited operating
zones lying between their minimum and maximum power outputs [11]. The operation in
these zones may result in amplification of vibrations in a shaft bearing which should be
avoided in practical application. Usually, each prohibited region makes the decision space
separate into disjoint subsets which then constitute a non-convex decision space.
Therefore, the corresponding economic dispatch problem becomes a non-convex
optimization problem [8].

2.3.3 Economic Dispatch Based on Traditional Optimization--- Bottlenecks

Generally the economic dispatch problem is solved as convex optimization
problem. Mathematical programming based optimization techniques have been developed
to solve economic dispatch (ED) problem. They usually require heat input, power output
characteristics of generators to be of monotonically increasing nature or of piecewise
linearity. These simplifying assumptions result in an inaccurate dispatch.

The actual economic dispatch problem is non-convex in nature due valve point
effect, multiple fuels and prohibited operating zones. Dynamic programming [12] has
been used to solve this problem, but due to curse of dimensionality it has limitations.
Thus for the accurate dispatch, the approaches independent of restrictions such as
continuity, differentiability, convexity are potential solution methodologies.

2.3.4 Genetic Algorithm Based Economic Dispatch

The power engineering community has been studying the feasibility and
application of Artificial intelligence tools for efficient solution of complex power system
problems. AI tools such as 1) Artificial Neural Network (ANN), 2) Expert Systems, 3)
Evolutionary Computation, and 4) Fuzzy Logic have been developed to formulate and
solve the convex and non-convex economic dispatch problem.

Genetic algorithm (GA) is one of the most popular paradigms of evolutionary

computation. It has been used to solve the economic dispatch problem independently and

12
in conjunction with other AI tools and optimization approaches. It has the superior global
searching capability in a complex searching surface using little information of searching
space, such as derivative, continuity thus providing potential tool for real economic
dispatch problem.

Genetic algorithm has inherent ability to reach the global minimum region of
search space in a short or affordable time [13-14]. But at the same time the disadvantages
are: 1) long computational time [15-16], 2) convergence speed near the global optimum
becomes slow, 3) provide near optimal solution [16] or non-convergence to global
solution [17]. There are several ways by which the performance of GA can be enhanced
[18]. Hybrid approach is one of the methodologies used efficiently for producing quality
results. The objective of hybridization is to overcome the weakness of one approach
during its application with the strengths of other by appropriately integrating them.

In the discussion to follow in this section brief review on GA based approaches

has been presented.

In 1993, Yoshimi, et al. [19] mapped ED problem in binary coded Genetic

Algorithm (GA) environment using typical reproduction operator crossover and mutation.
15-machine test system has been used for performance evaluation of GA in terms of
objective function and equality constraints satisfaction for four different case studies and
established the fact that GA performs function optimization in an adaptive manner so as
to schedule the power generation in order to minimize objective function.

In 1993, Walters, et al. [9] has used Simple genetic algorithm (SGA) with two
different encoding schemes for solving the ED problem including valve point effect. The
SGA has been tested for ED problem with & without losses for smooth and non-smooth
cost curves using three machine test system for five different cases by comparing the
results obtained from dynamic programming technique. He made useful observations
regarding the GA parameters during analysis and concluded that GA has the ability to
handle any type of unit characteristics.

In 1994, H. Ma, et al. [20] presented genetic algorithm based solution for ill
structured and multimodal economic dispatch problem including transmission loss

13
considering compensating generation plan provided in the Clean Air Act 1990. The
proposed algorithm has been tested on 9-units system (including 2 hydro units with fixed
outputs).

In 1994, Bakirtzis, et al. [21] presented two genetic algorithms GA–I & GA–II for
the solutions of the economic dispatch problem. Test results with systems of up to 72
generating units with nonconvex cost functions show that both genetic algorithms
outperform the dynamic programming solution to the economic dispatch problem.
However, the solution time of GA-II is system dependent and increases linearly with the
size of system.

In 1995, Po-Hung, et al. [22] presented genetic algorithm based approach for
large-scale economic dispatch subject to the network losses, ramp rate limits and
prohibited zone avoidance. He has selected the encoding scheme such that chromosome
contains only an encoding of the normalized system incremental cost thus making
chromosome size independent of number of units. This feature has been exploited as an
edge for large-scale economic dispatch. The two test systems: 3 machines system and 40
units Tai power systems have been used for demonstration of the approach. The
comparison of results with lambda –iteration method proved that proposed approach is
fast, robust in large systems.

In 1995, Sheble, et al. [23] presented genetic algorithm based solution for
economic dispatch with a view to get around the problems in classical optimization
theory. He proposed the techniques such as mutation prediction, elitism, interval
approximation and penalty factors to enhance the program efficiency and accuracy. Three
units test system has been test for Economic dispatch solution using proposed algorithm.

In 1996, P. Y. Wang, et al. [24] presented fuzzy logic controlled genetic algorithm
for economic dispatch problem. Two fuzzy logic controllers for adoptively adjustment
GA parameters (—crossover rate and mutation rate) have been proposed. Six generators
system has been solved for economic dispatch using the proposed algorithm and
compared the results with conventional GA.
In 1996, Orero, et al. [25] has developed the genetic algorithm solution for
nonconvex economic dispatch problem taking into account the prohibited operating zones

14
of the generators. Two different implementations of genetic algorithms: Standard Genetic
Algorithm & Deterministic Crowding Genetic Algorithm has been presented with a view
to improve the performance (--- robustness and few parameter settings) of genetic
algorithm. 15 generators practical power system with 4 of the units up to three prohibited
operating zones system has been tested on the proposed approaches.

In 1997, Li, et al. [26] investigated the capability of genetic algorithms on the
ramping rate constrained DED problem on 25 generators practical power supply system --
- Northern Ireland Supply (NIE). It has been established that the ramping rate constraint
not only put no additional burden on the genetic search, but also forced the search to
move within the operating feasible region, and gives smoother and more economical
operational strategy for the whole dispatch period. However, a too strict ramping rate
constraint will prevent a GA from obtaining an economic, feasible solution within a
reasonable time. An investigation into choice of appropriate ramping rates for the
generator units in a power system can be time consuming, but highly rewarding.

In 1997, Song, et al. [27] presented the application of a fuzzy logic controlled
genetic algorithm (FCGA) to environmental/economic dispatch. Two fuzzy controllers
have been designed to adaptively adjust the crossover probability and mutation rate
during the optimization process based on some heuristics. For the demonstration of
algorithm six machines system has been solved for economic load dispatch problem. The
results have been compared with conventional GA and Newton-Raphson Method for the
performance evaluation of proposed approach. It has also been recommended that this
algorithm can be applied to wide range of optimization problems.

In 1997, Song, et al. [28] proposed an advanced engineered-conditioning genetic

algorithm (AEC-GA) with a view to improve the performance of GA. This approach uses
combination strategy using local search algorithms and genetic algorithms. In addition to
this elite policy, adaptive mutation prediction, nonlinear fitness mapping and different
crossover techniques have also been explored. The proposed algorithm has been
demonstrated by solving the power economic dispatch problem. The six units test system
results have been for its validation.

15
 In 1998, Chira, et al. [29] has developed genetic algorithm based economic
dispatch solution for eastern region of Electricity Generating Authority of Thailand
(EGAT) system by using piecewise model of cost characteristics for the combined cycle
and the cogeneration power plants in the system. Conventional & parallel micro genetic
algorithms have been used for the solution with result that parallel micro genetic
algorithm offers faster convergence.

In 2000, Angela, et al. [30] developed framework “GAFrame” for constructing

genetic algorithms based on object oriented programming methodology. GAFrame
includes features: GUI, ease of use, code reusability, continuous & integer optimization
problem solution, flexible, and extensible. The two optimization problems: economic
dispatch and distribution system expansion planning has been solved by this framework
for validation of its features.

In 2001, T. Yalcinoz, et al. [31] proposed real coded genetic algorithm for the
solution of economic dispatch problem. Elitism, arithmetic crossover and mutation have
been used to generate solutions in successive generations. The approach has been tested
on 6 machine and 20 machine systems and results have been compared with an improved
Hopfield NN approach (IHN), a fuzzy logic controlled genetic algorithm (FLCGA), an
advance engineered-conditioning genetic approach (AECGA) and an advance Hopfield
NN approach (AHNN) and established the fact that the proposed technique improves the
quality of the solution.

In 2002, T. Yalcinoz, et al. [32] presented the environmental economic dispatch

problem using modified GA, which based on arithmetic crossover operator. The objective
function consists of three terms, which are the production cost and emissions functions.
The proposed algorithm has been tested on 3-unit system and a 10-unit system and results
compared with the taboo search (TS), the Hopfield NN and an improved Hopfield NN
approach.

In 2002, Ying-Yi, et al. [33] solved using binary coded GA for optimal dispatch
among multi-plant (cogeneration systems) with multicogenerators, which transmit MW to
designated buyers (load buses) via wheeling subject to the operation constraints in the
cogeneration systems and security constraints in the third party (transmission system

16
owner). Varying weighting coefficients for penalty functions (constraint handling) and
determination of gene variables for GA has been adopted. The algorithm has been tested
on IEEE 30 and 118-bus systems. From the test results, it has been shown that the
computational time required strongly depends on the number of the gene variables
(Lagrange multipliers) instead of the system size. Also, the proposed method can
efficiently obtain the global optimum with the concept of equal Lagrange multiplier.

In 2002, Jarurote, et al. [34] proposed an algorithm based on parallel micro

genetic algorithm for the solution of ramp rate constrained economic dispatch neglecting
the transmission loss for generating units with non-monotonic & monotonic incremental
cost functions. The proposed algorithm was programmed on thirty-two-processor Bewulf
cluster subject to schedule processors, computational loads, and synchronization overhead
for the best performance. The speedup upper bounds and the synchronization overheads
on the Beowulf cluster for different system sizes and different migration frequencies have
been shown. Claim has been made that PMGA is viable to the online implementation of
the constrained ED due to substantial generator fuel cost savings and high speedup upper
bounds.

In 2002, Patricia, et al. [35] presented GA based solution for the operational
planning of hydro-thermal power systems to get around the deficiencies in nonlinear
programming based approaches. The paper presents an adaptation of the technique and an
actual application on the optimization of the operation planning for a cascaded system
composed by interconnected hydroelectric plants. The fact has been established that by
using GA approach, for each additional problem constraint, only the equation used to
define each individual performance needs to be modified and proposed approaches can be
an efficient alternative or complementary technique for the planning of Hydrothermal
power system operation.

In 2003, Ioannis, et al. developed [36] real coded GA for the solution of network
constrained Power economic Dispatch for minimizing the dispatch cost subject to branch
power flow limits. 52 buses Greek island of Crete system has been tested with convex
cost functions and nonconvex cost functions for proposed algorithm to establish that the
algorithm retained the advantages of the GAs over the traditional ED methods but also

17
eliminated the main disadvantage of the binary coded GAs (long execution time) and thus
providing an efficient generic ED solution method.

In 2004, Liladhur, et al. [37] presented economic dispatch solution considering

valve point loading effect using simple genetic algorithm (SGA), SGA with generation-
apart elitism, SGA with atavism and atavistic genetic algorithm (AGA). These three
concepts have been compared on three test systems: 3-generator system, 13-generator
system and the standard IEEE 30-bus test system. It has been established that when
valve-point loading and ramping characteristics of the generators are taken into account,
AGA outperforms SGA, SGA with generation-apart elitism, SGA with atavism, Tabu
search and conventional Lagrangian multiplier method.

In 2004, Hosseini, et al. [38] discusses the centralized economic dispatch in

deregulated environment. Power economic dispatch problem has been using genetic
algorithm subject to the constraints such as are minimum and maximum power generation
of units, capacity of transmission lines and ramp rate limits. IEEE 30 bus system has been
tested for the proposed algorithm for its validation.

In 2005, Bakare, et al. [39] has made comparative investigation of both

conventional GA and micro-GA for Power economic dispatch problem using 6-bus IEEE
test system and 31-bus Nigerian grid Systems. The results obtained are satisfactory for
both approaches, but it is shown that the µGA performed better than CGA with reference
to the economic and computational time (average of 62% time reduction) for Nigerian
system.
In 2005, Chao-Lung Chiang, et al. [40] presented genetic algorithm solution for
economic dispatch taking into account the valve point effect and multiple fuels
simultaneously with multiplier update (IGA_MU). Basically this algorithm is the
integration of the improved genetic algorithm (IGA) and the multiplier updating (MU).
The proposed algorithm has been demonstrated by applying separately on: 13 units test
system considering valve-point effects only, 10 units considering multiple fuels only, and
ten units addressing both valve-point effects and multiple fuels. By comparison of results
of proposed algorithm with conventional genetic algorithm with multiplier update, it has
been established the fact that IGA_MU is more effective and can apply to the real world
problems.

18
 Wong, et al. in [41] presented Simulated Annealing (SA) based economic dispatch
algorithm. The algorithm was demonstrated on three generating units system only. In
comparison with Zoom Dynamic Programming (ZDP) method [42] economic dispatch
results obtained for the test system are more economical. Subsequently in [43] four
algorithms ( two GA based algorithms, Basic Genetic Algorithms (BGA) and Incremental
Genetic Algorithm (IGA), two Genetic simulated annealing based algorithms GGA and
GAA2) have been presented to determine the optimal or near optimum solution of the ED
problem. By modifying step 3 of four steps BGA algorithm, an IGA algorithm has been
derived which can find global optimum solution earlier in the solution process than BGA.
To avoid premature convergence in IGA, hybrid algorithm GAA is developed by
combining IGA and simulated annealing. Further GAA2 is developed to deal with the
memory requirement by reducing the population size to 2. 13 Machine real life system
has been tested and established the fact GAA2 is superior than other algorithms.

Ongsakul, et al. [44] proposed the micro genetic algorithm based on migration and
merit order loading solutions (MGAM-MOL) for solving the constrained ED with linear
decreasing and decreasing staircase IC functions. MGAM-MOL used a merit order
loading (MOL) solution as a base solution to reduce the search effort to the optimal
solution. The MGAM-MOL solutions were less expensive than those obtained from
simple genetic algorithm (SGA), micro genetic algorithm (MGA), and MOL.

Ongsakul, et al. in [45] presented a genetic algorithm based on simulated

annealing solutions (GA-SA) to solve ramp rate constrained Dynamic Economic Dispatch
(DED) problems for generating units with non-monotonically and monotonically
increasing IC functions. As the transmission line losses are incorporated, GA-SA is tested
on the 10 generating unit systems and compared to the zoom brute force (ZBF), ZDP, SA,
local search (LS), genetic algorithm based on MOL solutions (GA-MOL), and MOL
methods. To illustrate the quality of GA-SA solutions on larger generating unit systems,
the algorithm has also been tested and compared to the others on the 20 and 40 generating
unit systems.

Yalcinoze, presented [46] hybrid genetic algorithm (HGA) for solving the
economic dispatch problem. The algorithm incorporates the solution produced by an
improved Hopfield neural network (NN) as a part of its initial population. Elitism,

19
arithmetic crossover, and mutation are used in the GAs to generate successive sets of
possible operating policies. The technique improves the quality of the solution and
reduces the computation time, and is compared with the classical optimization technique,
an improved Hopfield NN approach (IHN), a fuzzy logic con-trolled GA (FLCGA), and
an improved GA (IGA).

Attaviriyanupap, et al. in [16] proposed hybrid methodology for solving DED.

The proposed method has been developed in such a way that a simple evolutionary
programming (EP) has been applied as a based level search, which can give a good
direction to the optimal global region, and a local search sequential quadratic
programming (SQP) has been used as a fine tuning to determine the optimal solution at
the final. A ten-unit test system with non-smooth fuel cost function has been used to
illustrate the effectiveness of the proposed method compared with those obtained from EP
and SQP alone.

Kumarappan, presented [47] Neuro hybrid Genetic Algorithm (GA) to solve an

economic dispatch problem. The algorithm has been proposed for minimum cost of
operating units. Working philosophy of algorithm is as: 1) Real Coded GA has been used
for global search 2) fine tunings by Tabu Search (TS) to direct the search towards the
optimal region and local optimization 3) For loss calculation Fast Decoupled Load Flow
(FDLF) conducted to find the losses by substituting the generation values to the
respective PV buses. Then the loss has been participated among all generating units using
participation factor method. Artificial Neural Networks (ANN) has been applied to the
Hybrid GA. The algorithm is tested on IEEE 6-bus system and 66-bus utility system.

2.3.4.1 Discussion
The focus of GA based approaches for economic dispatch solution is around the
aspects such as: convex [31] & nonconvex ED problem [9, 21, 22, 23, 25, 37],
environmental concern [20, 27, 32], power market [33,38] and handling of combined
cycle plant. Evolution model adopted includes: binary coded GA [9, 19, 20, 21, 22, 23]
with different schemes for picking bits in decoding step, real coded GA [31, 36], micro
GA [29] and parallel GA [34]. Standard test systems of various sizes in terms number of

20
machines [9, 19, 20, 21, 22, 23, 25] and utility systems [29, 36, 39] have been used during
investigations.

In the deregulated environment where cost minimizations not only the objective,
but at the same time profit maximization is the also the concern. Fast and accurate
economic dispatch solution is as usual the requirement in deregulated scenario as well.
Despite the advances in GA based approaches, potential avenues for further exploration
may be listed as:
1. Flexible and extensible computational framework as common environment for
implementing economic dispatch algorithms.
2. Some methodology for GA parameters for economic dispatch.
3. Mapping of biological mechanism in terms of GA operators in economic
dispatch problem.
4. Performance evaluation of GA evolution models for economic dispatch
problem.
5. Exploration of GA based ED approaches for the utility systems.
6. Hybrid methodology is the useful tool for efficient solution by exploiting the
strengths of GA with the powers of other techniques by appropriate integration
of both the approaches. So this sector is the potential area for developing
simple fast and efficient GA based hybrid methodologies.

21
 CHAPTER 3

Optimization Methods

3.1 Introduction
The process of optimization lies at the root of engineering, since the classical
function of engineer is to design new, better, more efficient, and less expensive systems
as well as devise plans and procedures for the improved operation of existing systems.
Electrical Power systems the critical infrastructures are designed and operated with spirit
and sole of optimization. This chapter presents the discussion on genetic algorithm,
particle swarm optimization (PSO), artificial neural network (ANN), Newton’s second
order approach (NSOA), sequential quadratic programming (SQP), and hybrid system
with the view to their use in the subsequent chapters.

3.2 Evolutionary Computation

The field of evolutionary computation (EC) is referred to by some as
computational intelligence (CI) and by others as soft computing. Evolutionary algorithms
(EAs) are computer-based problem-solving systems based on principles of evolution
theory. Different EAs have been developed and they all share a common conceptual base
of simulating the evolution of individual structures via the processes of selection,
mutation and recombination. The processes depend on the perceived performance of the
individual structures as defined by an environment.

Evolutionary computation paradigms generally differ from traditional search and

optimization paradigms in three main ways [48]:
1. Evolutionary computation paradigms utilize a population of points in their search.
2. Evolutionary computation paradigms use direct “fitness” information instead of
function derivatives or other related knowledge.
3. Evolutionary computation paradigms use probabilistic, rather than deterministic,
transition rules.

22
 In addition, evolutionary computation sometime encodes the parameters in binary
form or other symbols rather than working with the parameters themselves.

Regardless of the paradigm implemented, evolutionary computation tools often

follow a similar procedure [48]:
1. Initialize the population.
2. Calculate the fitness for each individual in the population.
3. Reproduce selected individuals to form a new population.
4. Perform evolutionary operations, such as crossover and mutation, on the
population.
5. Loop to step 2 until some condition is met.

The poplar evolutionary algorithms may include: genetic algorithm, evolution

strategies, evolutionary programming, genetic programming, simulated annealing,
particle swarm and tabu search. Being robust and powerful adaptive search techniques
there is growing interest in evolutionary algorithms. The discussion to follow will focus
around two paradigms: GA and PSO with reference to their use in the present research
work.

3.3 Genetic Algorithm

Genetic Algorithms (GAs) have been developed by John Holland and his
colleagues, and students at university of Michigan. These are the robust search
mechanisms based on the mechanics of natural genetics and are very different from most
of the traditional optimization methods. The main features of their working strategy are:
(1) population based search i.e., multipoint start, (2) operator inspired by biological
evolution, such as crossover and mutation, (3) probabilistic transition rules, (4) fast
convergence to near global optimum, (5) superior global searching capability in a
complex searching surface using little information of searching space, such as derivative,
continuity [49].

The working of GA’s is actually a balanced combination of exploration of new

regions in the search space and exploitation of already sampled regions. This balance
critically controls the performance of GA’s and achieved by the right choice of GA

23
control parameters: (1) Population size, (2) chromosome length, (3) probability of
crossover, and (4) probability of mutation. The essence of GA mechanic is based upon the
schema theory and the building-block hypothesis proposed by Holland and Gold-berg.
Figure 3.1 depicts the working of simple GA.

Binary coded GA works by encoding the problem variables into string of 1’s and
0’s. Holland’s genetic algorithm commonly called the Simple Genetic Algorithm (SGA)
works with a population of binary strings of 0s and 1s.

The genetic algorithm solution methodology requires GA parameter selection,

population of individuals for initialization, encoding, decoding, fitness evaluation for
selection, application of crossover and mutation operators to obtain new strings. The
subsections to follow covers brief discussion on the components of GA.
Step1: Read system data and select GA parameters
Step 2: Generate initial population
Step 3: decode chromosomes
Step 4: Evaluate Fitness for each chromosome
Step 5: if (termination criteria satisfied) GOTO Step 8
Step 6: else {
perform selection
crossover & mutation
new population
}
Step 7: GOTO Step 3
Step 8: Exit
Step 9: Print Final Results

Figure 3-1 Simple Genetic Algorithm Structure

3.3.1 Encoding & Decoding

Implementation of problem in genetic algorithm starts from the parameters
encoding. The encoding must be designed carefully to utilize the genetic algorithm’s
ability to efficiently transfer information between chromosome strings and objective
function of the problem. There are different encoding schemes for variables. In binary
encoding variables are represented by 1s and 0s. Binary coded strings are generated
randomly and equivalent decimal integer of binary string yj is obtained as [13]:

24
 l −1
y i = ∑ 2 j bij (3.1)
j =0

Actual value of the continuous variable xj in search space is given by fixed mapping rule
max min
min xi + xi
xi = xi + yi (3.2)
2 −1
l

Where; l is the length of the string

L is the number of strings of population size.
bij is the jth binary digit of the ith string
The subscript i in both the equations is 1, 2, …, L

3.3.2 Chromosome Length

The chromosome is formed by concatenating the strings of problem variables.
Each variable in the string may be of same fixed length or of different length [50] is given
by

l i ≥ log 2 ( xi max − xi min

∆x
)
+1 (3.3)
Equation (3.3) gives the number of binary digits needed to represent a continuous
variation in accuracy of ∆x with ximin , ximax the minimum and maximum value respectively
for the variable xi.

3.3.3 Fitness Function

Genetic Algorithm basically mimics the survival of the fittest principal of nature
for searching the search space. So the fitness evaluation has grate importance in GA
functioning. In order to emphasize the best chromosome and speed up the convergence of
evolutionary process fitness is normalized in to the range of 0 and 1. The fitness function
of ith chromosome [50] is given by
⎛ F ( x) ⎞
f i ( x) = 1 ÷ ⎜1 + k imin − 1⎟ (3.4)
⎝ F ⎠

Where;
Fi(x) is the solution corresponding to ith chromosome
Fmin is the solution of the highest ranking chromosome
k is the scaling factor

25
3.3.4 Genetic Algorithm Operators
The GA operators may be classified as:
1. Reproduction or Selection
2. Crossover
3. Mutation

3.3.4.1 Reproduction or Selection

Reproduction is the first operator applied on the population. Chromosomes are
selected from the population as parents to crossover and produce offspring in such a way
that best should survive and create offspring. These offsprings are the basis for the next
generation. So it is desirable that mating pool should consist of good individuals. A
selection strategy in GA is simply a process that favours the selection of better individuals
in the population for the mating pool.

There exist number of GA reproduction operators in the literature, but the main
idea in them is that the above average strings are picked from the current population and
their multiple copies are inserted in the mating pool in a probabilistic manner. The
various methods for selection [13] are:
1. Roulette wheel selection (RWS)
2. Stochastic remainder selection (SRS)
3. Tournament selection (TS)
4. Rank selection (RS)
5. Boltzmann selection
The SGA uses the Roulette wheel selection to implement proportionate selection.

Selection more than crossover and mutation is the operator responsible for
determining the characteristics of GAs [51]. Population diversity and selective pressure
are the two important issues in the evolution process of genetic search [52]. Higher the
selective pressure, the more the better individuals are favoured. The convergence rate of
GA is largely determined by selective pressure and population diversity. In fact higher
selective pressure results in higher convergence rate. If the selective pressure is too high,
there is chance of premature convergence. If the selective pressure is too low, the
convergence will be slow. An ideal strategy is to adjust the balance between selective
pressure and population diversity so as to fine tune the GA performance [13].

26
 The basic roulette wheel selection method is stochastic sampling with
replacement. Remainder sampling methods involve two distinct phase, individuals are
selected deterministically according to the integral part of their expected trials. The
remaining individuals are then selected probabilistically from fractional part of the
individual’s expected value. Remainder stochastic sampling with replacement uses
roulette wheel selection phase, individual’s fractional parts remain unchanged and
compete for selection between spin [50].

Fitness proportionate selection is likely to lead to two problems, namely

1. Stagnation of the search space because it lacks selection pressure, and
2. Premature convergence of search because it causes the search to narrow down too
quickly.

The tournament selection strategy provides selective pressure by holding

tournament competition among N individuals [53]. The winner from the tournament is the
one with highest fitness among N. The tournament competition is repeated until the
mating pool for generating new offspring is filled.

The roulette wheel will have the problem when the fitness values differ very
much. The low fitness chromosome will have little chance to select. In rank selection [13]
individuals are sorted according to their fitness values and the last position is assigned to
the best individual, while the first position is allocated to the worst one. The selection
probability is assigned to the individuals according to their ranks. All individuals get
different selection probability even when equal fitness values occur.

3.3.4.2 Crossover Operator

The crossover operator plays vital and central role in GA working, in fact it may
be considered to be one of the algorithm’s defining characteristics. This operator provides
mechanism for sharing information with probability of crossover between chromosomes
by combining two parents chromosome to produces offsprings with the possibility that
good chromosomes may generate better ones. There are three types of crossover
operators: (1) one point crossover, (2) multipoint crossover and (3) uniform crossover.
The SGA uses the single point crossover.

27
 In single point and multipoint crossover segment of bits are exchanged between
cross sites whereas uniform crossover exchanges bits of a string rather than segments. At
each string position, the bits are probabilistically exchanged with some fixed probability.

To compare the crossover techniques the concept of positional and distributional

biases has been used. Single-point crossover exhibits the maximum positional bias and
the least distributional bias whereas uniform crossover, at the other end has maximum
distributional bias and minimum positional bias.

3.3.4.3 Mutation
` It is the goal of learning algorithm to search always in regions not viewed before.
Therefore it is necessary to enlarge the information contained in the population. One way
to achieve this goal is mutation. Mutation operator changes 1 to 0 and vice versa with a
small probability. The operator injects new genetic material into the population to prevent
the premature convergence of GA to suboptimal solutions. The SGA treats mutation only
as a secondary operator with the role of restoring lost genetic material.

3.3.5 GA Parameters
The performance of GA depends on choice of GA parameters. While selecting
parameters following points are very important:
1. Higher crossover probability will increases recombination of building blocks, but
at the same time it will also increase the disruption of good strings.
2. Higher mutation probability tends to transform the genetic search into a random
search, but at the same time it also helps to add the new genetic material.
3. Higher population size increases its diversity and reduces the chances of
premature converge to a local optimum, but the time for the population to
converge to the optimal regions in the search space will also increase.
Thus the GA parameters selection requires the balanced compromise among various
options for a particular problem.

28
3.3.6 Termination Criteria
It may be a fixed number of generations or some rule specific to the problem.

3.4 Particle Swarm Optimization (PSO)

Particle swarm optimization (PSO) presented by Kennedy and Eberhart [54-55] is
one of the evolutionary computation technique based on the social behavior. Since its
introduction it has attracted lot of attentions from the researchers around the world. PSO
models problem as a set of particles each representing a dimension of solution space.
These particles move in solution space in search of optimal solution. The particles follow
three principles as described by Kennedy [56]:
Evaluating: Learning through self experience,
Comparing: Learning through comparative study,
Imitating: Learning through adapting the best trend.

In PSO algorithm a group of particles called swarm fly through the search space.
Each particle in the swarm represents the potential solution of the problem. The particle
of the swarm in the search space has two important attributes: velocity and position. The
position of the particle is influenced by the best position visited by it and the position of
best particle in its neighborhood. Based upon the neighborhood there are two PSO
algorithms: gbest and lbest.

In gbest model the entire swarm is the neighborhood and the best position in the
neighborhood is referred to as global best particle, whereas in lbest model small
neighborhoods are used. The performance of each particle is measured using a fitness
function that varies depending on the problem.

The social behavior of particle in a swarm is given by equation (3.5) as follows

[57]:

vi, j (t + 1) = vi, j (t ) + r ⎡ yi, j (t ) − xi, j (t )⎤ + r ⎡ yˆi, j (t ) − xi, j (t )⎤ (3.5)

1 ⎢⎣ ⎦⎥ 2 ⎢⎣ ⎦⎥
Where; xi : the current position of particle,

vi : the current velocity of particle,

29
 y i : the personal best position of the particle, and

ŷ i : an instance of xi that is visited by the particle and yielded best result so far or
the neighborhood best position of the particle.
There are following three components in equation (3.5):
1. First component vi, j (t ) represents the previous rate of change.

2. [ ]
Second component r1 yi , j (t ) − xi , j (t ) refers to self experience, comparing current

position against the personal best.

3. [ ]
The last part is r2 yˆ i , j (t ) − xi , j (t ) , which is the social cognitive term, in that ŷ is

the best among the local (or personal) best position found so far by any particles. This
position of particle is called global best or gbest.
If objective function is given by f then update function for yi is given by

⎧ y i (t ) if f ( x i (t + 1)) ≥ f ( y i (t ))⎫
y i (t + 1) = ⎨ ⎬ (3.6)
⎩ xi (t + 1) if f ( xi (t + 1)) < f ( y i (t )) ⎭

For the gbest model, the best particle is determined from the entire swarm by selecting the
best amongst the personal best position. If the position of global best particle is denoted
by vector ŷ in the swarm size s, then
yˆ (t ) ∈ { y 0 , y1 ,..., y s } = min{ f ( y 0 (t )), f ( y1 (t )),..., f ( y s (t ))}

The velocity and position update equations are given by

[ ] [ ]
vi , j (t + 1) = vi , j (t ) + c1r1, j (t ) yi , j (t ) − xi , j (t ) + c2 r2, j (t ) yˆ i , j (t ) − xi , j (t ) (3.7)

x i (t + 1) = x i (t ) + v i (t + 1) (3.8)

Where;
vi j represents the jth element of the velocity vector of ith particle.
r1 , r2 ∈ U (0,1) are independent random sequences which are scaled by constant

0< c1 , c 2 <=2. These constants are called acceleration coefficients.

The original formula developed by Kennedy and Eberhart was improved by Shi
and Eberhart [57-58] with the introduction of an inertia parameter w to prevent premature
convergence. Another parameter constriction factor χ was introduced with the hope that
it can ensure that PSO will converge [59-60]. The complete PSO formula is:

30
 [ ] [ ]
vi , j (t + 1) = χ (vi , j (t ) + c1r1, j (t ) yi , j (t ) − xi , j (t ) + c2 r2, j (t ) yˆ i , j (t ) − xi , j (t ) ) (3.9)

Where
2
χ= (3.10)
2 − ϕ − ϕ 2 − 4ϕ

Where ϕ = c1 + c 2 , ϕ > 4 (3.11)

3.4.1 Guaranteed Convergence PSO (GCPSO) Approach

Bergh [61] showed that a dangerous condition arises when a particle position
approaches closer to the global best position i.e., xi = yi = yˆ , the particle now will only
depends on the term wvi, j (t ) . In simple words when particle position coincides with the

global best it will move only if w and v are non zero. Therefore if the previous velocity
of particle is close to zero all particles will stop moving. To solve this issue Bergh
modified the equation used by the global best particle as follows:
vτ , j (t + 1) = − xτ , j (t ) + yˆ j (t ) + wvτ , j (t ) + ρ (t )(1 − 2r2, j (t )) (3.12)

Where τ is the index of global best particle so that yτ = yˆ . The term − xτ , j (t + 1) reset the

particle position so that the particle only depends on ŷ j . The term ρ (t )(1 − 2r2, j (t ))

generates random sample around ŷ j for searching better value of ŷ j . The position update

equation for the global best particle is given by

xτ , j (t + 1) = yˆ j (t ) + wvτ , j (t ) + ρ (t )(1 − 2r2, j (t )) (3.13)

The value of ρ (t ) is changed after each iteration using the following rule
⎧ 2 ρ (t ) if N S > s c
⎪ (3.14)
ρ (t + 1 ) = ⎨ 0 . 5 ρ (t ) if N F > f c
⎪ ρ (t ) otherwise
⎩

Where NS and NF denotes the number of consecutive successes and the number of failures
of the search in minimizing the objective function, sc and f c are upper threshold values.
The following two rules are required for implementation:
N S (t + 1) > N S (t ) ⇒ N F (t + 1) = 0

N F (t + 1) > N F (t ) ⇒ N S (t + 1) = 0

31
The optimal choice for the values of sc and f c depends upon the objective function. In
high dimension search spaces it is difficult to obtain better values using random search in
only a few iterations, so it is recommended to set s c = 15, f c = 5 .

3.4.2 Parameters Settings in PSO Model

The PSO algorithm with constriction factor can be considered as a special case of
PSO algorithm with inertial weight, while three parameters χ ,c1, and c2 are connected by
equations (3.10) and (3.11) [62]. The better approach is to use as the rule of thumb is to
utilize the PSO with constriction factor while limiting Vmax to Xmax, the dynamic range of
each variable on each dimension, or utilize PSO with inertial weight while selecting w, c1,
c2 according to the equations (3.10) and (3.11) [63]. For the Clerc’s constriction method φ
is commonly set 4.1.
Let c1 = c2 = 2.05, then from equation (3.11) φ=.2.05+2.05=4.11
Substituting the value of φ into (3.10) yields χ = 0.7298
Using this value of χ in equation (3.9) and dropping to the reference t, results in
vi , j (t + 1) = 0.7298(vi , j + 2.05r1, j ( y i , j − xi , j ) + 2.05r2, j ( yˆ i , j − xi , j ))

As 0.7298* 2.05 = 1.496 so this is equivalent to using c1 = c2 = 1.496 and χ = 0.7298 in

(3.9).
If the range of the search space is defined by [-,Xminx ,Xmax], the value of Vmax is set
so that Vmax = k* Xmax where 0.1≤ k ≤1 [64].

3.5 Artificial Neural Networks

Artificial Neural Networks (ANN) in general is highly interconnected network of
a large number of processing elements called neurons in an architecture, inspired by the
brain. ANN exhibits characteristics such as mapping, generalization, robustness, and fault
tolerance. Neural network learn from examples and various learning laws exist of which
supervised and unsupervised are popular. For a particular application the neural network
is defined by its architecture and learning rule.

32
3.5.1 Feedforward Neural Networks
Error back propagation is a systematic and well known method of training
multilayer feedforward artificial neural networks. In literature, there are several forms of
backpropagation, all of them requiring different levels of computational efforts. The
conventional backpropagation method is however, the one based on gradient descent
algorithm [65]. Here the backpropagation searches on the error surface by means of
gradient descent technique in order to minimize error criterion given by
1
E=
2
∑ (T j − O j ) 2 (3.15)

Where T j the target output, and O j is the output calculated by the network.

Backpropagation training algorithm as the most universal network training algorithm has
the intrinsic defects such as depending on choosing of initial weights, slow training speed
and existence of local minimum points, what lead to the inconsistency and
unpredictability of the output of neural network [66].

PSO has been applied successfully for the training of neural networks while
addressing the shortcomings in the backpropagation learning. Particle swarm can directly
evolve ANN by considering it as an optimization problem. For a given problem, neural
network can be designed based on swarm intelligence learning rule [65]. PSO have been
used to train multilayered neural network to solve variety of problems. Gudise, et al. [67]
presented the comparison of PSO and backpropagation algorithms for training the neural
network and showed that swarm learning rule is better than backpropagation learning.

3.6 Optimization Using Newton's Approach

The object of optimization is to select the best possible decision for a given set of
circumstances without having to enumerate all the possibilities. Three basic steps required
in the optimization problem may be listed as follows:
1. Mathematical model of the problem.
2. The objective function or performance index of the problem.
3. The optimization procedure.

This section is intended to present the theoretical basis for Newton's method for
optimization. The case for finding the minimum of a single variable quadratic function,

33
the minimum of the multivariable function is presented first and then the method is
extended to find the minimum of a multivariable function subject to equality constraints.
This is the general non linear optimization problem of minimizing a Lagrangian function.
The minimum of the Lagrangian function is found in exactly the same way as minimizing
a multivariable nonlinear function. The discussion on the Newton’s approach has been
presented based on [68-70].The discussion on single variable and multivariable functions
is given in appendix B. The discussion to follow covers the minimization of Lagrangian
function.

3.6.1 Minimizing the Lagrangian Function

In the general optimization problem, it would not only be necessary to optimize an
objective function but the solution also has to satisfy some constraints, in the form of
equalities and inequalities. The inequalities are made active, when necessary, by changing
them to equalities. Thus the general optimization problem is to find the optimum of a
function subject to satisfying a set of equalities. This problem is formulated as the
minimization of a Lagrangian function. The objective function is nonlinear and some or
all of the constraints may be nonlinear. Any point that satisfies the equality and inequality
constraints is feasible. A local minimum is a feasible point that minimizes the objective
function within some neighborhood. In the practical problems, the search for minimum is
usually confined to the neighborhood in which the local minimum is the sought after
global minimum.

The formulation of the Lagrangian equation for a given active set and the
determination of its minimum using Newton's method are very briefly discussed here.
The problem may be expressed as;
Minimize F ( y)

Subject to α i ( y) − β i = 0 for i = 1, 2,…, M.

Where;
y is the vector of N variables,
F ( y ) is the nonlinear objective function,

α i ( y ) − β i = 0 is a linear or nonlinear equation in y, with ß as the constant.

34
The number of equations M must be less than the number of variables N, so that some of
the variables are free to optimize F(y).
The Lagrangian function for the general problem can be written as:
L(y ,λ ) = F(y) - λ T [ ai (y) - β i ] (3.16)
Where;
L(y,λ) is the Lagrangian function
λT is the transposed vector of Lagrangian multipliers
The negative signs in equation (3.16) are merely conventions.

The problem is to find y* and λ* so that L(y*,λ*) is minimum. At this point the M
equations will be solved for y* and F(y*) will be at its minimum subject to the solution of
M equations. Since both y* and λ* are unknowns, they are both variables of the
Lagrangian function. In order to solve for y*, it is also necessary to find λ*. In most
practical problems λ* is not merely a meaningless quantity. It also provides useful
information about the optimum. Each Lagrangian multiplier λi measures the rate of
change of the objective function with respect to the equality constraint i at the minimum
y*.

3.6.2 General Non-Linear Optimization Problem

For quadratic function with active set consisting of linear equations, the Lagrangian
function can be written as:
L(y ,λ ) = [ 12 y T G y + b T ( y)] - λ T [ Ay - β ] (3.17)

Where; G: symmetric (NxN) Hessian matrix.

b: vector of N constraints
A: (MxN) coefficient matrix of M linear eqations.
β: vector of constants for M linear equations
Since L(y, λ) is quadratic in (y, λ) it can be minimized by using Newton’s approach as
minimizing multivariable quadratic function.
The gradient or first derivative of Lagrangian function L(y, λ) is given by

⎡ ∇L(y) ⎤ ⎡Gy + b - A λ ⎤
T

∇L(y, λ ) = ⎢ =
⎥ ⎢ ⎥ (3.18)
⎣∇L( λ )⎦ ⎢⎣ - Ay - β ⎥⎦

35
Where; ∇L(y, λ ) : vector of (N+M) function.
∇L(y) : function of y and λ.
∇ L( λ ) : function of y only.
The second partial derivatives of Lagrangian function L(y, λ) forms (N+M) x (N+M)
matrix, which apparently does no have ay name called here as W-matrix. This matrix is
subdivided into four submatrices as below;
⎡ G - AT ⎤
W = ∇ L(y,λ ) = ⎢
2
⎥ (3.19)
⎣− A 0 ⎦
W is constant in the case of quadratic functions. The gradient vector is set equal to zero
for stationary points, and following pair of equations are obtained:
∇L(y) = Gy + b - A λ T
(3.20)
∇L( λ ) = − Ay + β

⎡ G - A T ⎤ ⎡ y ⎤ ⎡− b ⎤
*

⎢ ⎥⎢ ⎥ = ⎢ ⎥ (3.21)
⎣− A 0 ⎦ ⎣⎢λ* ⎦⎥ ⎣− β ⎦

When W is nonsingular, equation (3.21) has unique solution (y*,λ*). This solution point is
either stationary point or true minimum. In general cases where F(y) may higher order,
then quadratic and the active set is partly or completely nonlinear, the expressions for
∇L(y, λ ) and ∇ 2 L(y,λ ) in equations (3.18) and (3.19) respectively still apply. ∇L(y, λ ) is
no longer linear in y and λ and W is no longer a constant matrix.
For the general nonlinear problem, the gradient vector is given by
⎡∇L(y) ⎤ ⎡ H(y,λ )y + d - J T (y)λ ⎤
∇L(y,λ ) = ⎢ ⎥= ⎢ ⎥ (3.22)
⎣∇L( λ )⎦ ⎣ - ai (y) - β i ⎦
W-matrix can be written as:
⎡ H(y,λ ) - J T (y)⎤
W ( y, λ ) = ∇ L( y, λ ) = ⎢
2
⎥ (3.24)
⎢⎣ - J(y) 0 ⎥⎦
Where;
1. Subvector d comes from the linear terms, if any, of F(y).
2. J(y) is the Jacobian matrix of the equations in the active set. It is of the order
(MxN) matrix of first partial derivatives of M nonlinear equations, [ai(y) - ßi] with
respect to y and is function of y.
3. H(y,λ) is the (NxN) Hessian matrix of second partial derivatives of L(y,λ) with
respect to y and is the function of y and λ.

36
 4. W-matrix is symmetric because H(y,λ) is symmetric. It always has an (MxM) null
matrix at its lower right corner because subvector ∇L(λ) is a function of y only.
5. The (NxN) matrix G(y) of F(y) is one of the components of H(y,λ). Therefore the
second partial derivatives of the form (ð2L/ðλiλj) vanishes because λ is not a
variable in the gradient vector element (ðL/ðλi).
6. [ai(y) - ßi] is a vector consisting of M nonlinear equations
When W(y,λ) is not a constant matrix, it is not possible to solve directly for
L(y*,λ*). However, by using Newton’s method for finding the minimum of higher order
multivariable function, it is possible to develop an iterative scheme for finding (y*,λ*) the
optimal point.
The matrix equation for iterative correction (∆y,∆λ), in applying Newton's method
for minimizing a general Lagrangian function, is given by
⎡ H(y,λ ) - J T (y)⎤ ⎡ ∆y ⎤ ⎡ - ∇L(y) ⎤
⎢ ⎥⎢ ⎥ = ⎢ ⎥ (3.25)
⎢⎣ - J(y) 0 ⎥⎦ ⎣∆λ ⎦ ⎣- ∇L( λ )⎦

The matrix and vector functions have to be evaluated for a given (y,λ) in order to
obtain the numerical values. For the sake of convenience of algorithm description vector
Z composed of subvectors y and λ is introduced in such a way that g(Z) = ∇L(y,λ). The
equation for iterative solution at iteration K can be written as:
W( Z K )∆ Z K +1 = - g( Z K )
(3.26)
Z = Z +∆Z
K +1 k K +1

An outline of Newton’s algorithm for minimizing a general Lagrangian function is

given as follows:
1. Initialize the iteration count K = 0
2. Estimate Z* as ZK
3. Evaluate the gradient g(ZK) and W-matrix W(ZK)
4. Test for the minimum:
• If ZK = Z*, the problem is solved ; EXIT
• ELSE CONTINUE
5. Solve W(ZK) ∆ZK+1 = - g(ZK) for ∆ZK+1
6. Update the approximate solution to ZK+1 using:
ZK+1 = ZK + ∆ZK+1 and update iteration count K = K + 1

37
 7. Go to step 3.

3.7 Sequential Quadratic Programming (SQP)

Generally in constrained optimization, the aim is to transform the problem into an
easier sub-problem, which can then be solved and used as the basis of an iterative
process. A characteristic of a large class of early methods is the translation of the
constrained problem to a basic unconstrained problem by using a penalty function for
constraints, which are near or beyond the constraint boundary. With this strategy the
constrained problem is solved using a sequence of parameterized unconstrained
optimizations, which in the limit converge to the constrained problem. These methods are
now considered relatively inefficient and have been replaced by methods, which have
focused on the solution of the Kuhn-Tucker (KT) equations [71]. The KT equations are
necessary conditions for optimality for a constrained optimization problem. If the
problem is a convex problem, that is f(x) and g (x) are convex functions, then the KT
equations are both necessary and sufficient for a global solution point.
For the problem
Minimize f (x ) (3.27)
subject to
g i ( x) = 0 i = 1,2,3,..., me (3.28)

g i ( x) ≤ 0 i = me + 1,..., m (3.29)
The Kuhn-Tucker equations can be stated as [71]:
∇L( x * , λ* ) = 0 (3.30)

λ*i g i ( x * ) = 0 i = 1,2,3,...,me (3.31)

λ*i ≥ 0 i = me + 1,..., m (3.32)

The Lagrangian function L and is given by
m
L ( x, λ ) = f ( x ) + ∑ λ i g i ( x ) (3.33)
i =1

Where λ is the Lagrangian multiplier, me and m the number of equality and inequality
constraints respectively.
The solution of the KT equations forms the basis to many nonlinear programming
algorithms including SQP.

38
 SQP methods represent the state of the art in nonlinear programming methods. As
with most optimization methods, SQP is not a single algorithm, rather conceptual method
from witch several specific algorithms have evolved.

The main idea in SQP is to obtain a search direction by solving a quadratic program; that
is a problem with a quadratic objective function and linear constraints. This approach is a
generalization of Newton’s method for unconstrained minimization [72].
The formula for Newton’s Method is given by
⎡ x k +1 ⎤ ⎡ x k ⎤ ⎡ p k ⎤
⎢λ ⎥ = ⎢λ ⎥ + ⎢v ⎥ (3.34)
⎣ k +1 ⎦ ⎣ k ⎦ ⎣ k ⎦
Where; p k and v k are the search directions and obtained as the solution to the linear

system given by
⎡ pk ⎤
∇ 2 L( x k , λk ) ⎢ ⎥ = −∇L( x k , λ k ) (3.35)
⎣v k ⎦
In the absence of inequality constraints, the linear system takes the form as given below:

⎡∇ 2 L( x k , λ k ) - ∇g ( x k ) ⎤ ⎡ p k ⎤ ⎡− ∇L( x k , λ k ) ⎤
⎢ ⎥⎢ ⎥ = ⎢ (3.36)
⎢⎣ - ∇g ( x k ) 0 ⎥⎦ ⎣v k ⎦ ⎣ g ( x k ) ⎥⎦

This system of equations represents the first-order optimality conditions for the
optimization problem

1
2
[ ]
p T ∇ 2xx L( x k , λk ) p + p T [∇ x L( x k , λk )] (3.37)

subject to
[∇g ( xk )]T p + g ( xk ) = 0 i = 1,..., me (3.38)

with vk the vector of Lagrangian multipliers.

When inequality constraints are considered, (3.39) will be added into the problem.
[∇g i ( xk )]T p + g i ( xk ) ≥ 0 i = me + 1,...m (3.39)
This optimization problem is a quadratic program; i.e., it is the minimization of a
quadratic function subject to linear constraints. In SQP method at each iteration a

39
quadratic function is solved to obtain ( pk , vk ) . These are used to update ( xk , λk ) , and the
process repeats at the new point.

3.7.1 MATLAB Sequential Quadratic Programming

The SQP method described in sub-section 3.7 is the classical simple form.
However when solving optimization problem, SQP is not often used in this form. There
are two major reasons for this. First, it is not guaranteed to converge to a local solution to
the optimization problem. Second, it is expensive. The method described above requires
the Hessian of the objective function and the constraint function. Once these have been
obtained, a quadratic program must be solved to determine the updates search direction. It
would be desirable to reduce these requirements for derivatives, and to reduce the number
of arithmetic operations required at each iteration of the algorithm. If large problems are
being solved, it would also be desirable to reduce the storage requirements.
In present work SQP approach used in MATLAB has been adopted. Its implementation
consists of three stages:
1. Quadratic Programming (QP) problem solution
2. Line search & merit function
3. Updating of Hessian matrix of the Lagrangian function

a. Quadratic Programming (QP) Sub-Problem Solution

The QP program gives at each iteration a search direction pk +1 using procedure in
[71].
b. Line Search and Merit function
Convergence results for the SQP method are obtained by insisting that ( xk +1 , λk +1 )
be a “better” estimate of the solution than ( xk , λk ) . In the unconstrained case
progress is measured in terms of the objective function. In the constrained case it
is common to measure progress in terms of an auxiliary merit function. Usually, a
merit function is the sum of terms that include the objective function and the
amount of infeasibility of the constraints. If the new point reduces the objective
function and reduces infeasibility, then the value of the merit function will
decrease [73]. The merit function used here [71] is as follows:
me m
ψ ( x) = f ( x) + ∑ ri g i ( x) + ∑ r . max{0, g ( x)}
i i (3.40)
i =1 i = me +1

40
 The line search with respect to the merit function gives the new iterate as:
⎡ x k +1 ⎤ ⎡ x k ⎤ ⎡ pk ⎤
⎢ λ ⎥ = ⎢ λ ⎥ + α ⎢v ⎥ (3.50)
⎣ k +1 ⎦ ⎣ k ⎦ ⎣ k ⎦
where α is a step length and determined in order to produces a sufficient decrease
in the augmented Lagrangian function.
c. Quasi-Newton update
One way to reduce the expense of this algorithm is to use a quasi-Newton
approximation to the Hessian of the Lagrangian . If this is done, the second
derivatives need not be computed. At each iteration, H k is approximated by Bk ,
and calculated using BFGS quasi-Newton [71] method as given below:
q k q kT H kT H k
Bk +1 = Bk + − (3.51)
q kT s k q kT H k s k
Where;
s k = x k +1 − x k
n
⎛ n
⎞
q k = ∇f ( x k +1 ) + ∑ λi .∇g i ( x k +1 ) − ⎜ ∇f ( x k ) + ∑ λi .∇g i ( x k ) ⎟
i =1 ⎝ i =1 ⎠
λ is an estimate of the Lagrangian multiplier.

3.8 Hybrid GA
Hybrid methodology deals with synergistic integration of two approaches. The
objective of hybridization is to overcome the weakness of one approach during its
application with the strengths of other by appropriately integrating them. The integration
may be sequential, auxiliary, and embedded.

The GA excels at gravitating towards the global minimum. But especially it is not
fast at finding the minimum when in a locally quadratic region [74]. Thus hybrid
approach integrating the strengths of the GA with local optimizer can solve complex
convex and nonconvex problems effectively. Hybrid GA may take the following forms:
1. Running the GA until it slows down, then letting a local optimizer take over.
2. Running the GA as base line search for some specified number of generations or
till convergence to near optimal solution and then local optimizer run to find the
minimum.

41
3. After every some fixed number of iteration, running a local optimizer on the best
solution or best few solutions and adding the resulting chromosome to the
population.

42
 CHAPTER 4

Computational Framework for Economic

Dispatch Problem

4.1 Introduction
Generally the framework is platform which provides support for the development
of a specific domain of applications. This chapter presents the design and implementation
of framework targeting the economic dispatch solution.

4.2 Framework
Framework is a set of cooperating classes that makeup reusable design for a
specific class of software and defines overall structure of application, and it’s partitioning
in classes and objects. Architecture of a set of applications acts as basic conceptual model
for framework design.

Frameworks have basic intentions and purpose for reuse. There are two common
reuse strategies: white box reuse & black box reuse. White box reuse emphasizes the
reuse on source code thus necessitating the understanding on source code level for
developers, whereas in black box reuse, reusable components are in binary form and
application developer does not access to source code. This type of reuse is normally more
safe and easy to implement [75].

A framework can be customized to a particular application by creating application

specific subclasses of abstract classes targeting a specific group of applications, which
have similar architecture design [76]. A fairly large number of useful frameworks are
available for application builders. One of these frameworks is Microsoft’s Document-
View Architecture. It provides windows based menu driven application framework for
data handling through documents and data visualization through views. Microsoft

43
Foundation Classes (MFCTM) provides extensive functionality for such primitive tasks.
Standard Template Library (STL) is a toolkit, which provides fine implementation of
collections and common algorithms.

4.3 Computational Framework “PED_Frame” --- Need & Scope

Economic dispatch is vital step in electrical power system operation. It is being
solved by conventional optimization techniques and artificial intelligence tools based
approaches. Generally the algorithms for these techniques are implemented as stand alone
programs. The common software environment integrating the various algorithms is the
need for analysis in research and development activities. The Framework “PED_Frame”
[77] is the platform providing common environment for implementing the algorithms for
economic dispatch solution. This framework has the potential and ability to reuse the
existing code. The version one of the “PED_Frame” targets the implementation of
traditional, GA based and hybrid methodologies for economic dispatch used in the
present research work.

4.4 Features and Characteristics of Proposed Framework

In economic dispatch solution many common requirements are encountered in
most of the algorithms. PED_Frame provides basic structure to fulfill these requirements
and exposes the following functionalities in such a way that all algorithms can use them
easily:
1. Easy integration of new algorithm
2. Options for application of algorithms and their parameters
3. Ability to accept input data from standard format
4. Results in the form of formatted text files
5. Analysis progress visualization mechanism
6. Activity logging mechanism for view of intermediate results.
7. Handling of hybrid methodologies
8. Comparisons of results
9. Common services.

44
4.5 PED_Frame Design
The proposed framework “PED_Frame” targets power economic dispatch
software application development. It provides white-box reuse mechanism for
incorporating new algorithms. It provides application developer with capability of data
input in the form of machines cost curves and other algorithm specific information, which
is standard for all economic dispatch applications. PED_Frame also provides outputs in a
standard format. All inputs and outputs are handled through a grid. The class diagram of
PED_Frame is shown in Figure 4.1

Figure 4-1 Class diagram of PED_Frame

Analysis status visualization is useful during economic dispatch calculations,

therefore PED_Frame provides extremely efficient component for status visualization. As
most of the electrical engineering application are using vector and matrix algebra.
Standard Template Library (STL) is used as base for the development of the vector-
matrix library. This library gives efficient algorithm implementation for vectors and
matrices [78].

4.5.1 MFC Document -View Architecture

PED_Frame is primarily based upon doc-view architecture of Microsoft
Foundation Class. It acts as container for PED_Frame and provides base services. These

45
services are primarily in the form of messaging mechanisms, and threading support. The
base and derived classes are listed in table 4.1.
Table 4-1 Classes in PED_Frame
S. No. Class Base Class Remarks Description
This class manages overall
1 CPEDApp CWinApp
framework.
This class manages the main
2 CMainFrame CFrameWnd
window.
This class provides basic services
for visualization of application
3 CPEDView CScrollView
data. It also contains grid objects
and manages their functionality
This class works as data
repository and its management. It
also contains instances of matrix
class for input matrix, B-
coefficient matrix and output
4 CPEDDoc CDocument matrix. The reference of these
matrices is passed to respective
CPEDAnaslysisBase derived
class, which performs analysis by
using these matrices and updates
output matrix.
It acts as base class for different
5 CPEDAnaslysisBase CWinThread classes which implement
algorithms.
It manages the current status of
6 CPEDAnalysisStatus CDilog execution and visual components
like progress bar.
This class acts as base class for
7 CPEDGABase CPEDAnaslysisBase
GA based applications
This class acts as base class for
8 CPEDCABase CPEDAnaslysisBase
CA based applications

46
4.5.2 Input/Output Using Grid
CPEDGrid class acts as interface for grid, which provides functionality like input,
formatted output, print and data validation checks. It extends grid functionality to link its
services with other parts of the PED_Frame. These are:
a. Data present in Grid is transferred to a matrix, which can act as data
communicator.
b. Data present in matrix is formatted in grid for display and printing purposes.
c. Data editing and data validation checks.
PED_Frame creates three instances for grid class --- Machine data, B-coefficients and
output.

4.5.3 Analysis Status Visualization

Analysis status visualization is implemented in class CPEDAnalysisStatus derived
from CDialog class of MFCIM. This dialog window contains progress indicators showing
current state of convergence, iteration count, and time elapsed.

4.5.4 Vector-Matrix Library

This library includes template based CPEDVector and CPEDMatrix classes.
CPEDVector is derived from vector<class T> and it implements functionality for
initialization, resizing, addition, multiplication, searching, sorting and debugging routines.
CPEDMatrix class is derived from CPEDVector <CPEDVector <class T>> and it
implements functionality for initialization, resizing, and matrix algebra operations such as
addition, multiplications, division, transpose, inverse.

4.5.5 File Read Writes

The format for input and output has been standardized. The data files for all the
test systems have been prepared according to this standard and data is read through these
files. The results may be written with different options in output files.

47
4.6 Framework Implementation
Analysis is the process of computing ED solution using the techniques in the
PED_Frame. This part of the framework is responsible for actual implementation of
algorithm. This work is accomplished through following classes:
a. CPEDAnalysisBase class is derived from CWinThread so acting as independent
thread of execution. It contains references to matrices for inputs such as power
demand, machine cost curves co-efficient, B-coefficient, GA parameters,
encoding schemes, GA operators, maximum iterations count, error tolerance and
output. It also provides virtual functions for start analysis, communicating status
information and formatting output.
b. CPEDGABased is derived from CPEDAnalysisBase and works as base for
implementation of genetic algorithm. This class contains information for Genetic
Algorithm (GA) parameters. A new class can be derived from this class to
implement any GA based solution for Economic dispatch.
c. CPEDCABase is also derived from CPEDAnalysisBase and works as base for
conventional algorithm implementation. It contains information like initial start
and step length etc. A new class can also be derived from this class to implement
any mathematical programming based solution for Economic dispatch.

4.7 Steps for Implementation of Customized Solutions Using PED_Frame

PED_Frame is useful if we implement customized solution based on generic
services provided by it. We need to perform some specific task to use these services in
customized solution.

4.7.1 Economic Dispatch Using Mathematical Programming based Approach

It can be implemented as follows:
a. Derive a class from “CPEDCABase” and declare algorithm specific parameters in
it.
b. Write code for new algorithm in StartAnalyse function, which is in overridden
form.
c. Call UpdateStatus with error value and current iteration.

48
 d. Override “FormatResults” to transfer results from analysis structure to output
matrix.
e. Create a new menu item and write its message handler in CPEDDoc. Create an
object of Derived class and set parameters.
f. Create an object of CPEDAnalysisStatus and pass this object to it; calling its
DoModal( ) method . This will start analysis.
g. On completion of analysis just call “UpdateAllViews”, this will update grid
control with results.

4.7.2 Economic Dispatch Using GA

a. Derive a class from “CPEDGABase”.
b. Remaining steps are same as those in conventional approach.

4.7.3 Economic Dispatch Using Hybrid Approach

Sequential hybrid combinations have developed in such a way that control is
transferred from one algorithm to another after specified number of iterations or
convergence.

4.8 DOS Mode Operation

In the DOS mode all requisite data is read form text file and out put is written on
text files with following provisions:
• Output results with complete detail for every generation along with input data
• Output results of final generation with or without input.
• In case of multiple runs complete detail for every run with or without the
summary of the final results of all the runs.
• In case of hybrid methodology complete output results for each generation for
each algorithm or final results of each algorithm with or without input and
summary of final results of all algorithms. Similar options exits for multiple runs
of application.
Tracking of the time used in computation is also performed.

49
4.9 PED_Frame --- Execution Mode
The PED_Frame in Visual C environment appears as shown in Figures 4.2.

Figure 4-2 PED_Frame in visual C++ Environment

The File pull down menu as shown in Figure 4.3 is used for data read, save and print.

Figure 4-3 File Pull Down Menu in PED_Frame

50
The generators cost curves are loaded either from data file or may be entered through
keyboard and appears in the PED_Frame as shown in Figure 4.4

Figure 4-4 Cost curves of machines read from data file in PED_Frame Environment

The B-coefficients or loss coefficients are loaded from data file and shown in Figure 4.5.

Figure 4-5 B-coefficient read from data file in PED_Frame Environment

51
The Economic Dispatch menu gives the selection for different algorithms as shown in
Figure 4.6.

Figure 4-6 λ Iteration method in execution mode in PED_Frame

The parameters settings in PED_Frame for economic dispatch using λ Iteration method
are shown in Figure 4.7.

Figure 4-7 Parameter settings for λ Iteration method in PED_Frame

The output of ED studies is shown in Figure 4.8. The output includes the power
generation of the machines with their generation cost, total production cost, transmission
loss and algorithm convergence parameters.

52
 Figure 4-8 λ Iteration: Output results in PED_Frame

ED based on λ Iteration using GA has been selected as shown Figure 4.9 in PED_Frame
environment including transmission loss.

Figure 4-9 λ Iteration using GA in PED_Frame

53
The GA parameters for ED studies are set in PED_Frame as given Figure 4.10.

Figure 4-10 GA parameter settings in PED_Frame

The progress during execution is given by progress indicator as shown in Figure (4.11).

Figure 4-11 Progress indicator in PED_Frame

54
Output in PED_Frame appears as shown in Figure (4.12) and (4.13).

Figure 4-12 λ Iteration using GA: Output results in PED_Frame

Figure 4-13 Economic Dispatch using GA based on Real Power Search: Output
results in PED_Frame

55
 CHAPTER 5

Economic Dispatch Solution using Genetic

Algorithm

5.1 Introduction
This chapter presents the modeling aspects of genetic algorithm based economic
dispatch with a view to develop different evolution models to solve convex economic
dispatch problem and non-convex economic dispatch due to valve point effects. The
remaining discussions in this chapters focus on: algorithm development and their
implementations, case studies on convex ED, and finally the performance evaluation of
nonconvex economic dispatch problem.

5.2 Economic Dispatch Problem --- Mathematical Formulation

The economic dispatch problem can be broadly classified as convex and non-
convex economic dispatch.

5.2.1 Convex Economic Dispatch

In the convex economic dispatch input-output characteristics are assumed
monotonically increasing or piecewise linear. Mathematically in its simplest form with
one objective function and two constraints economic dispatch is formulated [79] as:
N N
Minimize CT = ∑Ci (Pi ) = ∑ai + bi Pi + ci Pi 2 (5.1)
i =1 i =1

Subject to Equality constraint:

N
PD + PL = ∑ Pi (5.2)
i =1

and inequality constraints:

Pi (min) ≤ Pi ≤ Pi (max) (5.3)

CT is the total cost

56
Ci is the fuel cost of the ith generating unit with ai, bi, and ci cost coefficients.
PD is the total load.
Pi(min) is the minimum active power generation.
Pi(max) is the maximum active power generation.
PL is the transmission loss.

5.2.2 Nonconvex Economic Dispatch

Nonconvex economic dispatch may results due to prohibited operating zones,
valve point effect, and multiple fuels. Here the nonconvex economic dispatch due valve
point will be addressed. Typically, the valve point results in the ripples when each steam
valve starts to open. The ED cost objective function taking into account the valve point
effects is generally modeled as superposition of sinusoidal function and second order
polynomial functions. This non-smooth curve mathematically [9] represented as:
C i ( Pi ) = a i + bi Pi + ci Pi 2 + ei sin( f i ( Pi (min) − Pi )) (5.4)

where ei , fi are the cost coefficients for ith generator reflecting valve point effects.

5.2.3 Transmission Loss in Economic Dispatch

Economic dispatch problem may be solved neglecting transmission loss. Present
day utility systems serve over a vast area of relatively low density. The transmission loss
may vary from 5 to 15 % of total load. Therefore it is essential to take into account
transmission loss in economic dispatch calculations [50].

Simple and approximate method of expressing transmission loss in terms of

generators power is through B-coefficients. The Kron’s formula [79] for loss calculation
is given by,
ng ng ng
PL = ∑ ∑ Pi Bij Pj + ∑ B 0i Pi + B00 (5.5)
i =1 j =1 i =1

Bij = constants called B coefficients or loss coefficients and form a square matrix B of
same dimensions as P.
B0i = form the vector B0 of same dimensions as P.
B00 = constant.

57
In the present work loss has been included using Kron’s formula. B-coefficients where
required have been calculated using the following equation [79]:
PL = PGT1ψ T C T Rbus C *ψ * PG*1

PL = PGT1ℜ[H ]PG*1 (5.6)

Where ℜ[H ] is the real part of the Hermitian matrix? The steps involved in evaluating
the B-coefficients are listed as follows:
Step 1: Load flow analysis for initial operating states
Step 2: Determination of individual load current ILD, total load current ID and ratio lk = ILD
/ ID
Step 3: Formation of Zbus
Step 4: B-coefficients from equation (5.6)

5.2.4 Economic Dispatch Using GA--- Search Methodology

Two search methodologies have been adopted namely as follows:
a. λ –Search: scans the search space genetically for incremental cost λ for
optimum generation schedule. Population of chromosomes representing λ is
generated.
b. Real Power Search: searches the power generation of the machines
genetically for optimum schedule. Each chromosome is formed by
concatenation of binary strings representing power generation of machines.
λ–Search has been used for convex economic dispatch, where as real Power
Search has been used for both convex and nonconvex ED problems. Both search
methodologies follow the GA working philosophy summarized in Figure 5.1.
Genetic Algorithm ( )
{
Population initialization;
Population evaluation;
while (Termination criterion satisfied)
{
Selection for next population;
Perform Crossover;
Perform Mutation;
Population evaluation;
}
}

Figure 5-1 Genetic Algorithm Working Steps

58
5.2.5 Objective Functions
The objective functions for Convex ED problem and nonconvex ED problem are
the cost curves given by equation (5.1) and equation (5.4) respectively.

5.3 Economic Dispatch Solution --- GA Based Modeling

Genetic algorithm is one of the most developed paradigm in evolutionary
computation. It basically implements the theory of evolution as an algorithm. The method
of selecting or picking parents and method of inserting offsprings back into population
taken together is called model of evolution. The following issues are important while
applying GA to practical problems:
1. Choosing basic implementation issues such as:
a. Representation
b. GA parameters selection
c. Selection
d. Crossover & mutation operators
2. Fitness function
3. Termination criteria

5.3.1 Initial Population Generation

GA may be binary coded or real coded depending upon the representation of
problem variable. In binary coded GA, problem variables are coded in the form {0,1}bits
whereas in real coded variables are represented by floating point number.
Deciding about the composition of population is very difficult. The aspects
pertaining to the population may be the population size, sampling, chromosome age and,
size variations [74]. Here the focus is on the initial population generation with fixed
population size. Three following schemes for binary coded GA have been adopted:
a. Random Encoding (RE)
Initial population is randomly generated. Some times this scheme results in
over-sampling of some regions and sparse sampling in others in search space
[74].

59
 b. Complementary Encoding (CE)
Half of the population is generated randomly and remaining half of the
population is the complement of the first half. This approach ensures diversity
by having every bit to assume both one and zero within the population. This
scheme helps the algorithm adequately to sample the objective function in
early stages and may reduce convergence time and prevent premature
convergence in a local minimum [74].
c. Gray Encoding (GE)
Ordinary binary number representation of variables may slow down the
convergence of a genetic algorithm. The two very close decoded values 128 &
129 have binary representation exactly opposite to each other and the
offsprings resulting from these parents with cross-site three are 159 & 56,
which are quite different. These values should be converging but actually it is
not so. Increasing the bits representation magnifies the problem. Gray
encoding of the variables gives the solution to the problem. It redefines the
binary numbers so that the consecutive numbers have hamming distance of
one. If the same numbers are redefined at hamming distance one with same
cross-site produces offsprings 128,127. The parents are the good solutions so
we may expect offsprings as good solution with Gray encoding [74].

5.3.2 Reproduction Operators

Reproduction operator is usually the first operator applied on the population. The
chromosomes are selected based on the Darwin’s evolution theory of survival of the
fittest. Best are selected for crossover. There are number of GA reproduction operators in
the literature but the main idea in them is that the above average strings are picked from
the current population and their multiple copies are inserted in the mating pool in a
probabilistic manner. The selection methods used in this work are: Roulette Wheel
Selection (RWS), Tournament Selection (TS), Rank Selection (RS), and Stochastic
Remainder Selection (SRS).

5.3.3 Crossover Operators

Crossover operator is applied to the mating pool with a hope that it will produce
good offsprings. The aim of this operator is to search the parameter space in such a way

60
that the information stored in the parents is preserved to maximum extent. Number of
crossover operators exists in the GA literature. Single point to five point crossover,
uniform crossover with & without mask has been investigated for economic dispatch
problem.

5.3.4 Inversion (INV)

Inversion is a phenomenon in which a portion of a gene or a group of genes which
have been cut out, turn through an angle of 180 and get reattached in reverse order.
Bridges first discovered inversion in 1923. In inversion the gene sequence in an inverted
segment is exactly the opposite of that in its normal homologous (non-inverted) segment.
In case of an inversion, there is no net loss or gain of genetic material and thus a
heterozygous condition is perfectly all right, but linkage relationships change [80].

This phenomenon of inversion has been mapped into economic dispatch problem
as an operator and applied with the probability of crossover. The chromosome is selected
from the population and undergoes inversion such a way that the string of bits is selected
from the two randomly generated sites in order in the chromosome and inverted as shown
below.
1 0 1 1 0 1 1 0 1
1 0 0 0 1 0 0 0 1

5.3.5 Deletion -Regeneration (DEL-REG)

Deletions are caused when a chromosome breaks and loses a segment. Deletions
lead to decrease in gene number. Bridges first discovered deletion in 1917. Chromosomes
with a deletion cannot be reverted back to wild type conditions [80]. This process has
been approximated as an operator to be applied with crossover probability. The bits
between two randomly generated cross-sites in order in the selected chromosome are
deleted and regenerated randomly as shown below.
1 0 1 1 0 1 1 0 1
1 0 1 0 1 1 0 0 1

61
5.3.6 Chromosome Length
It is kept fixed and formed by concatenating the string of fixed numbers of bits for
each variable. The number of binary digits needed to represent a continuous variation in
accuracy of ∆x for each variable is calculated [50] as:

l i ≥ log 2 ( xi max − xi min

∆x +1 ) (5.7)

5.3.7 Decoding
The decimal value of the chromosome is obtained by decoding the binary string
as:
l −1
y i = ∑ 2 j bij (5.8)
j =0

Where b is the binary bit and l is the length of chromosome.

Actual value of the power generation in search space is given by fixed mapping rule
max min
min Pi + Pi
Pi = Pi + yi (5.9)
2 −1
l

5.3.8 Fitness Function

In order to emphasize the best chromosome and speed up convergence of the
evolutionary process, fitness is normalized in the range between 0 and 1 [50]. Fitness
function for ED problem is given by
j ⎛ ε j ⎞⎟
f = 1 ÷ ⎜1 + k (5.10)
⎜ PD ⎟
⎝ ⎠

Where PD is the total load demand, k is scaling factor and ε is power mismatch
tolerance.

5.3.9 Termination Criteria

Termination criteria is either fixed number of generation or power mismatch
tolerance
N
ε i = PD + PL j − ∑ Pi j <= 0.01 (5.11)
i =1

62
5.4 Economic Dispatch Using GA --- (λ –Search)
The constrained economic dispatch (ED) problem is converted into unconstrained
problem using Lagrange multiplied method. The augmented function is given by [50]
⎛ ng
⎞
L( Pi , λ ) = F ( Pi ) + λ ⎜⎜ PD + PL − ∑ Pi ⎟⎟ (5.12)
⎝ i =1 ⎠
Here λ the incremental cost is the Lagrangian multiplier in equation (5.12). The
necessary conditions for this problem are:
∂L( Pi , λ ) ∂F ( Pi ) ⎛ ∂P ⎞
= + λ ⎜⎜ L − 1⎟⎟ = 0 (5.13)
∂Pi ∂Pi ⎝ ∂Pi ⎠

∂L( Pi , λ ) gn

= PD + PL − ∑ Pi = 0 (5.14)
∂λ i =1

The incremental transmission loss is obtained by differentiating equation (5.5) as

n
∂PL g

= ∑ 2 Bij Pj (5.15)
∂Pi j =1

Substituting the values of incremental loss, incremental cost in equation (5.13) gives
ng

2(ai + λBii ) Pi + λ ∑ 2 Bij Pj = λ − bi where i = 1, 2,...n g (5.16)

j =1
j ≠i

The system of linear equation given by equation (5.16) can be solved to obtain Pi
if λ is known.

In this method equal system incremental cost λ is obtained using GA and used as
coding parameter. The number of bits required to represent the variable λ may either be
taken arbitrarily a fixed number or be calculated using equation (5.7). The number of bits
will entirely be independent of number of generating units. This is particularly attractive
in large scale systems. The pseudo code for λ –Search algorithm is shown in Figure 5.2.

63
 Figure 5-2 Pseudo Code for λ –Search Methodology

5.5 Economic Dispatch Using GA --- Real Power –Search (RP-Search)

In this approach each generating unit loading is represented by a binary numbers.

System consisting of N generating units each loaded within its limits Pi ( max
, Pi
min
), the
number of bits to represent the value Pi of unit i is given by equation (5.7). The length of
chromosome is obtained by concatenating the bits representing each unit. The real power
search algorithm is the same as shown in Figure 5.2 except λ j
is replaced Pi. The
chromosome is decoded by using equation (5.8) and actual Pi is obtained using equation
(5.9).

5.6 GA Based Hybrid Approach for Convex Economic Dispatch

The philosophy of hybrid methodology is to exploit the best of the integrating
approaches by balancing the weakness of one with the strength of the other. GA based
sequential hybrid methodology [81] has been developed with the strategy, “GA Lambda

64
search after specified number of generation passes control to conventional lambda
iteration method, which then produces the final result”. This approach provides new
systematic method for initial selection of λ rather than guess work.

5.7 Computer Implementation

The algorithms have been implemented in the computational framework
PED_Frame and results have been taken on P-IV Personal Computer.

5.8 Case Studies --- Convex Economic Dispatch

Test Systems: The description and data of test systems used for
investigation in the case studies is given in Appendix A.
GA based ED: For GA based economic dispatch each test system runs 20
times for both λ-Search and Real Power-Search and results
corresponding to minimum generation cost has been
recorded and compared with the λ iteration method. The
transmission loss calculation has been done by using B-
coefficients.
Output Results: The detailed output results in tabular form are given in
Appendix C. The summary and comparison of results is
given in this chapter.

5.8.1 3-Machines Test Systems

The following three test systems have been investigated:
a. Test System I --- 3-Machine 6-Bus System with PD = 210 MW
b. Test System II --- 3-Machine 6-Bus System with PD = 850 MW
c. Test System III --- 3-Machine 5-Bus System with PD = 150 MW
The input data of these systems is given in Tables A.1, A.2 and A.3 placed in
Appendix A. The power mismatch tolerance is 0.0001 for λ iteration and 0.001 for λ-
Search and Real Power-Search methodologies.

65
Evolution Model: Evolution model with GA parameters for three test systems for both λ
–Search and RP-Search are given in Table C.1 placed in appendix C
Output Results: Tables C.2, C.3 and C.4 placed in appendix C lists the details of output
results of the test systems I, II, and III respectively. The table 5.1 presents the comparison
of results. The observations are as follows:
• In all three test systems the generation cost is almost the same for λ-Iteration and λ-
Search techniques.
• In RP-Search the generation cost is slightly higher than the other two techniques.
• The execution time is increasing in the order λ-Iteration, λ-Search techniques and RP-
Search.
• The standard deviation is very low in λ-search as compared to RP-Search.

Table 5-1 Comparison of the Results for 3-Machines Test Systems I, II, and III
GA Results After 20 Runs
Corresponding to Minimum
3 Machine Test Cost
λ-Iteration
System
λ-Search RP-Search

Test System I 3167.1620 3167.2395 3188.6666

Gen. Cost
Test System II 8515.1515 8514.037 8562.3129
($/h)
Test System III 1599.9834 1599.9887 1603.5225
Test System I 0.015 0.7790 4.0620
Execution
Test System II 0.015 0.1250 10.5810
Time(sec)
Test System III 0.015 0.0620 0.0470
Test System I --- 0.0059 0.4370
Standard
Test System II --- 0.0046 143.2415
Deviation
Test System III ---- 0.0029 0.5715
% Difference Test System I --- -0.0024 -0.6790
in Gen. Cost Test System II --- 0.0131 -0.5539
from λ Test System III
----
Iteration -0.0003 -0.2212

5.8.2 IEEE Test Systems

Following two IEEE test systems have been selected for analysis:
a. 5-Machines 14-Bus System with PD = 259 MW
b. 6-Machines 30-Bus System PD = 283.4 MW
Two cases for each of the system based on different cost curves have been tested.
The input data of these systems for case I & case II is given in Tables A.4, A.5 and A.6

66
placed in appendix A. The power mismatch tolerance is 0.0001 for λ iterations and 0.001
for λ-Search and Real Power-Search methodologies.
Evolution Model: Evolution model with GA parameters for these test systems for both λ
–Search and RP-Search are given in Table C.5 and C.6 respectively placed in appendix C.
Output Results: Tables C.7, C.8, C.9 and C.10 placed in appendix C lists the details of
output results of the 5-machine and 6-machine system for both the cases respectively. The
Table 5.2 and 5.3 present the comparison of the results. The observations are as follows:
• The generation cost is almost the same for λ-Iteration and λ-Search techniques for
both test systems for cases I & II.
• In RP-Search the generation cost is higher than the other two techniques.
• The execution time is increasing in the order λ-Iteration, λ-Search techniques and
RP-Search except for case II: 5-Machine system.
• The standard deviation is very low in λ-search as compared to RP-Search.

Table 5-2 Comparison of the Results for IEEE 5-Machines Test System
GA Results After 20 Runs
IEEE 5 Corresponding to Minimum
Machines 14 Cost
λ-Iteration
Bus Test
System λ-Search RP-Search

Gen. Cost Case – I 833.64750 833.5897 887.9356

($/h) Case – II 720.4242 720.5030 740.7419
Execution Case – I 0.015 1.6880 2.1250
Time Case – II 0.015 4.2970 0.1870
Standard Case – I --- 0.0017 21.0689
Deviation Case – II --- 0.0020 91.5675
% Difference Case – I --- 0.0069 -6.5121
in Gen. Cost Case – II
from λ --- -0.0109 -2.8202
Iteration

Table 5-3 Comparison of the Results for IEEE 6-Machines Test System
GA Results After 20 Runs
IEEE 6 Corresponding to Minimum
Machines 30 Cost
λ-Iteration
Bus Test
System λ-Search RP-Search

Gen Cost Case - I 914.59347 914.7356 963.3991

($/h) Case - II 802.36290 802.7253 872.8807

67
 Execution Case - I 0.015 0.5780 1.6870
Time Case - II 0.015 3.8880 4.2500
Standard Case - I --- 0.0016 25.2307
Deviation Case - II --- 0.0017 53.1444
% Difference Case - I --- -0.0155 -5.3363
in Gen . Cost
from λ Case - II --- -0.0452 -8.7888
Iteration

5.8.2.1 6-Machines 26-Bus System with PD = 1263 MW

The generators data and loss coefficients are given in Table A.7 and Table A.8
respectively placed in appendix A.
Evolution Model: Evolution model with GA parameters for both λ –Search and RP-
Search are given in Table C.11 placed in appendix C.
Output Results: Tables C.12 placed in appendix C lists the details of output results of the
test systems. The Table 5.4 presents the comparison of results. The observations are as
follows:
• The generation cost is almost the same for λ-Iteration and λ- Search techniques.
• In RP-Search the generation cost is slightly higher than the other two techniques.
• The execution time is increasing in the order RP-Search, λ-Iteration, and λ-Search
techniques.
• The standard deviation is very low in λ-Search as compared to RP-Search.

Table 5-4 Comparison of the Results for 6-Machines 26 Bus System

GA Results After 20 Runs Corresponding

6-Machines 26 Bus to Minimum Cost
λ-Iteration
System
λ-Search RP-Search

Gen Cost
15689.78647 15689.9361 15708.7556
($/h)
Execution Time 1 0.6250 0.4840
Standard 66.0514
--- 0.0090
Deviation
% Difference in
Gen. Cost from λ --- -0.0010 -0.1209
Iteration

68
5.8.3 GA λ-Search Methodology with Mismatch Tolerance 0.0001

The GA λ-Search methodology with mismatch tolerance of 0.0001 was also

tested, but in that case GA takes longer time to converge.

5.8.4 Summary and Conclusion for the Case Studies

From the observations recorded in all test cases the trend established is
summarized as:
1. The generation cost for λ-Iteration and λ-Search techniques are very close to
each other despite the different mismatch tolerances (i.e., 0.0001 for λ-
Iteration and 0.001 for λ-Search).
2. The generation cost in RP-Search is higher than other two techniques. For
standard test systems the execution time in λ-Search technique is higher than λ
Iteration method, and lesser than RP-Search. However, in some cases RP-
Search time less than λ-Search technique.

These case studies shows that in convex economic dispatch RP-Search gives near
optimal results and for λ-Search results are close to the λ-iteration and with higher
mismatch tolerance it will take longer time to converge.

5.9 Case Studies --- Convex Economic Dispatch Using Hybrid Approach
Following two test system have been selected for investigation:
1. Test System I --- 3-Machine 6-Bus System with PD = 210 MW
2. 20-Machines system with PD=2500 MW

The input data of Test System-I is given in Table A.1 placed in Appendix A. The
cost curves and B-coefficients for the second system are given in Tables A.9 and A.10
respectively placed in appendix A.
The working strategy of this approach is as follows:
Step 1 Run λ-Search for Specified Number of Generation (SNG).
Step 2 Use the answer from step 1 as a starting point and solve ED dispatch
problem using λ-Iteration.
Step 3 Get the final results and quit program.

69
 Step 4 Comparison of results of hybrid approach with λ-Search and λ-Iteration.
The Evolution model and GA parameters are given in Table C.13 placed in appendix C.
The output results of 3-Machines and 20 Machines systems have been tabulated in Tables
C.14 and C.15 respectively placed in appendix C. The comparison of the results is given
in Table 5.5. The results show that the total generation cost for λ-Iteration and hybrid
approach remains the same. However, the number of iterations is reduced in hybrid
approach.
Table 5-5 Comparison of the Results: Convex ED using Hybrid Approach
3-Machine System
λ-Iteration GA λ-Search Hybrid
Gen Cost ($/h) 3167.16442 3155.4267 3167.164
PL 8.8362 8.0073 8.8362
Tol 0.0001 0.001 0.0001
Iter. or Gen. 08 6496 06
λ 13.0056 12.9544 13.0056
20-Machine System
λ-Iteration GA λ-Search Hybrid
Gen.Cost ($/h) 62453.3662 62453.7890 62453.3660
PL 91.7170 91.7182 91.7170
Tolerance 0.0001 0.1 0.0001
Iter. or Gen. 16 909 13
λ 20.9539 20.9540 20.9539

5.10 Nonconvex Economic Dispatch

Thirteen generators system with valve point effect and demand of 1800 MW has
been selected. The cost curves are given in Table A.11 placed in appendix A.
The performance evaluation of GA based economic dispatch solution based on
empirical results for different evolution model with a view to establish the trend for the
parameters of interests has been carried out. The working strategy adopted is as follows:

• GA Parameters fixed.
• Max. No. of runs fixed.
• Set of GA Evolution models is formed in such a way that each encoding scheme
with each selection method has been tested for eight different crossover operators.
• A termination criterion is power balance equation (5.2).
• Program runs twenty times for each case and results have been tabulated in Tables
5.6 and 5.7. Column 1, 2 & 3 represents the minimum generation cost, minimum

70
 execution time, and number of convergences in 20 trials respectively. Table 5.8
lists the minimum number of generation for convergence after 20 trials.

The parameters of interest include: generation cost, execution time, number of

generation for convergence, number of convergence in fixed no of runs, and mismatch
tolerance. The strategy for comparison includes intra-comparison and inter-comparison.
In intra-comparison parameter of interest has been compared for three encoding scheme
with each of four selection method against crossover operator, whereas in case of inter-
comparison four selection method with each of the three encoding schemes against
crossover operators.
Table 5-6 Generation Cost, Minimum Execution Time & No. of Convergences for
Roulette Wheel Selection and Tournament Selection against Crossover Operators
with Three Encoding Schemes
Encoding Roulette Wheel Selection Tournament Selection
Schemes Min. No. Min. No.
S.No Operators Generation Generation
Exe. of Exe. of
Cost Cost
Time Conv. Time Conv.
RE 19123.697 0.071 20 19181.484 0.030 20
One Point
1 GE 18993.967 0.060 20 19113.296 0.040 20
Crossover
CE 19006.306 0.080 20 18644.243 0.050 20
RE 18860.694 0.110 20 19187.549 0.010 20
Two Point
2 GE 18862.772 0.331 20 19087.879 0.010 20
Crossover
CE 18941.772 0.060 20 19022.125 0.020 20
RE 18972.097 0.080 20 19056.880 0.030 20
Three Point
3 GE 18951.427 0.110 20 19048.249 0.070 20
Crossover
CE 19045.969 0.111 20 19258.512 0.010 20
RE 18968.956 0.180 20 19139.809 0.060 20
Five Point
4 GE 19020.457 0.121 20 19048.249 0.070 20
Crossover
CE 18939.271 0.001 20 19068.460 0.090 20
Uniform RE 19032.748 0.210 20 18934.987 0.100 20
Crossover GE 18875.262 0.010 20 18993.428 0.060 20
5
without
CE 19029.501 0.010 20 19016.410 0.060 20
mask
Uniform RE 18999.313 0.070 20 19024.916 0.040 20
6 Crossover GE 18920.957 0.120 20 19046.933 0.050 20
with mask CE 19005.506 0.130 20 19106.402 0.060 20
RE 19008.231 0.110 20 19126.267 0.130 20
7 Inversion GE 19012.665 0.070 20 19060.114 0.060 20
CE 19119.527 0.140 20 18959.989 0.191 20
RE 19012.910 0.290 20 18901.961 0.460 20
Deletion &
8 GE 18961.688 0.581 20 18939.581 0.051 20
Regeneration
CE 18996.082 0.180 20 19073.803 0.060 20

71
 Table 5-7 Generation Cost, Minimum Execution Time & No. of Convergences for
Rank Selection and Stochastic Remainder Selection against Crossover Operators
with Three Encoding Schemes
Encoding Rank Selection Stochastic Remainder
Schemes Selection
S.No Operators Min. No. Min. No.
Generation Generation
Exe. of Exe. of
Cost Cost
Time Conv. Time Conv.
RE 19054.205 0.050 17 19185.209 0.011 20
One Point
1 GE 19056.388 0.040 20 18933.225 0.060 20
Crossover
CE 18982.232 0.030 20 18998.593 0.030 20
RE 18903.387 0.000 20 19131.330 0.080 20
Two Point
2 GE 19064.576 0.050 20 19046.593 0.040 20
Crossover
CE 18997.681 0.040 20 18860.589 0.020 20
RE 18934.945 0.001 20 18660.011 0.060 20
Three Point
3 GE 18990.309 0.060 20 19172.235 0.171 20
Crossover
CE 19219.935 0.040 20 19142.567 0.090 20
RE 19053.374 0.060 20 18804.900 0.130 20
Five Point
4 GE 18904.763 0.070 20 19041.905 0.080 20
Crossover
CE 19084.054 0.030 20 19020.222 0.060 20
Uniform RE 19078.522 0.030 20 18979.744 0.060 20
Crossover GE 19044.779 0.030 20 19013.387 0.030 20
5
without
CE 18943.724 0.080 20 19087.088 0.071 20
mask
Uniform RE 19572.497 0.240 20 18850.258 0.001 20
6 Crossover GE 18852.632 0.060 20 19133.045 0.070 20
with mask CE 19041.642 0.050 20 18857.418 0.100 20
RE 19035.649 0.120 20 18987.803 0.110 20
7 Inversion GE 19008.501 0.090 20 18948.923 0.181 20
CE 18966.386 0.130 20 19100.301 0.230 20
RE 19089.825 0.030 20 18915.405 0.100 20
Deletion &
8 GE 18980.331 0.040 20 18917.715 0.151 20
Regeneration
CE 19128.201 0.080 20 18908.902 0.010 20

Table 5-8 Minimum No. of Generations for Convergence After 20 Trials

Roulette Tournament Rank
Stochastic
Wheel Selection Selection
Remainder
Selection Selection
S.No Operators Encoding Encoding Encoding Encoding
Schemes Schemes Schemes Schemes
RE GE CE RE GE CE RE GE CE RE GE CE
One Point
1 29 23 38 12 19 19 17 16 14 03 31 12
Crossover
Two Point
2 41 122 26 07 10 08 03 20 17 49 17 6
Crossover
Three Point
3 30 37 47 23 33 04 03 22 14 29 69 28
Crossover
4 Five Point 72 56 01 31 33 38 27 23 15 50 27 22

72
 Crossover
Uniform
5 Crossover 83 03 03 39 23 24 20 15 29 25 12 26
without mask
Uniform
6 Crossover 57 44 44 10 23 23 11 24 16 03 24 31
with mask
7 Inversion 107 28 72 71 28 75 41 30 49 35 55 86
Deletion &
8 107 246 77 181 33 29 15 10 41 44 63 09
Regeneration

5.10.1 Intra Comparison --- Performance Indicators Variations (PIV)

Minimum generation cost and minimum execution time profiles after twenty trials for one
selection method with three encoding scheme against crossover operators have been
plotted and shown in figures.
a. PIV for Roulette Wheel Selection with Three Encoding Schemes ~
Crossover Operator.
The cost profile Figure 5.3(a) indicates stable result with comparable and
slight difference with hundred percent convergences in case of all operators.
Lowest & highest minimum cost occurs with Gray and random encoding
respectively. Two encodings have comparable costs among themselves but in
slight difference with third encoding scheme in all operators. There is
similarity in the pattern of variation in Gray & random encodings and
generation cost with Grey encoding remains low amongst all the three
encodings except five point crossover.

The execution time profile Figure 5.3(b) is with peaks, dips and few
saddles. Lowest & highest time occurs with complementary & Gray encodings
respectively. All the three encodings have comparable times in operators: one
point, three point, uniform crossover and marked difference amongst
themselves in rest of the operators.
b. PIV for Tournament Selection with Three Encoding Schemes ~ Crossover
Operator
The cost profiles Figure 5.4(a) has moderate variation in cost except
marked difference in operator one point & three point crossover with hundred
percent convergences in all operators. Lowest & highest minimum cost occurs

73
 in complementary encoding with operators one point & three point
respectively. The Gray encoding has smooth profile as compared to the other
encodings.
The execution time profiles Figure 5.4(b) of all the three encodings have
comparable times except inversion, deletion & regeneration operators. Lowest
& highest time occurs with Gray & random encodings respectively.
c. PIV for Rank Selection with Three Encoding Schemes ~ Crossover
Operator
Dispersed variations in the cost profile as shown in Figure 5.5(a). Lowest
& highest minimum cost occurs with random & complementary encoding
respectively. The variation in gray encoding is less than other encodings.
Slight variations exist in time Figure 5.5(b) for all the operators. Lowest &
highest minimum time occurs with random and complementary encodings
respectively.
d. PIV for Stochastic Remainder Selection with Three Encoding Schemes ~
Crossover Operator.
Variations in cost profiles Figure 5.6(a) with peaks and dips along with
few comparables and hundred percent convergences. Lowest & highest
minimum cost is with random encodings. Dispersed variation in time profiles
Figure 5.6(b) with lowest & highest minimum time in complementary
encoding.

5.10.2 Inter Comparison --- Performance Indicators Variations (PIV)

Minimum generation cost & minimum execution time profiles after twenty trials
for each of the three encoding scheme with four selection methods against crossover
operators have been plotted and shown in figures.
a. PIV for Random Encoding with Four Selection Method ~ Crossover
Operator
Figure 5.7(a) shows cost variation profiles. For roulette wheel and rank
selection the cost varies closely to each other except at uniform crossover with
mask and there is increasing trend except dips at three point crossover and
uniform crossover with mask operator. The variations in case of tournament
selection are well behaved and remain high as compared to the other selection

74
 method except at uniform crossover without mask. The minimum cost occurs
with stochastic remainder selection with non-smooth variations.

Figure 5.7(b) shows the time variation profiles. The execution time for
roulette wheel selection remains high as compared to other three selection
methods. Whereas for rank & tournament selection time varies closely to each
other except at deletion & regeneration operator. In stochastic remainder
selection time variations follows the variation pattern of roulette wheel
selection and remains low from it. Marked difference in time appears at
deletion & regeneration amongst all selection methods.
b. PIV for Gray Encoding with Four Selection Method ~ Crossover
Operator.
Figure 5.8(a) shows cost variation profiles. Dispersed cost variation
profiles for all the selection methods such that there are peaks & dips except
tournament selection where the variations are smooth. At operator deletion &
regeneration the cost are close to each as compared to the others where there is
marked dispersion. Lowest & highest minimum costs occurs with rank
selection & stochastic remainder selection respectively.

Figure 5.8(b) shows the time variation profiles. There are slight variations
in time profiles except marked variations in case of roulette wheel at two point
crossover and deletion & regeneration operators. The time is very close to
each other for one point crossover and uniform crossover without mask for all
selection methods.
c. PIV for Complementary Encoding with Four Selection Method ~
Crossover Operator
Figure 5.9(a) shows cost variation profiles. There is variation in the cost
profiles with similarities in trend of variations. Tournament selection & rank
selection form one pair and roulette wheel & stochastic remainder selection
form second pair. Lowest & highest cost occurs in tournament selection.

Figure 5.9(b) shows the time variation profiles. There are similarities in
trend of variations of execution time profiles. For inversion operator time is on

75
 higher side for all selection methods. Lowest & highest minimum time is with
roulette wheel & stochastic remainder selection respectively.

Roulette Wheel Selection

RE GE CE
19300

19100
Gen. Cost

18900

18700
1 2 3 4 5 6 7 8
Crossover Operators

(a)

Roulette Wheel Selection

RE GE CE
0.7
Min. Exe. Time

0.3

-0.1
1 2 3 4 5 6 7 8
Crossover Operators

(b)

Figure 5-3 Minimum Generation Cost and Minimum Execution Time against
Crossover Operators for Roulette Wheel Selection

76
 Tournament Selection
RE GE CE

19300

19100
Gen. Cost

18900

18700

18500
1 2 3 4 5 6 7 8
Crossover Operators

(a)

Tournament Selection

RE GE CE

0.7
Min. Exe. Time

0.3

-0.1
1 2 3 4 5 6 7 8
Crossover Operators

(b)

Figure 5-4 Minimum Generation Cost and Minimum Execution Time against
Crossover Operators for Tournament Selection

77
 Rank Selection
RW GE CE
19300

19100
Gen. Cost

18900

18700
1 2 3 4 5 6 7 8
Crossover Operators

(a)

Rank Selection
RW GE CE

0.15

0.10
Min. Exe. Time

0.05

0.00

-0.05

-0.10
1 2 3 4 5 6 7 8
Crossover Operators

(b)

Figure 5-5 Minimum Generation Cost and Minimum Execution Time against
Crossover Operators for Rank Selection

78
 Stochastic Remainder Selection

RW GE CE

19400

19100
Gen. Cost

18800

18500
1 2 3 4 5 6 7 8
Crossover Operators

(a)

Stochastic Remainder Selection

RW GE CE
0.3
Min Exe. Time

0.1

-0.1
1 2 3 4 5 6 7 8
Crossover Operators

(b)

Figure 5-6 Minimum Generation Cost and Minimum Execution Time against
Crossover Operators for Stochastic Remainder Selection

79
 Random Encoding
RW TS RS SRS
19300

19200

19100
Gen. Cost

19000

18900

18800

18700

18600
1 2 3 4 5 6 7 8
Crossover Operators

(a)

Random Encoding
RW TS
0.50 RS SRS

0.45
0.40
0.35
Execution Time

0.30
0.25
0.20
0.15
0.10
0.05
0.00
1 2 3 4 5 6 7 8
Crossover Operators

(b)

Figure 5-7 Minimum Generation Cost and Minimum Execution Time against
Crossover Operators for Random Encoding Scheme

80
 Gray Encoding

RW TS RS SRS
19200

19150

19100

19050
Gen. Cost

19000

18950

18900

18850

18800
1 2 3 4 5 6 7 8
Crossover Operators

(a)

Gray Encoding

RW TS RS SRS
0.7

0.6
Execution Time

0.5

0.4

0.3

0.2

0.1

0
1 2 3 4 5 6 7 8
Crossover Operators

(b)

Figure 5-8 Minimum Generation Cost and Minimum Execution Time against
Crossover Operators for Gray Encoding Scheme

81
 Complementary Encoding

RW TS RS SRS
19300

19200

19100

19000
Gen. Cost

18900

18800

18700

18600

18500
1 2 3 4 5 6 7 8
Crossover Operators

(a)

Complementary Encoding

0.25 RW TS RS SRS

0.20
Execution Time

0.15

0.10

0.05

0.00
1 2 3 4 5 6 7 8
Crossover Operators

(b)

Figure 5-9 Minimum Generation Cost and Minimum Execution Time against
Crossover Operators for Complementary Encoding Scheme

82
5.10.3 Convergence Profiles
The termination criterion is equation 5.2 with tolerance of 0.01. The convergence
profiles have been drawn between power mismatch against number of generations for
each operator with one selection method and different encoding schemes. For example for
1-Point operator with Roulette wheel selection (RW) there are three profiles for three
encoding schemes namely RWRE1PT, RWGE1PT & RWCE1PT. The tolerance profiles
for four operators (one point, uniform crossover without mask, inversion, deletion &
regeneration) have been drawn during one trial with fixed GA parameters.
a. Roulette Wheel Selection
For roulette wheel selection with one point and uniform crossover
operators, variations in the convergence profiles as shown in Figure 5.10(a-c)
and Figure 5.11(a-c) prevail right from start till to the convergence. Spike like
variations pattern in the convergence profile for Inversion, Deletion &
Regeneration operators as shown in Figure 5.12(a-c) and Figure 5.13(a-c).
b. Tournament Selection
There are variations during the initial generations and then smooth
convergence to the solution for one point crossover and uniform crossover
without mask as shown in Figure 5.14(a) and Figure 5.14(b) respectively.
Spike like variations for Inversion operator with random encoding and
Gray encoding, whereas for complementary encoding convergence profile is
with peaks and flats and fastest amongst the three encodings. Figure 5.15(a-c)
highlights the picture.
In Deletion & Regeneration operator spike like variations with random
encoding whereas for gray and complementary encoding relaxed variation
with peaks, flats and saddles. Figure 5.16(a-c) gives the picture. The
convergence is fastest with gray encoding amongst the three encodings.
c. Rank Selection
There are variations during the initial generations and then smooth
convergence to the solution for one point crossover and uniform crossover
without mask as shown in Figure 5.17(a) and Figure 5.17(b) respectively.
Spike like variations in all three encodings as shown in Figure 5.18(a-c)
for Inversion operator.

83
 In Deletion & Regeneration operator spike like variations with random
encoding whereas for gray and complementary encoding relaxed variation
with peaks, flats and saddles. Figure 5.19(a-c) gives the picture. The
convergence is fastest with gray encoding amongst the three encodings.
d. Stochastic Remainder Selection
For one point and uniform crossover without mask variations exist in
almost three quarter portion of the convergence profile and then stable
convergence to solution as shown in Figure 5.20(a) and Figure 5.20(b)
respectively.
Spike like variations in Inversion, Deletion & Regeneration operators for
all three encodings except random encoding with Inversion where
convergence profile is relaxed with peaks and few flats. Figure 5.21(a-c) and
Figure 5.22(a-c) gives the picture.

5.10.4 Discussion
In the light of empirical analysis the important observations may be listed as:
1. In random encoding scheme the tournament and rank selection methods gives
consistent performance as compared to the other methods and both are fast in
converging to the solution. The costs are comparable at one point crossover,
uniform crossover, inversion and deletion & regeneration operators.
2. In Gray encoding the tournament selection maintains the stable pattern of
generation cost at all operators as compared to the other three selection
methods which have diverse pattern of cost values except deletion &
regeneration operator. The tournament selection method is fastest amongst all
the selection methods for all operators.
3. In complementary encoding all selection methods have comparable generation
costs for all operators.
4. The tournament & rank selection methods are faster as compared to the other
two methods in terms to the convergence to the solution.
5. Three types of convergence profiles exists:
i) Variations initially and then smooth transitions towards convergence.
ii) Relaxed variations with peaks, flats and saddles.
iii) Spikes like variations.

84
 Type i) is mostly with one point and uniform crossover operators and type ii)
& iii) with Inversion, and Deletion-Regeneration operators.
The generation cost obtained from different evolution models for this case study
give the near optimal result when compared with other technique such as PSO, thus
requires fine tuning to reach to the optimal solution.

The 3-Machine system named as Test System 2 in appendix A was investigated

with different evolution models by tuning the GA parameters for the total generation cost.
The solution hits close to the lowest cost 8234.4 obtained in the literature with evolution
model using Inversion operator, and requiring fine tuning. Thus in nonconvex search
space GA gives near optimal results. The results can be fine tuned to optimal value by
adopting appropriate methodology. In chapter seven three hybrid models have been
proposed for nonconvex economic dispatch problem.

85
 Roulette Wheel Random Encoding One
Point Crossover

190

140

Mismatch
90

40

-10
1 101 201 301 401 501 601 701
Generations

(a)
Roulette Wheel Gray Encoding One Point
Crossover

140
Mismatch

90

40

-10
1 101 201 301 401 501
Generations
(b)
Roulette Wheel Complementary Encoding
One Point Crossover

240
190
Mismatch

140
90
40
-10
1 751 1501 2251 3001
Generations
(c)

Figure 5-10 Power Mismatch against No. of Generations with Roulette Wheel
Selection, Three Encoding Schemes and One Point Crossover

86
 Roulette Wheel Random Encoding Uniform
Crossover without Mask

390

290

Mismatch
190

90

-10
1 501 1001 1501 2001 2501 3001 3501
Generations

(a)
Roulette Wheel Gray Encoding Uniform
Crossover without Mask

190

140
Mismatch

90

40

-10
1 101 201 301 401 501 601
Generations

(b)
Roulette Wheel Complementary Encoding
Uniform Crossover without Mask

490
390
Mismatch

290
190
90
-10
1 201 401 601 801 1001
Generations

(c)

Figure 5-11 Power Mismatch against No. of Generations with Roulette Wheel
Selection, Three Encoding Schemes and Uniform Crossover

87
 Roulette Wheel Random Encoding
Inversion

45
35

Mismatch
25
15
5
-5
1 501 1001 1501 2001 2501 3001
Generations

(a)
Roulette Wheel Gray Encoding Inversion

60
50
40
Mismatch

30
20
10
0
-10
1 101 201 301 401
Generations

(b)
Roulette Wheel Complementary Encoding
Inversion

35

25
Mismatch

15

-5
1 101 201 301
Generations

(c)

Figure 5-12 Power Mismatch against No. of Generations with Roulette Wheel
Selection, Three Encoding Schemes and Inversion

88
 Roulette Wheel Random Encoding Deletion
& Regeneration

150

110

Mismatch
70

30

-10
1 251 501 751 1001
Generations

(a)
Roulette Wheel Gray Encoding Deletion &
Regeneration

150

110
Mismatch

70

30

-10
1 201 401 601
Generations

(b)
Roulette Wheel Complementary Encoding
Deletion & Regeneration

110

80
Mismatch

50
20

-10
1 201 401 601
Generations

(c)

Figure 5-13 Power Mismatch against No. of Generations with Roulette Wheel
Selection, Three Encoding Schemes and Deletion & Regeneration

89
 Tournament Selection One Point Crossover
35
RE GE CE

25
Mismatch

15

-5
1 11 21 31 41
Generations

(a)

Tournament Selection Uniform Crossover without Mask

45
RE GE CE

35

25
Mismatch

15

-5
1 11 21 31 41
Generations
(b)

Figure 5-14 Power Mismatch against No. of Generations with Tournament Selection,
Three Encoding Schemes, One Point & Uniform Crossover

90
 Tournament Selection Random Encoding
Inversion

55

Mismatch
35

15

-5
1 201 401 601 801
Generations

(a)
Tournament Selection Gray Encoding
Inversion

35
Mismatch

25
15
5
-5
1 101 201 301 401 501 601
Generations

(b)
Tournament Selection Complementary
Encoding Inversion

60
Mismatch

40

20

0
1 21 41 61 81
Generations

(c)

Figure 5-15 Power Mismatch against No. of Generations with Tournament Selection,
Three Encoding Schemes and Inversion

91
 Tournament Selection Random Encoding
Deletion & Regenration

30

Mismatch
20
10
0
-10
1 101 201 301
Generations

(a)
Tournament Selection Gray Encoding
Deletion & Regenration

55
Mismatch

35

15

-5
1 11 21 31 41
Generations

(b)
Tournament Selection Complementary
Encoding Deletion & Regenration

40
Mismatch

30
20
10
0
1 11 21 31 41 51 61 71 81 91
Generations

(c)

Figure 5-16 Power Mismatch against No. of Generations with Tournament Selection,
Three Encoding Schemes and Deletion & Regeneration

92
 Rank selection One Point Crossover

RE GE CE
35

25
Mismatch

15

-5
1 11 21 31 41 51 61
Generations

(a)
Rank Selection Uniform Crossover without Mask

RE GE CE
35

25
Mismatch

15

-5
1 11 21 31 41 51 61 71
Generations

(b)

Figure 5-17 Power Mismatch against No. of Generations with Rank Selection, Three
Encoding Schemes, One Point & Uniform Crossover

93
 Rank Selection Random Encoding
Inversion
90
70

Mismatch
50
30
10
-10
1 501 1001 1501 2001 2501 3001
Generations

(a)
Rank Selection Gray Encoding Inversion

70

50
Mismatch

30

10

-10
1 51 101 151 201 251 301 351 401 451
Generations

(b)
Rank Selection Complementary Encoding
Inversion

40
30
Mismatch

20
10
0
-10
1 51 101 151 201 251 301 351
Generations

(c)

Figure 5-18 Power Mismatch against No. of Generations with Tournament Selection,
Three Encoding Schemes and Inversion

94
 Rank Selection Random Encoding Deletion
& Regeneration

40
30

Mismatch
20
10
0
-10
1 51 101 151 201 251 301 351
Generations

(a)
Rank Selection Gray Encoding Deletion &
Regeneration

50
40
Mismatch

30
20
10
0
-10
1 6 11 16 21 26 31 36 41
Generations

(b)
Rank Selection Complementary Encoding
Deletion & Regeneration

20
15
Mismatch

10
5
0
-5
-10
1 11 21 31 41 51 61
Generations

(c)

Figure 5-19 Power Mismatch against No. of Generations with Tournament Selection,
Three Encoding Schemes and Deletion & Regeneration

95
 Stochastic Remainder Selection One Point Crossover

70 RE GE CE

50
Mismatch

30

10

-10
1 11 21 31 41 51 61 71
Generations

(a)

Stochastic Remainder Selection Uniform Crossover

Without Mask

RE GE CE
70

50
Mismatch

30

10

-10
1 11 21 31 41 51 61 71 81
Generations

(b)

Figure 5-20 Power Mismatch against No. of Generations with Stochastic Remainder
Selection, Three Encoding Schemes, One Point & Uniform Crossover

96
 Stochastic Remainder Selection Random
Encoding Inversion

50
40

Mismatch
30
20
10
0
-10
1 11 21 31 41
Generations

(a)
Stochastic Remainder Selection Gray
Encoding Inversion

40
30
Mismatch

20
10
0
-10
1 301 601 901 1201 1501 1801 2101
Generations

(b)
Stochastic Remainder Selection
Complementary Encoding Inversion

40
30
Mismatch

20
10
0
-10
1 201 401 601
Generations

(c)

Figure 5-21 Power Mismatch against No. of Generations with Stochastic Remainder
Selection, Three Encoding Schemes and Inversion

97
 Stochastic Remainder Selection Random
Encoding Deletion & Regeneration

50
40

Mismatch
30
20
10
0
-10
1 101 201 301 401
Generations

(a)
Stochastic Remainder Selection Gray
Encoding Deletion & Regeneration

70

50
Mismatch

30

10

-10
1 201 401 601 801 1001
Generations

(b)
Stochastic Remainder Selection
Complementary Encoding Deletion &
Regeneration

40
30
Mismatch

20
10
0
-10
1 51 101 151 201 251
Generations

(c)

Figure 5-22 Power Mismatch against No. of Generations with Stochastic Remainder
Selection, Three Encoding Schemes and Deletion & Regeneration

98
 CHAPTER 6

Economic Dispatch Studies on NTDC System

6.1 Introduction
This chapter gives discussion on the National Transmission & Dispatch Company
(NTDC) network, its operational problem, and proposed test circuits for economic
dispatch studies. Finally there are economic dispatch studies on NTDC test systems.

6.2 WAPDA System ---- Brief Overview

Need for the optimum development of water and energy resources can not be
over-emphasized in the view of the cardinal role of agriculture in the national economy,
therefore water and power take a very high priority. The Federal Ministry of Water and
Power with two main functionaries─── Water and Power Wings, besides all policy
matters for the developments of theses resources perform certain specific functions. The
Water and Power Development Authority (WAPDA) was created in 1958 with power and
water wings. The charter of duties, which reads as: investigation, planning and execution
of schemes in the fields of Water and Power [82]. WAPDA worked in vertically
integrated environment up to 2001.
The corporatization of WAPDA’s power wing started in 2002. It was divided into
four generation companies, transmission & dispatch-company, nine distribution
companies and one itself WAPDA. The public sector hydel and thermal generation are in
the control of WAPDA and gencos respectively, Independent power producers (IPPs) are
in private sector.

6.3 National Transmission & Dispatch Company

National transmission and dispatch company links the power generation units and
load centers dotting the entire country, thus creating one of the largest contiguous grid
systems of the world. At present NTDC is operating and maintaining nine 500 kV and

99
twenty four 220 kV grid stations along with 10,167 km length of associated transmission
lines [83].
NTDC power system has following mainly four types of power stations connected to the
national grid system:
a. Hydro power stations
b. Steam power stations
c. Gas turbine power stations
d. Combined cycle power plant stations.
e. Nuclear power stations.

The generating units are loaded according to the merit order determined by their
cost of operation and synchronized with system with the rising trend of load curve.
However sometimes it is necessary to take generation at high operational cost subjected
to constraints such as less transmission or transformer capacity. All the functions of
500/220 kV power system and power houses are monitored by SCADA system through
RTUS installed at the grid/power stations.

NTDC is responsible to purchase the power from hydel stations in the north,
thermal units in public and private sectors installed mostly in the central southern regions
of the country and to sell power to distribution companies through its large network of
transmission lines and grid stations of 500 kV and 220 kV voltage capacities.

6.4 Operational Constraints in NTDC System

6.4.1 Hydro-Electric Generation Constraints

Pakistan is the one of the most fortunate countries of the world having the lot of
water potentials. Their total estimated capacity is about 30,000 MW, and all are almost in the
northern areas of Pakistan. WAPDA's hydel power stations consists of five major station
located at Tarbela, Mangla, Warsak, Chashma, Ghazi Boratha and nine small hydel power
stations. Two hydel power stations are presently commissioned in private sector namely
Jagran (AJK) and Malakand( NWFP).
Tarbela and Mangla dams are multi-purpose projects with main emphasis on
irrigation under the indus water treaty. A schematic of WAPDA hydel power station system

100
is shown in Figure 6.1 indicating that the Tarbela and Mangla reservoirs are the major
feature of the WAPDA system, representing the only significant capacity for the seasonal
storage of water. There are two principal effects of wide seasonal variations in hydel
generation capability. These are the variations in water releases from the reservoirs and in
hydraulic heads available for power generation. Operation of the NTDC system is in practice
dominated by these variations.
Water management i.e., the use of reservoir storage and planning of water release is
dominated by irrigation needs rather than power requirements. Seasonal water management
schedules are derived on the basis of data of current and historic water levels in the
reservoirs.
The seasonal variation of the reservoir is briefly analyzed as follows:
1. Prior to the start of the monsoon and snow melting on mountains, water levels in the
reservoirs are low, hence the power output capabilities of the turbines are
significantly less than the installed capacity, however, limited release is required by
the essentially rain-fed areas early part of Kharif cropping season. The hydro-electric
power plant operation is further constrained by the necessary adjustments to planned
releases to prevent upstream or downstream flooding or to makeup for deficient
rainfalls.
2. Water flows to the reservoirs increase rapidly at the beginning of summer, first on
account of snow melt and then with monsoon rains. About beginning of June for
Mangla and ending of July for Tarbela, water is usually abundant and plant can run
on base load. Some times it may be necessary to spill water from the reservoirs in
order to restrict water levels at the dams and allow space for flood control.

The seasonal pattern of reservoir and hydro electric plant operation discussed above
has considerable implications for WAPDA system operation in general. The system
experiences shortages of:

1. Generation capacity and energy during late winter to early summer when reservoir
levels are low.

2. Capacity during the period later in summer, before the reservoirs are filled.

3. Energy during the canal closure period in January to early February when reservoir
releases are severely restricted.

101
Considerable load shedding takes place on NTDC system during the above periods. The load
distribution pattern of NTDC system is such that 75% of the total load is located to the north
of the Multan. So primary transmission system has to transfer the blocks of power from
north to south when full hydel generation capacity is available and from south to north in
winter when water is in short supply and reliance has to be placed on thermal generation.

6.4.2 Thermal Generation Constraints

Seasonal variations in thermal plant capability also have considerable influence on
the system operation. Gas turbines are inherently sensitive to ambient temperature. Hence
the power capabilities of gas turbine components of combined cycle plant are significantly
lower in summer.

6.4.3 Effects of Transmission Constraints

Various transmission constraints have effected the system operation in recent years.
Power flows on NTDC network are from north to south in summer and from south to north
in winter and both the times power transfer is limited by transmission capability subjected to
the constraints such as line or transformer overloading, stability problems and difficulty with
control of voltage or reactive power.

6.4.4 Seasonal Variations in Power Demand

The seasonal variation in electricity demand also effects the system operation. This
variation arises principally from increased air-conditioning loads in summer and heating
load in winter.

102
 Figure 6-1 WAPDA Hydroelectric Generation System

6.5 Test Circuits for NTDC System

The following four test circuits close to the original network of NTDC have been
prepared:
1. Test System 1: 77-bus system shown in Figure 6.2
2. Test System 2: 55-bus system shown in Figure 6.3
3. Test System 3: 44-bus system shown in Figure 6.3
4. Test System 4: 32-bus system shown in Figure 6.4
The bus-bar data has been prepared from the original loadings of the system for
the Test System 2, 3, and 4 [84-85]. However, due to constraints of data the work on test
system 1 is left for future exercise. Test systems 2 and 3 may be extended up to 59 and 50
buses respectively. However, presently 55 and 44 buses have been considered. Test
systems 2, 3, and 4 have been tested for their performance using load flow analysis based
on Newton-Rapshon algorithm. Summary of load flow analysis is given in Table 6.1. The
voltage profiles of these systems are shown in Figure 6.5.

103
Figure 6-2 77-Bus Test System

Figure 6-3 55-Bus Test System

104
 Figure 6-4 44-Bus Test System

Figure 6-5 32-Bus Test System

Table 6-1 Summary of Load Flow Results

S.No. Test System Generation Load Loss No. of
(MW) (MW) (MW) Iterations
1 32- Bus 6096.96 5979.00 117.22 06
2 44-Bus 7719.46 7544.16 175.51 09
3 55-Bus 7381.64 7219.00 162.62 07
4 32-Bus 1625.85 1600 25.85 10

105
 Voltage Profile NTDC 32-Bus Test Circuit

1.20

Bus Voltage (pu)

1.10

1.00

0.90

0.80
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31
Bus Bars

Figure 6-6 Voltage Profile NTDC 32-Bus Test Circuit

Voltage Profile NTDC 44-Bus Test Circuit

1.25
Bus Voltage (pu)

1.15

1.05

0.95

0.85

0.75
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43
Bus Bars

Figure 6-7 Voltage Profile NTDC 44-Bus Test Circuit

Voltage Profile NTDC 55-Bus Test Circuit

1.25
Bus Voltage (pu)

1.15

1.05

0.95

0.85

0.75
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55
Bus Bars

Figure 6-8 Voltage Profile NTDC 55-Bus Test Circuit

106
 The test system of NTDC network for economic dispatch using thermal machines
has been prepared by selecting 32-bus system. This system has generation at 12 points in
the system. The system has been scale down in terms of power generation; system data
and load demand keeping the number of lines and number of generation point same.
Using the tuned bus-bar and system-data, load flow analysis of this system has been
carried. The loss coefficients have been calculated using steps given in chapter 5.
Summary of the load flow analysis is given in Table 6.2 and voltage profile is shown in
Figure 6.6. The 12-Machines 32–bus test circuit with load demand of 1600 MW is thus
proposed for ED dispatch studies
The following three test systems with cost curves close to the original machines in
the system have also been prepared for ED analysis neglecting transmission losses:
1. 15-Machine NTDC Test Circuit with PD = 2400 MW
2. 25-Machine NTDC Test Circuit with PD = 2800 MW
3. 34-Machine NTDC Test Circuit with PD = 3000 MW
The potential of the application of AI tools for WAPDA system [86] has also been
explored. However in the discussion to follow the GA based economic dispatch problem
will be investigated.
Table 6-2 Summary of Load Flow Result: NTDC 32-Bus Test Circuit
For ED studies
Test System Generation Load Loss No. of
(MW) (MW) (MW) Iterations
32-Bus 1625.85 1600 25.85 10

Voltage Profile NTDC Test Circuit

1.20
Bus Voltage (pu)

1.10

1.00

0.90

0.80
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31
Bus Bars

Figure 6-9 Voltage Profile NTDC Test Circuit

107
6.6 Case Studies
The cost curves and B-coefficients for 12-Machine system are given in Tables
A.12 and A.13 respectively placed in appendix A. The cost curves of the 15, 25 and 34
machines are given in Tables A.14, A.15 and A.16 respectively placed in appendix A.
For GA based economic dispatch each test system runs 20 times for both λ-Search
and Real Power-Search and results corresponding to minimum generation cost has been
recorded and compared with the λ-iteration method. The transmission loss has been
calculated using B-coefficients.
The detailed output results in tabular form are given in appendix C. The summary
and comparison of results is given in this chapter.

Evolution Model: Evolution model with GA parameters for four test systems for both λ–
Search and RP-Search are given in Table C.16 and Table C.17 respectively placed in
appendix C.
Output Results: Tables C.18, C.19, C.20 and C.21 placed in appendix C lists the details
of output results of the 12-Machine, 15-Machine, 25-Machines, 34-Machines system
respectively. The Table 5.4 presents the comparison of the results. The observations are
as follows:
• The generation cost is almost the same for λ-Iteration and λ-Search techniques for
all test systems.
• In RP-Search the generation cost is higher than the other two techniques.
Maximum increase is 2.7% with 12-Machines system and minimum increase is
0.7% with 15-Machine system.
• The execution time is higher in λ-Search techniques and RP-Search from λ-
Iteration.
• The increasing trend in time varies from system to system.
• The standard deviation is very low in λ-Search as compared to RP-Search
• The trends for standard test systems in chapter 5 for convex ED studies are almost
repeated on the NTDC test circuits.

108
 Table 6-3 Comparison of the Results for NTDC System
GA Results After 20 Runs
Corresponding to Minimum
Test System λ-Iteration Cost

λ-Search RP-Search

12-Machines 4659491.9674 4664527.9400 4785261.2186

Gen Cost 15-Machines 5512088.5870 5512087 5546108.2504
($/h) 25-Machines 6654911.2924 6654911.7586 6712903.1928
34-Machines 7073116.03408 7073164.5506 7225943.9370
12-Machines 0.062 1.3440 0.3600
Execution 15-Machines 0.031 13.907 2.4690
Time 25-Machines 0.078 21.0150 0.3130
34-Machines 16.3440 0.9060
12-Machines ---- 1.3123 114251.7916
Standard 15-Machines --- 0.00000 60195.1588
Deviation 25-Machines --- 14.7640 81079.3851
34-Machines --- 12.2968 38343.8
% Difference 12-Machines --- -0.1081 -2.6992
in Gen. Cost 15-Machines ---- 0.0000 -0.6172
from λ 25-Machines ---- 0.0000 -0.8714
Iteration 34-Machines ---- -0.0007 -2.1607

109
 CHAPTER 7

Proposed Hybrid Models for Nonconvex

Economic Dispatch Problem

7.1 Introduction
Hybrid methodology deals with the synergistic integration of two or more
approaches. The integrated use of approaches has resulted in effective problem solving in
comparison with each approach used individually and exclusively. Genetic algorithms
have the inherent ability to reach the global minimum region of search space very soon,
but then take longer time to converge the solution. GA based hybrid approaches get
around this problem. This chapter presents three hybrid models for the solution of
nonconvex economic dispatch problem.

7.2 GA Evolution Model for Hybrid Approaches

The method of selecting or picking parents and method of inserting offsprings
back into population taken together is called model of evolution for genetic algorithm.
The working strategy of GA can be considered a balance mix of exploration and
exploitation of search space. This balance controls the performance of GA and is
determined by the suitable selection of GA parameters: population size, probabilities of
crossover and mutation. The selection of the values of control parameters in GA is one of
the most important and promising area of research in evolutionary computation [87]. The
GA models and GA parameters adopted in hybrid models have been selected keeping in
view the investigations carried out in chapter 5.

The transmission loss in GA subprogram has been taken into account using B-
coefficients and in Newton’s second order approach subprogram loss is calculated exactly
due to the inclusion of load flow equation in the modeling. The subsequent sections
present the details of three proposed hybrid models separately.

110
7.3 Hybrid Model-I: A Synergy of GA and Newton’s Second Order Approach
This hybrid approach solves the economic dispatch with valve point effect and
combines the genetic algorithm (GA) with economic dispatch based on the Newton’s
second order approach (NSOA). Three standard test systems have been tested for
validation of the approach.

7.3.1 Economic Dispatch Using NSOA ---- Problem formulation

The Newton’s second order approach is the most effective and fundamental
approach for solving Optimal Power Flow problems with non-separable objective
functions. It tries to minimize the Lagrangian function by approximating the solution
space with quadratic and finding the minimum of quadratic at each of the iteration during
solution.
Mathematically with one objective function and constraints the ED is formulated [68] as:
N
Minimize CT = ∑Ci (Pi ) (7.1)
i =1

Subject to:
i) active and reactive power balance equations for N node system with M
generation buses
∆Pi = Pi (V , θ ) − pi = 0 where i = 1,2,3…, N (7.2)
∆Qi = Qi (V , θ ) − qi = 0 where i = M + 1,…, N (7.3)
Where
Pi (V , θ ) = V i 2 ( g ii + ∑ t ij 2 g ij ) + V i ∑Vj t ij | Y ij | cos ( θ i - θ j - γ ij )
j j

Qi = −Vi 2 ( bii + ∑ t ij 2 bij ) + Vi ∑ V j t ij | Y ij | sin( θ i - θ j - γ ij )

j j

pi = Pg i − Pd i

q i = Q g i − Qd i

ii) active and reactive power generation constraints for all units
Pi (min) ≤ Pi ≤ Pi (max) (7.4)

Qi (min) ≤ Qi ≤ Qi (max) (7.5)

iii) voltage magnitude constraints for all nodes

Vi (min) ≤ Vi ≤ Vi (max)

111
 where:
Pi (V, θ) actual active power injections at bus i
Qi (V, θ) actual reactive power injections at bus i
pi scheduled active power injections at bus i
qi scheduled reactive power injections at bus i
Pg i , Q g i active and reactive power generation at bus i

Pd i , Qd i active and reactive demand at bus i

Vi, Vj voltage magnitude of each bus i, j

Vi(min) , Vi(max) minimum and maximum voltage of bus i
Pi (min) , Pi (max) minimum and maximum active generation of unit i

Q i (min) , Qi (max) minimum and maximum reactive generation of unit i

tij transformer tap ratios

Yii = gii + bii driving point admittance of branch i
Yij ∠γ ij = g ij + bij transfer admittance of branch (i,j)

The Lagrangian function can be written as:

[ ] [ ]
N N
L( y, λ ) = F ( Pg i ) + ∑ λ pi Pi (V , θ ) − Pg i + Pd i + ∑ λqi Qi (V , θ ) − Qg i + Qd i
i =1 i =1

+ Penalty functions for inequalities on variables

+ Penalty functions for functional inequalities (7.6)
Where λ pi and λ qi are Lagrangian multipliers for active and reactive power flow equations

respectively?

While solving, each of the iteration of the algorithm is a simultaneous solution for
all the unknowns in quadratic approximation of Lagrangian function. The unknown
variables include control variables, state variables and Lagrangian multipliers. The second
partial derivatives of Lagrangian function with respect to all variables and Lagrangian
multipliers produces a symmetric matrix called W-matrix having block structure with
each (4x4) block corresponding to an element of nodal admittance matrix of the network
being solved. For less storage and computer time, the use of sparsity based storage and
solution of the linear system has been observed. The bi-factorization technique [88] is
used for factorization and solution of linear systems.

112
 Considering equality constraints only, with active set specified, the equations for
linear system are given by,
W(Z)∆Z = - g(Z ) (7.7)
Equation (6.7) can be expressed in its principle submatrix and subvector components as;
⎡ H(y,λ ) - J T (y)⎤ ⎡ ∆y ⎤ ⎡ - g(y) ⎤
⎢ ⎥⎢ ⎥ = ⎢ ⎥ (7.8)
⎢⎣ - J(y) 0 ⎥⎦ ⎣∆λ ⎦ ⎣- g( λ )⎦

where
H(y, λ) the Hessian matrix of the Lagrangian
J(y) the Jacobian matrix of the power flow equation
g(y) the gradient with respect to y
g(λ) the gradient with respect to λ
W(Z) Composite matrix of J(y) and H(y, λ) called W-matrix
The quadratic penalty functions have been used to handle the inequality constraints which
become active during solution process. The flow chart for implementing the algorithm is
shown in Figure 7.1.

7.3.2 A Synergy of GA-NSOA --- Methodology

In the proposed approach genetic algorithm acts as a global optimizer gives near
optimal generation schedule, which becomes the input for generation buses in the Newton
second order algorithm. This algorithm acting as local search technique dispatches the
generated active power of units for minimization of cost and gives optimum generation
schedule. Three machines 6-bus, IEEE 5-machines 14-bus, and IEEE 6-mchines 30-bus
systems have been tested for validation of the approach. The strategy is illustrated in flow
chart shown in Figure 7.2.

113
 Start

Initialise
x=0
Zk = Z0

Calculate Initial Line Flows

Calculate Active
& Reactive
Power for all Buses

Check
Reactive Power Yes Remove Penalty from
(Q) Limit Lagrangian Multiplier
Violation λq

No
Calculate Gradient Vector
g(Zk)

Exit and Print the Final

Check for Yes Results
Convergence

No
Calculate the
W-Matrix W(Zk)

Solve the Linear system

W(Zk)∆Zk + 1 = -g(Zk) for ∆Zk + 1

Update;
Zk + 1 = Zk + ∆Zk + 1

Update the line flows

Active Set Selection for the Next

Iteration

k=k+1

Figure 7-1 Flow Chart for Economic Dispatch Using NSOA

114
 Start

Read System Data

Initialization

ED using GA

No Termination
Criteria
Reached

Yes

Apply NSOA for ED to obtain

Generation Schedule

Calculate Cost of Generation

End

Figure 7-2 Flow Chart for GA-NSOA Hybrid Approach

7.3.3 Test Results

Following three standard test systems have been selected for the illustration of the
effectiveness of the proposed approach:
1. 3-Machines 6-Bus System
2. IEEE Test Systems
• 5-Machines 14-Bus System
• 6-Machines 30-Bus System

115
 The description about these test systems and there data is given in appendix A.3-
Machine system selected here is described as Test System–I in the appendix. For IEEE
test systems two cases case-I and case-II based on different cost curves have been
defined. The details of the tables placed in appendix A showing the input data is as
follows:
Table A.1: Lists the Generators data of 3-Machine system
Table A.4: Lists the cost curves of 5-Machines System for Case-I and Case-II
Table A.5: Lists the cost curves of 5-Machines System for Case-I and Case-II
Table A.6: Lists the B-coefficients of IEEE Test Systems
The bus-bar and line data of three machine system & IEEE Test systems have been
taken from [128] and [135] respectively..

GA Model and GA Parameters

GA evolution model and GA parameters for 3-Machines and IEEE Test systems
Case-I are given in Table 7.1 and for IEEE Test systems Case –II in Table 7.2
Input Data
The input data for GA and Newton’s second order approach (NSOA) include the
following:
GA : GA Parameters, Power Demand, Machine’s Cost Curves, Loss-
coefficients, Tolerance, Maximum Generation, No. of Runs
NSOA : Base MVA, Tolerance, Bus-bar data, Line-data, Transformer tap settings,
Machine’s Cost Curves, Voltage limits, Active and reactive power limits.
Computer implementation
The NSOA has been integrated with GA approach and implemented in the DOS mode of
PED_Frame to run on PC. The results have been taken on the P-IV 1.60 GHz machine.

Table 7-1 GA – Models & GA Parameters 3 – Machines System & IEEE Test
Systems Case - I
Case – I: IEEE Test Sytems
Test Systems 3-Machines 6-
. 5 Machines 6-Machines
=> Bus
14-Bus 30-Bus
Encoding
Complementary Complementary Complementary
Scheme
GA Model
Selection Tournament Tournament Tournament

116
 Crossover Inversion Inversion Inversion

Mutation Bit Flip Bit Flip Bit Flip

Population size 20 20 20
Chromosome
53 80 95
GA Length
Parameters Crossover
0.8 0.75 0.75
Probability
Mutation
0.001 0.0001 0.0001
Probability

Table 7-2 GA – Models & GA Parameters IEEE Test Systems – Case - II

Case - II
S. No. Test Systems => IEEE 14-Bus IEEE 30-Bus
5-Machines 6-Machines
Encoding Scheme Random Random

Selection Tournament Selection Tournament Selection

GA Model
Crossover Inversion Inversion

Mutation Bit Flip Bit Flip

Population size 20 20

Chromosome Length 85 102

GA
Parameters Crossover
0.75 0.75
Probability
Mutation Probability 0.0001 0.0001

7.3.3.1 Output Results

The results of three test systems are shown in the following tables:
Table 7.3 lists the results of 3-Machines 6-Bus System
Table 7.4 and 7.5 lists results of IEEE 5-Machines 14-Bus System for Case – I and
Case-II respectively.
Table 7.6 and 7.7 lists results lists results of IEEE 6-Machines 30-Bus System for
Case–I and Case-II respectively.

117
 Table 7.8 gives percentage reduction in total production cost from GA to hybrid
approach.

The results of the test systems in the table correspond to minimum generation cost
after 50 runs. The total generation cost variation profiles of the three systems for both GA
and hybrid approaches are shown in Figures from 7.3 to 7.6.

The salient features of the proposed approach in the light of the observations from the
results are as follow:
1. The total production cost profiles shown in Figures 7.3 to 7.6 for test systems
under consideration indicate:
• In all the test cases the total production cost by hybrid approach remains
low as compared to GA alone.
• The GA cost profiles are fluctuating with peaks, dips and flats in the wider
band. Whereas the cost profiles for hybrid approach fluctuate in very
narrow band and remains low from GA.
2. The solution time of the second order approach is low as compared to the GA
in all cases and it takes only 3 to 4 iterations for convergence.
3. Percentage reduction in cost from minimum in GA to minimum in hybrid
approach with respect to minimum in GA varies in the range 0.8-11 %.for the
test system under investigations.
4. Minimum execution time in hybrid approach varies from 0.04 seconds to 10
seconds. The tendency of higher time is due to higher time of corresponding
GA input.
5. The transmission loss is calculated exactly due to inclusion of load flow
equations.
This approach is simple, fast and gives fair amount of reduction in generation cost as
compared to GA alone.

118
Table 7-3 Output Results GA-NSOA: 3-Machines 6-Bus Test System After 50 Runs
GA Results Hybrid Approach Results
Corresponding to Corresponding to
Max. Cost Mini. Cost Max. Cost Mini. Cost
P1 68.0342 98.8968 50.0000 50.0000
P2 76.5807 44.0979 86.0366 86.1270
P3 75.1004 74.5504 79.7428 79.6530
∑Pi (MW) 219.7154 217.5451 215.7793 215.7800
Cost ($/h) 3441.8947 3310.2501 3206.0015 3205.9661
PLoss 9.7158 7.5456 5.7793 5.7800
Tolerance 0.0004 0.0006 0.0000 0.0000
Iterations 2877 7372 3 3
Time (sec.) 1.6250 4.0780 0.0160 0.0160
Time
0.5940 4.0940
GA-NSOA
Standard Deviation 29.0228 0.0082

3-Machines 6-Bus Test System

GA Hybrid

3500

3400
Generation Cost

3300

3200

3100

3000
1 5 9 13 17 21 25 29 33 37 41 45 49
No. of Runs

Figure 7-3 Total Generation Cost Profiles: GA & GA-NSOA Approaches for 3–
Machines System

119
Table 7-4 Output Results GA-NSOA Case-I: IEEE 5-Machines 14-Bus System After
50 Runs with PD=259 MW
GA Results Hybrid Approach Results
Corresponding to Corresponding to
Max. Cost Mini. Cost Max. Cost Mini. Cost
P1 127.0665 151.3922 180.6417 181.126
P2 37.9675 47.6567 46.7289 46.7552
P3 44.6850 36.2761 19.1582 19.1521
P4 30.2956 14.8402 10.5424 10.1912
P5 29.1272 19.6219 10.8886 10.7730
∑Pi (MW) 269.1417 269.7871 267.9599 267.9975
Cost ($/h) 1037.2599 915.3274 906.5187 905.5535
PLoss 10.1417 10.7869 8.9599 8.9975
Tolerance 0.0000 0.0002 0.0000 0.0000
Iterations 263 326 4 4
Time (sec.) 0.4060 0.5310 0.016 0.016
Time GA-NSOA 2.329 2.219
Standard Deviation 37.3834 0.1312

Case-I:IEEE 5 Machines 14 Bus System

GA Hybrid

1050
Generation Cost

1000

950

900

850

800
1 5 9 13 17 21 25 29 33 37 41 45 49
No. of Runs

Figure 7-4 Total Generation Cost Profiles: GA & GA-NSOA Approaches for
Case-I IEEE 5-Machines 14 Bus System

120
Table 7-5 Output Results GA-NSOA Case-II: IEEE 5-Machines 14-Bus System
After 50 Runs with PD=259 MW
GA Results Hybrid Approach Results
Corresponding to Corresponding to
Max. Cost Mini. Cost Max. Cost Mini. Cost
P1 68.4602 92.8362 180.5352 181.1019
P2 81.7015 96.3369 46.7258 46.7552
P3 98.8178 16.7218 19.1613 19.1533
P4 18.7764 52.1543 10.6094 10.2025
P5 29.0295 12.0150 10.9198 10.7827
∑Pi (MW) 296.7854 270.0640 267.9514 267.9956
Cost ($/h) 1327.6141 834.6972 744.9592 744.2422
PLoss 37.7862 11.0648 8.9514 8.9956
Tolerance 0.0009 0.0007 0.0000 0.0000
Iterations 2163 7296 4 4
Time (sec.) 3.3910 11.4380 0.0160 0.0160
Time (sec)
2.0310 4.6030
GA-NSOA
Standard Deviation 109.3692 0.0171

Case - II : IEEE 5-Machines 14 - Bus system

GA Hybrid

1200
Generation cost

1000

800

600
1 5 9 13 17 21 25 29 33 37 41 45 49
No of Runs

Figure 7-5 Total Generation Cost Profiles: GA & GA-NSOA Approaches for
Case-II: IEEE 5-Machines 14 Bus System

121
Table 7-6 Output Results GA-NSOA Case-I: IEEE 6-Machines 30-Bus System After
50 Runs with PD=283.4 MW
GA Results Hybrid Approach Results
Corresponding to Corresponding to
Max. Cost Mini. Cost Max. Cost Mini. Cost
P1 114.1100 158.5465 177.2313 182.5352
P2 63.6466 56.6262 48.6799 48.3554
P3 36.4689 29.9757 20.2517 19.8544
P4 19.9727 19.5614 21.8357 17.1149
P5 18.1661 14.6370 12.0142 13.6485
P6 38.1295 12.5529 13.3451 12.3359
∑Pi (MW) 290.4938 291.8998 293.3579 293.8443
Cost ($/h) 1100.3515 993.3848 992.4302 984.8000
PLoss 7.0942 8.4998 9.9579 10.4443
Tolerance 0.0003 0.0001 0.0000 0.0000
Iterations 499 21 3 4
Time (sec.) 0.9530 0.0320 0.0310 0.0160
Time(sec)
0.2030 0.0480
GA-NSOA
Standard Deviation 22.4336 1.7129

Case-I:IEEE 6 Machines 30 Bus System

GA Hybrid

1150
Generation Cost

1100
1050
1000
950
900
1 5 9 13 17 21 25 29 33 37 41 45 49
No. of Runs

Figure 7-6 Total Generation Cost Profiles: GA & GA-NSOA Approaches for
Case-I: IEEE 6-Machines 30 Bus System

122
Table 7-7 Output Results GA-NSOA Case-II: IEEE 6-Machines 30-Bus System
After 50 Runs with PD=283.4 MW
GA Results Hybrid Approach Results
Corresponding to Corresponding to
Max. Cost Mini. Cost Max. Cost Mini. Cost
P1 52.3506 75.8363 181.6886 175.7710
P2 31.2671 66.8846 48.3137 48.6879
P3 97.9248 35.3572 19.8677 20.3210
P4 58.8753 51.9620 17.4501 22.7443
P5 26.4126 33.2975 13.9438 12.2412
P6 19.9424 24.7958 12.5082 13.4911
∑Pi (MW) 286.7727 288.1332 293.7720 293.2566
Cost ($/h) 1297.2246 924.4308 840.0704 837.6799
PLoss 3.3718 4.7332 10.3720 9.8566
Tolerance 0.0009 0.0001 0.0000 0.0000
Iterations 1422 11381 4 3
Time (sec.) 2.7500 22.0480 0.0310 0.0310
Time (sec)
1.6820 9.8311
GA-NSOA
Standard Deviation 83.6342 0.8500

Case - II : IEEE 6 - Machines 30 - Bus system

GA Hybrid

1250
Generation Cost

1150
1050
950
850
750
1 5 9 13 17 21 25 29 33 37 41 45 49
No. of Runs

Figure 7-7 Total Generation Cost Profiles: GA & GA-NSOA Approaches for
Case-II: IEEE 6-Machines 30 Bus System

123
 Table 7-8 Percentage Reduction in Total Generation Cost in GA-NSOA Hybrid
Approach
% Red.
% Red.
GA Max. GA Min. Hybrid From
. Systems From GA
Cost Cost Mini. Cost GA Min.
Max. Cost
Cost
3-Machines
3441.8947 3310.2501 3205.9661 6.8546 3.1503
System
5-Machines
1037.2599 915.3274 905.5535 12.6975 1.0678
Case-I
IEEE

System
6-Machines
1100.3515 993.3848 984.80 10.5013 0.8717
System
5-Machines
1327.6141 834.6972 744.2422 43.9414 10.8369
Case -II
IEEE

System
6-Machines
1297.2246 924.4308 837.6799 35.4252 10.3561
System

7.4 Hybrid Model –II: Neuro-Genetic Hybrid Approach

ED dispatch problem has been addressed independently by GA [89] and artificial
neural networks [90]. The proposed neuro-genetic approach combines the genetic
algorithm (GA) with the artificial neural network (ANN) using swarm intelligence (SI)
learning rule. Three machines standard test system has been tested for validation of the
proposed approach. The discussion to follow in next sections of this approach is with
reference to context of the discussion in chapter 3.

In this approach GA based economic dispatch works as global optimizer and

produces near optimal solution based training data for neural network model. The neural
network fine tunes the data subject to the target values of power generations of machines.
The best tuned solution is considered the required solution. The target values of the
machine output may be taken by randomly generated values around the near optimal
solutions satisfying both equality and inequality constraints.

The swarm intelligence learning rule based feed forward neural network has been
used for fine tuning the near optimal GA results. The standard back propagation learning
rule neural network has also been used in the hybrid approach for the comparison of
results. Flow chart in Figure 7.8 highlights the steps in the hybrid methodology.

124
 Start

Read System Data

Initialization

ED using GA Generates
Training & Test Data

ANN using SI Learning to obtain

Generation Schedule

No Check For
Comparison ANN
using BP Learning

Yes

ANN using BP Learning to obtain

Generation Schedule

Calculate Cost of Generation

End

Figure 7-8 Flowchart for Neuro-Genetic Approach

7.4.1 Artificial Neural Network using SI Learning Rule

PSO trains the ANN by changing its weight such that mean square error (MSE) for
the training set of an ANN model is reduced. It is a feedback process which runs until the
batch error falls under an acceptable range or iteration cutoff threshold is reached. In the
process of training ANN, particle position corresponds to the weights in the network. The
fitness function corresponds to the mean square error (MSE) of the network [91]. So each

125
particle represents a possible solution network. Particles cooperate with each other to find
an optimal network with minimal MSE. Guaranteed convergence PSO approach [61] has
been used with steps shown in Figure 7.9 for training ANN.

Repeat:
For each particle in swarm
If MSE( current network ) < MSE (personal best network)
then Personal best network = current network
If MSE ( current network )< the MSE (global best network)
then Global best network = current network
Update network weight
Until epoch < max. epochs or MSE (global best network)< acceptable
error
Figure 7-9 Steps for SI Learning ANN

7.4.2 Test Results

Three machine standard test system described as Test System 2 in the appendix A
has been selected for investigation. The cost curves of the machines are given in Table
A.2 placed in appendix A.

7.4.2.1 GA Based Economic Dispatch

a. Input data
The inputs are standard nonconvex machine cost curves; power demand and
GA parameter settings.
b. GA Evolution Model & GA parameters
Random and complementary encoding schemes for initial population
generation with four selection method, seven crossover operators and bit flip
mutation have been adopted. A termination criterion is power balance
equation (i.e., sum of generation = Load) with mismatch tolerance 0.01. The
GA parameters used are: Population size = 20, Chromosome Length = 20 bits,
Crossover Probability pc = 0.8, and Mutation Probability pm = 0.001.

126
 c. Output Results GA based Economic Dispatch
The economic dispatch for the three machines for the total demand of 850
MW has been conducted neglecting transmission loss. Table 7.9 tabulates the
results of minimum total generation cost and corresponding generation
schedule during the 20 runs of the program for each selection method against
four crossover operators with initial population generation with two encoding
schemes. Using the different evolution models the training and test data of
generation schedules like the shown in Table 7.9 has been prepared.

7.4.2.2 ANN Model

In the analysis in the ANN model three inputs (P1, P2, P3) the power generation of
the machines and three outputs P1, P2, P3 have been taken.

Training data and test data consisting of sixty samples of near optimal solution for
three test cases have been generated by GA. The target has been taken as the best solution
[92] attained for this system so far in the literature.
Parameters Settings
a. PSO Model
The parameter settings are: No. of Particles = 10, Constriction factor
χ = 0.7298 , c1 = c2 = 1.496, s c = 15, f c = 5 , the range for individual
dimension of particle position is 21 ≤ x ≤ −21 , the range for individual
dimension of particle velocity is 2.1 ≤ v ≤ −2.1 , max. epochs = 10000, and
acceptable error = 1.0e-12.
b. Backpropagation Algorithm
The learning rate η and momentum rate α both influence the rate of
convergence. The value of both of theses rates lie between 0 and 1. Typical
values lie in the range of 0.5-0.9. The number of samples and hidden layers
have selected through experimentation for different combinations. The
parameters selected are: No. of samples = 60, No of inputs and outputs = 3, α
= 0.85, η = 0.8, No of hidden layers = 2.

127
7.4.2.3 Output Results
Sixty samples consisting of three inputs to the neural network produce fine tuned
sixty out puts each consisting of three power generations P1, P2, P3. The total production
cost for all samples and its percentage error with respect to target production cost have
been calculated.

The best fine tuned results generated from both the ANN models amongst the
sixty samples and corresponding near optimal input generated by GA have been tabulated
in Table 7.10 and Table7.11 respectively. The percentage error of the generation cost has
been calculated with reference to the cost corresponding to the target generation schedule.

The row seven, eight, nine and ten in Table7.10 tabulate the minimum/maximum
percentage errors of total production cost in the set of sixty samples for ANN models and
GA respectively. The plot of percentage error for three test cases against the number of
samples is shown in Figures 7.10, 7.11 and 7.12. The results in the Table 7.10 and Table
7.11 and percentage error profiles shows: (1) Neuro-Genetic approach with both the ANN
models gives better results as compared to GA alone, (2) Nero-Genetic approach based on
Swarm Intelligence learning rules is superior to backpropagation learning.

Table 7-9 GA Based Economic Dispatch --- Best Solution After 20 Runs
Crossover Encoding Roulette Wheel Selection
Operators Schemes P1 P2 P3 ∑ Pi Cost
(MW) (MW) (MW)
One Point Random 575.49 218.67 55.84 850.00 8740.30
Complementary 594.70 104.63 150.67 850.00 8449.11
Uniform Random 402.64 300.20 147.16 850.01 8426.52
without Complementary
Mask 305.45 389.84 154.71 850.01 8396.58
Uniform Random 495.35 270.39 84.27 850.01 8537.50
With Mask Complementary 483.53 253.40 113.08 850.01 8503.02
Inversion Random 598.18 159.14 92.68 850.00 8516.41
Complementary 402.99 396.69 50.32 850.00 8286.70
Crossover Encoding Tournament Selection
Operators Schemes P1 P2 P3 ∑ Pi Cost
(MW) (MW) (MW)
One Point Random 406.57 256.07 187.36 849.99 8454.23
Complementary 391.62 336.86 121.51 850.00 8512.01
Uniform Random 401.83 333.17 114.99 849.99 8413.77
without Complementary 393.73 312.62 143.65 850.00 8399.99

128
Mask
Uniform Random 396.55 354.54 98.90 850.00 8422.60
With Mask Complementary 312.54 390.75 146.70 850.00 8431.58
Inversion Random 404.18 308.86 136.96 850.00 8473.30
Complementary 493.81 304.99 51.19 849.99 8441.18
Crossover Encoding Rank Selection
Operators Schemes P1 P2 P3 ∑ Pi Cost
(MW) (MW) (MW)
One Point Random 510.08 212.21 111.85 834.15 8506.26
Complementary 487.60 253.41 109.00 850.01 8440.32
Uniform Random 312.09 399.46 138.44 850.00 8430.41
Without
Complementary 407.98 333.97 108.05 850.00 8429.74
Mask
Uniform Random 391.00 315.43 143.59 850.01 8404.27
With Mask Complementary 406.16 337.44 106.40 850.00 8425.27
Inversion Random 379.16 330.23 140.62 850.01 8503.43
Complementary 408.45 317.72 123.84 850.01 8486.58
Crossover Encoding Stochastic Remainder Selection
Operators Schemes P1 P2 P3 ∑ Pi Cost
(MW) (MW) (MW)
One Point Random 502.46 203.17 144.37 850.00 8517.59
Complementary 506.40 240.89 102.71 850.00 8401.45
Uniform Random 384.87 329.65 135.48 850.00 8487.22
Without
Complementary 518.49 174.50 157.01 850.00 8516.73
Mask
Uniform Random 398.28 345.57 106.17 850.01 8417.96
With Mask Complementary 394.43 323.11 132.46 850.00 8383.71
Inversion Random 292.57 393.93 163.51 850.00 8448.28
Complementary 307.59 396.77 145.64 850.00 8350.41

Table 7-10 Best Results by Hybrid Approach amongst the Sixty Samples
Test Case 1 Test Case 2 Test Case 3
Target ANN Model ANN Model ANN Model
SI BP SI BP SI BP
P1 300 300.01 299.72 300.01 299.71 299.99 299.73
P2 400 399.98 400.20 399.99 400.2 400 400.20
P3 150 150.02 150.07 149.99 150.04 150 150.05
∑Pi (MW) 850 850.01 849.99 849.99 849.95 849.99 849.98
Cost ($/h) 8234.06 8234.43 8234.03 8234.04 8233.36 8234.04 8233.67
Mini.%Error --- 0.0000 0.0010 0.0004 0.0046 0.0004 0.0048
Max. %Error --- 0.0142 0.2420 0.0042 0.1656 0.0094 0.1638
Mini. % --- 0.350 0.214 0.214
Error GA
Max. % --- 7.528 6.815 6.815
Error GA

129
Table 7-11 GA Input Corresponding to the Best Results by Hybrid Approach in
Table 7.11
Test Case 1 Test Case 2 Test Case 3
ANN Model ANN Model ANN Model
SI BP SI BP SI BP
P1 593.41 505.65 411.98 503.68 477.35 373.11
P2 158.98 250.48 253.85 255.12 251.46 295.93
P3 97.61 93.77 184.17 91.2 121.17 180.96
∑Pi (MW) 850 850 850 850 850 850
Cost ($/h) 8513.76 8359.33 8502.96 8398.99 8569.24 8758.67

Test Case 1
SI BPA

0.3

0.2
% Error

0.1

-0.1
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57
No. of Samples

Figure 7-10 Test Case 1: Comparison of Percentage Error of the Generation Cost

Test Case 2
SI BPA

0.2
% Error

0.1

-0.1
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61
No. of Samples

Figure 7-11 Test Case 2: Comparison of Percentage Error of the Generation Cost

130
 Test Case 3
SI BPA

0.2

% Error
0.1

-0.1
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57
No. of Samples

Figure 7-12 Test Case 3: Comparison of Percentage Error of the Generation Cost

7.5 Hybrid Model–III: A Hybrid of GA and Sequential Quadratic Programming

(SQP)
This approach explores the use of GA with Sequential Quadratic Programming
(SQP). Case study has been conducted by taking standard three machines, five machines
and six machines systems.

7.5.1 Sequential Quadratic Programming (SQP)

Sequential quadratic programming methods represents the state of the art in
nonlinear programming methods. Its implementation involves following three stages:
1. Updating of the Hessian matrix of the Lagrangian function
2. Quadratic programming problem solution
3. Line search and merit function calculation

In the present work SQP has been used by adopting MATLAB function
“fmincon”. This function attempts to find a constrained minimum of a scalar function of
several variables starting at an initial estimate using SQP for medium scale optimization.
It solves a quadratic programming (QP) subproblem at each iteration. An estimate of the
Hessian of the Lagrangian is updated at each iteration using the BFGS formula.

131
7.5.2 Genetic Algorithm – Sequential Quadratic Programming (GA-SQP) Hybrid
Methodology
The GA working as base line search acts as global optimizer and produces good
potential solutions. These solutions become input for SQP subprogram and produces
ultimately optimal results.

7.5.3 Case Studies

a. Test systems
Three following test systems have been investigated for the illustration of
the effectiveness of the proposed approach.
1. 3-Machines System with power demand PD = 850 MW.
2. IEEE 5-Machines & 6-Machines Test Systems with cost curves
of Case-I.
The description & generator’s data is given in the appendix A. The Tables A1,
A4 and A5 give the generator’s data of 3-Machines, IEEE 5-Machines and 6-
Machines system respectively.
b. Transmission Loss
The transmission loss has been neglected.
c. GA Model and GA Parameters
GA model and GA parameters for 3-Machines and IEEE Test systems are
given in Table 7.13.
d. Input Data
The input data for GA and SQP algorithms include the following:
GA : GA Parameters, Power Demand, Machine’s Cost
Coefficients, Tolerance, Maximum Generation, No. of
Runs.
Quazi : Machine’s Cost Coefficients, Active Power Limits,
Newton Power Demand and Tolerance.
e. Computer implementation
The GA has been implemented in PED_Frame. Economic dispatch
solution based on SQP has been modeled using Matlab “fmincon” function in
Matlab programming environment. The results have been taken on the P-IV
1.60 GHz machine.

132
 Table 7-12 GA–Models & GA Parameters for GA-SQP Hybrid Approach
IEEE Test Systems
Test Systems => 3-Machines 5-Machines 14- 6-Machines
Bus 30-Bus
Encoding
Complementary Complementary Complementary
Scheme
Selection Tournament Tournament Tournament
GA Model
Crossover Inversion Inversion Inversion

Mutation Bit Flip Bit Flip Bit Flip

Population size 20 20 20
Chromosome
56 85 102
GA Length
Parameters Crossover
0.8 0.75 0.75
Probability
Mutation
0.001 0.0001 0.0001
Probability

f. Output results
Output results have been recorded after 50 trails and the Tables 7.13, 7.14 and
7.15 list results of 3-Machines, 5-Machines 6-Machines systems respectively.
Table 7.16 gives percentage reduction in total production cost. The total
generation cost profiles for GA based ED and GA-SQP hybrid approaches for
the three systems are shown in Figures 7.13, 7.14, and 7.15.
The salient features of the proposed approach in the light of the observations from
the results are as follow:
1. The total production cost profiles shown in Figures 7.13 to 7.15 for test
systems under consideration indicate that in all the test cases the total
production cost by GA-SQP hybrid approach remains low as compared to GA
alone.
2. The GA cost profiles are fluctuating with peaks and dips in the wider band.
Whereas the cost profiles for hybrid approach fluctuate in very narrow band
and remains low from GA.

133
 3. The results of 3-Machine systems are the better than the results obtained by
GA [9] so far and same as that obtained by evolutionary programming [93]
and improved particle swarm optimization [94].
4. The SQP approach takes less iteration for convergence as compared to GA.

Percentage reduction in cost with respect to maximum and minimum in GA shows

the contrast improvement in the results and hence the promise of the approach.

Table 7-13 Output Results GA-SQP Approach 3-Machines Test System after 50
Runs for PD=850 MW
GA Results Hybrid Approach Results
Corresponding to Corresponding to
Max. Cost Mini. Cost Max. Cost Mini. Cost
P1 544.3046 399.4438 589.2 300.2662
P2 139.8398 251.3055 163.3038 400
P3 165.8595 199.2504 96.4963 149.7338
∑Pi (MW) 850.0038 849.9997 850.0001 850.0001
Cost ($/h) 8936.2848 8256.3990 8517.6 8234.1
Tolerance 0.0038 0.0003 0.0001 0.0001
Iterations 333 1878 15 46
Time (sec.) 0.25 1.5940 2.2813 3.125
Time (sec) GA-SQP 2.3903 5.297
Summary of Results
Min Cost Mean Cost Max Cost Std Cost Mean Time
GA 8256.3990 8629.2807 8936.2838 153.3127 2.9737
GA-SQP 8234.1000 8270.5560 8517.6000 71.7253 7.8625

3-Machines System
GA Hybrid
Generation Cost

8900
8700
8500
8300
8100
7900
1 5 9 13 17 21 25 29 33 37 41 45 49
No. of Runs

Figure 7-13 Total Generation Cost Profiles: GA & GA-SQP Approaches for
3-Machines System

134
 Table 7-14 Output Results GA-SQP Approach IEEE 5-Machines 14-Bus System
After 50 Runs For PD=259 MW
GA Results Hybrid Approach Results
Corresponding to Corresponding to
Max. Cost Mini. Cost Max. Cost Mini. Cost
P1 120.6317 86.2886 150.0029 154.0001
P2 22.4225 73.5074 28.9962 54.9999
P3 88.9871 35.3066 15 15
P4 13.4882 47.4658 54.9999 25
P5 13.4714 16.4309 10 10
∑Pi (MW) 259.0009 258.9993 259 259
Cost ($/h) 1051.8277 798.7345 722.6106 697.8751
Tolerance 0.0009 0.0007 0 0
Iterations 3608 8178 41 64
Time (sec.) 4.6560 10.5620 3.4688 4.6719
Time (sec) GA-SQP 12.4688 9.3279
Summary of Results
Min Cost Mean Max Cost Std Cost Mean Time
GA 798.7345 898.6096 1051.8277 59.1970 4.1939
GA-SQP 697.8751 701.4576 722.6106 3.5258 7.4848

IEEE 5-Machines 14 Bus System

GA Hybrid
Generation Cost

1050
950
850
750
650
1 5 9 13 17 21 25 29 33 37 41 45 49
No. of Runs

Figure 7-14 Total Generation Cost Profiles: GA & GA-SQP Approaches for
5-Machines System

135
 Table 7-15 Output Results GA-SQP Approach IEEE 6-Machines 30-Bus System
after 50 Runs for PD=283.4 MW
GA Results Hybrid Approach Results
Corresponding to Corresponding to
Max. Cost Mini. Cost Max. Cost Mini. Cost
P1 58.2482 98.4461 150.0029 200.0044
P2 38.3316 85.6557 26.3972 36.3956
P3 92.0688 35.2794 15 15
P4 39.2213 12.3154 69.9999 10
P5 40.2411 30.9697 10 10
P6 15.2887 20.7346 12 12
∑Pi (MW) 283.3996 283.4008 283.4 283.4
Cost ($/h) 1232.2153 873.9894 817.7333 781.3363
Tolerance 0.0003 0.0008 0 0
Iterations 4321 8166 34 22
Time (sec.) 6.8440 12.7660 3.5 2.7656
Time (sec) GA-SQP 6.4380 18.4686
Summary of Results
Min Cost Mean Max Cost Std Cost Mean Time
GA 873.9894 1011.6711 1232.2153 74.0979 7.2593
GA-SQP 781.3363 784.4656 817.7333 7.7434 11.0309

IEEE 6-Machines 30 Bus System

GA Hybrid
Generation Cost

1150
1050
950
850
750
650
1 5 9 13 17 21 25 29 33 37 41 45 49
No. of Runs

Figure 7-15 Total Generation Cost Profiles: GA and GA-SQP Approaches for
6-Machines System

136
 Table 7-16 Percentage Reduction in Total Generation Cost in GA-SQP Hybrid
Approach
GA % Red. % Red.
S. GA Hybrid
Systems Min. From GA From GA
No. Max. Cost Mini. Cost
Cost Max. Cost Min. Cost
3-Machines
1 8936.2848 8256.399 8234.1 7.8577 0.2701
System
IEEE
2 5-Machines 1051.8277 798.7345 697.8751 33.6512 12.6274
System
IEEE
3 6-Machines 1232.2153 873.9894 781.3363 36.5909 10.6012
System

137
 CHAPTER 8

Conclusions and Suggestions

This chapter is primarily concerned with more general aspects of the work
described in the thesis. The detailed aspects of the material contained in the thesis have
been discussed in the respective chapters.

The research work presented has been entirely computer oriented and the main
motivation has been to develop economic dispatch solution using GA independently and
GA based hybrid models.

ED is the essential and vital step in power system operational planning. It is

carried out for both off-line as well as on-line studies. Actual ED problem is nonconvex
in nature. Generally it is solved as convex ED problem in mathematical programming
(MP) environment by taking the monotonically increasing quadratic cost curve resulting
in an inaccurate dispatch. However, the nonconvex ED problem in mathematical
programming environment has been addressed using dynamic programming, but due to
curse of dimensionality it is not preferred for large systems.

Genetic algorithm is potential solution methodology for ED problem, as it

addresses the convex and nonconvex economic dispatch equally. Genetic algorithm is fast
in exploring the global minimum regions but then takes longer time to converge the local
minimum solution. The performance of GA can be enhanced either independently at its
components level by adopting new encoding schemes, new operators, good control of GA
parameters etc., or by integrating it with other techniques to counter its weaknesses.

GA based ED problem has been formulated using two search methodologies

namely λ-Search and real power search. Former has been used for convex ED whereas
latter has been used for both convex and nonconvex ED solutions.

138
 Flexible and extensible the computational framework namely “PED_Frame” has
been developed as common environment for implementations of different algorithms for
economic dispatch solution. The algorithms used in the present work have been
implemented in the PED_Frame. The convex ED dispatch has been conducted in its
visual environment, whereas nonconvex ED has been solved using its DOS mode.

Different evolution models for GA have been developed by adopting three

encoding schemes for initial population generation, four selection methods, and different
crossover operators. Two biological methods namely inversion and deletion-regeneration
has been implemented as an operator with crossover probability. Based on these evolution
models ED studies have been conducted.

Pakistani utility system and its operational problems have been reviewed with a
view to carry out its ED dispatch studies using λ-iteration and GA based models. Its four
load flow test circuits consisting of 77-bus, 55-bus, 44-bus and 32-bus have prepared. The
load flow of the last three test circuits has been carried out by implementing Newton’s
Rapshon methods to validate their performance. Finally the data of 12-Machines 32-bus
test system has been scale down and tuned for ED dispatch test circuit. This is the test
circuit with standard inputs such as machine’s cost curves, B-coefficients, and load
demand. In addition to this test circuit 15, 24, 34 Machines test systems have also been
prepared to carry out ED neglecting transmission loss.

ED studies have been conducted using the following test systems:

1. Three different 3-Machine test systems
2. IEEE 5-Machine 14-Bus system
3. IEEE 6-Machne 30-Bus System
4. American Power company 6-Machine 26-Bus system
5. Pakistani Utility Systems

Based on case studies conducted for GA based convex economic dispatch it is

concluded that RP-Search gives near optimal results, λ-Search results are close to the λ-
iteration and with higher mismatch tolerance it takes longer time, and the standard

139
deviation is very low in λ-Search as compared to RP-Search. The execution time is low
for λ iteration as compared to GA search methodologies.

Hybrid methodology for convex ED is proposed in which after specified number

of generations in GA subprogram control is passed to λ-iteration which generates final
results. This approach gives new method for initial selection of λ through systematic
calculation rather than guess work. The approach has been tested on 3-Machine and 20-
Machine systems. The results show that the total generation cost for λ-Iteration and
hybrid approach remains the same. However, the number of iterations is reduced in
hybrid approach and hence the time.

For nonconvex economic dispatch total production cost has been investigated by
using ninety six evolution models with a view to observer its variation subject to
encoding schemes, selection methods and crossover operators.
• In case of random encoding the costs are comparable at one point crossover,
uniform crossover, inversion and deletion & regeneration operators. The
variations are consistent with tournament and rank selection as compared to
the other two methods.
• For Gray encoding scheme there are stable variations in generation cost as
compared with other three selection methods.
• In complementary encoding the generation costs are comparable for all four
selection methods against all eight operators.

Tournament and rank selection methods are fast as compared to the other two
selection methods in terms of convergence.

Initially variations and then smooth transitions towards convergence profiles

occur for one point and uniform crossover operators with all four selection methods.
Whereas spikes like variations and relaxed variations with peaks, flats and saddles
profiles are with inversion and deletion-regeneration operators.

GA in the nonconvex ED with different evolution models give near optimal

results. To fine tune these results three GA based hybrid model have been proposed.

140
 Hybrid model-I combines GA based ED with ED using Newton Second order
approach. GA works as global optimizer produce near optimal results. The economic
dispatch based on Newton’s second order approach using these results produces optimal
results. The Newton’s approach for ED is the direct application of classical Newton’s
method for minimization of multivariable nonlinear function. Each iteration of the
solution algorithm is the simultaneous solution of all the unknowns in the quadratic
approximation of the Lagrangian function of the problem. The bi-factorization technique
for the solution of linear system and sparsity storage has been observed in order to use
less storage and computer time. The hybrid model has been tested on three machine
system, IEEE 5-Machine 14-Bus and IEEE 6-Machine 30-Bus system. For IEEE test
systems two different sets of nonconvex cost curves have been used. The results of
standard test systems obtained using this approach show the comprehensive reduction in
cost thus indicating the promise of the approach.

Hybrid Model-II integrates the GA with swarm intelligence learning based

artificial neural network (ANN). For a given load demand GA generates the training and
test data which is fine tuned by feedforward neural network using swarm intelligence
learning rule based on guaranteed convergence Particle Swarm Optimization to desired
targets. The best tuned result in the given set of sample is the desired generation schedule.
For comparison same test and training data presented to standard backpropagation
learning rule neural network. The case studies for three machine test cases show that
swarm intelligence based learning ANN produce better results than backpropagation
learning neural network.

The Hybrid Model-III combines GA with sequential quadratic programming

(SQP). The near optimal results are fed to SQP subprogram which ultimately generates
the optimal results. In the present work SQP has been used by adopting MATLAB
function “fmincon”. This function attempts to find a constrained minimum of a scalar
function of several variables starting at an initial estimate using SQP for medium scale
optimization. It solves a quadratic programming (QP) subproblem at each iteration. An
estimate of the Hessian of the Lagrangian is updated at each iteration using the BFGS
formula. Three test systems have been tested. Final results are better than those obtained
from GA alone and comparable to those obtained in the literature.

141
 The demonstration of validity of applying the proposed models to the solution of
economic dispatch problem in this thesis gives rise to the number of topics for further
research in this area. Some of the recommendations for future research may be
summarized as follows:
• Performance enhancement of GA for ED at its components level by
developing new operators, encoding schemes, and control of GA parameters.
• Exploration of new hybrid models.
• Exploration of Hybrid model I & III for on-line applications.
• Exploration of Hybrid Model-I for large scale nonconvex ED with application
to utility system.
• Exploration for nonconvex large scale ED using GA by machine clustering.
• The practical application of the programs developed can greatly be enhanced
by extending PED_Frame with following directions:
1. Enhancement of visual environment to provide flexible, highly interactive
and convenient procedures for system inputs, outputs and control priorities
for program control functions.
2. Intermediate interruption of execution of algorithm and resumption of
execution of the algorithm from the same status.
3. Graphics aid for analysis.
• ED Solution of Pakistani utility system using hybrid models.

It is hoped that that in future the nonconvex ED calculations will be extensively be

used, as power industry would continuously seek to maintain a secure supply of electrical
energy at maximum economy and efficiency. This trend will accelerate if improved
methods combined with clearer application oriented formulations and models can be
developed into high quality production grade programs for on-line/offline studies.

142
 References

[1]. E. R. Laithwaite and L. L. Freris, Electrical Energy: its generation,

transmission and use, McGRAW-HILL Book Company (U.K), 1980.
[2]. H. H. Happ, “Optimal power dispatch --- A comprehensive survey,” IEEE
Transactions on PAS, vol. PAS-96, No. 3, pp. 841-854, May June 1977.
[3]. IEEE Working Group Report, “Description and bibliography of major
economy-security functions Part I: Description,” IEEE Transactions PAS, pp.
211-214, January 1981.
[4]. IEEE Working Group Report, “Description and bibliography of major
economy-security functions Part II: Bibliography (1959-1972),” IEEE
Transactions PAS, pp. 215-223, January 1981.
[5]. IEEE Working Group Report, “Description and bibliography of major
economy-security functions Part III: Bibliography (1973-1979),” IEEE
Transactions PAS, pp. 224-235, January 1981.
[6]. B. H. Chowdhury and S. Rahman, “A review of recent advances in economic
dispatch,” IEEE Transactions on Power Systems, vol. 5 No.4, pp. 1248-1259,
November 1990.
[7]. Zi-X. Liang and J. D. Glover, “Improved cost functions for economic dispatch
computations,” IEEE Transactions on Power Systems, vo1.6, No. 2, pp. 821-
829, May 1991.
[8]. W. M. Ling, F. S. Cheng, and M. T. Tsay, “Nonconvex economic dispatch by
integrated artificial intelligence,” IEEE Transactions on Power Systems, vol.
16, No. 2, pp. 307-311, May 2001.
[9]. D. C. Walters and G. B. Sheble, “Genetic algorithm solution of economic
dispatch with valve point loading,” IEEE Transactions on Power Systems, vol.
8 , Issue 3, pp. 1325–1332, Aug. 1993.
[10]. C. E. Lin and G. L. Vivani, “Hierarchical Economic dispatch for piecewise
quadratic cost functions,” IEEE Transaction on Power Apparatus and System,
vol. PAS-103, No.6, June 1983.

143
[11]. S. O. Orero and M. R. lrving, “Economic dispatch of generators with
prohibited operating zones: a genetic algorithm approach,” IEE Proceedings
Gener. Transm. Disrib., vol. 143, No. 6, pp.529-534, November 1996.
[12]. Z. X. Liang and J. D. Glover, “A zoom feature for a dynamic programming
solution to economic dispatch including transmission losses,” IEEE
Transactions on Power Systems, vol. 7, No. 2, pp. 544-550, 1992.
[13]. S. Rajasekaran and G. A. V. Pai, Neural Networks, Fuzzy Logic, and Genetic
Algorithms, Prentice Hall of India, New Delhi, 2004.
[14]. Grimaldi, E. A, Grimaccia, F., Mussetta, M., and Zich, R. E., "PSO as an
effective learning algorithm for neural network applications," Proceedings of
3rd International Conference on Computational Electromagnetic and its
Applications, 2004, pp. 557-560.
[15]. K. P. Dahal, S. J. Galloway, G. M. Burt, and J. R. McDonald, “Generation
scheduling using genetic algorithm based hybrid technique,” Proceedings of
Large Engineering System Conference 2001, July 2001, pp. 74-78.
[16]. P. Attaviriyanupap, H. Kita, E. Tanaka and J. Hasegawa, “A hybrid EP and
SQP for dynamic economic dispatch with nonsmooth fuel cost function,”
IEEE Transactions on Power Systems, vol. 17, No. 2, pp. 411-416, May 2002.
[17]. T. K. A. Rahman, Z. M. Yasin and W. N. W. Abdullah, “Artificial-Immune-
Based for solving economic dispatch in power system,” Proceedings of
National Power & Energy Conference (PECon) 2004, Malaysia, 2004, pp.31-
35.
[18]. Y. H. Song and C. S. V. Chou, “Advanced engineered-conditioning genetic
approach to power economic dispatch,” IEE Proc. Gener. Transm. Distrib.,
vol. 144, No.3, pp. 285-292, May 1997.
[19]. M. Yoshimi, K. S. Swarup, and Y. Izui, “Optimal economic power dispatch
using genetic algorithms,” Proceedings of the Second International Forum on
Applications of Neural Networks to Power Systems, 19-22 April 1993, pp.
157 – 162.
[20]. H. Ma, A. A. El-Keib and R. E. Smith, “A genetic algorithm-based approach
to economic dispatch of power systems,” Proceedings of IEEE Conference on
Creative Technology Transfer – A Global Affairs, April 1994, pp. 212-216.

144
[21]. A. Bakirtzis, V. Petridis, S. Kazarlis, “Genetic algorithm solution to the
economic dispatch problem,” IEE Proc. Gener. Transm. Distrib., vol. 141, No.
4, pp.377-382, July 1994.
[22]. Po-Hung Chen and Hong-Chan Chang, “Large-scale economic dispatch by
genetic algorithm,” IEEE Transactions on Power Systems, vol. 10. No. 4, pp.
1919-1926, November 1995.
[23]. G. B. Sheble and K. Brittig, “Refined genetic algorithm -- economic dispatch
example,” IEEE Transactions on Power Systems, vol. 10, No. 1, pp. 117-124,
February 1995.
[24]. P. Y. Wang, G. S. Wang, Y. H. Song and A. T. Johns, “Fuzzy logic controlled
GA,” Proceedings of International Conference on Fuzzy Systems, Sept. 1996,
vol. No. 2, pp. 972 –979.
[25]. S. O. Orero and M. R. lrving, “Economic dispatch of generators with
prohibited operating zones: a genetic algorithm approach,” IEE Proc. Gener.
Transm. Distrib., vol. 143, No. 6, pp. 529-534, November 1996.
[26]. F. Li, R. Morgan and D. Williams, “Towards more cost saving under stricter
ramping rate constraints of dynamic economic dispatch problems - a genetic
based approach,” Second International Conference on Genetic Algorithms in
Engineering Systems: Innovations & Applications, Sept. 1997, Conf. Pub. No.
446, pp. 221-225.
[27]. Y. H Song, G. S. Wang, P. Y. Wang and A. T. Johns,
“Environmental/economic dispatch using fuzzy logic controlled GA,” IEE
Proc. Gener. Transm. Distrib., vol. 144 No. 4, pp. 377-382, July 1997.
[28]. Y. H. Song and C. S. V. Chou, “Advanced engineered-conditioning genetic
approach to power economic dispatch,” IEE Proc. Gener. Transm. Distrib.,
vol. 144, No. 3, pp. 285-292, May 1997.
[29]. C. Achayuthakan and S. C. Srivastava, “A genetic algorithm based economic
load dispatch solution for eastern region of EGAT system having combined
cycle and cogeneration plants," Proceedings EMPD 98, March 1998, vol. 1,
pp. 165-170.
[30]. A. S. Chuaung and F. Wu, “An extensible genetic algorithm framework for
problem solving in a common environment,” IEEE Transactions on Power
systems, vol. 15, No. 1, pp. 269-275, February 2000.

145
[31]. T. Yalcinoz,, H. Altun, and M. Uzam, “Economic dispatch solution using a
genetic algorithm based on arithmetic crossover,” Power Tech Conference,
Portugal, Sept. 2001.
[32]. T. Yalcinoz and H. Altun, “Environmentally constrained economic dispatch
via a genetic algorithm with arithmetic crossover,” IEEE African 2002, pp.
923-928.
[33]. Ying-Yi Hong and Chih-Yuan Li, “Genetic algorithms based economic
dispatch for cogeneration units considering multiplant multibuyer wheeling,”
IEEE Transactions on Power Systems, vol. 17, No. 1, pp. 134-140, February
2002.
[34]. J. Tippayachai, W. Ongsakul, and I. Ngamroo, “Parallel micro genetic
algorithm for constrained economic dispatch,” IEEE Transactions on Power
Systems, vol. 17 No. 3, pp. 790-797, August 2002.
[35]. P. T. Leite, A. A. de F. M. Carneiro, and A. C. de P. L. F. de Carvalho,
“Energetic operation planning using genetic algorithms,” IEEE Transactions
on Power Systems, vol. 17, No. 1,173 – 179, February 2002.
[36]. A. G. Damousis, A. Bakirtzis and P. S. Dokopoulos, “Network-constrained
economic dispatch using real-coded genetic algorithm,” IEEE Transactions on
Power Systems, vol. 18, No. 1, pp. 198-205, February 2003.
[37]. L. G. Sewtohul, R. T. F. Ah King and H. C. S. Rughooputh, “Genetic
algorithms for economic dispatch with valve point effect,” Proceedings of the
2004 IEEE International Conference on Networking, Sensing & Control
Taipei, Taiwan, March 2004, pp. 1353-1363.
[38]. S. H. Hosseini and M. Kheradmandi, “Dynamic economic dispatch in
restructured power systems considering transmission costs using genetic
algorithm,” Proceedings of Electrical & Computer Engineering, Canadian
Conference May 2004, vol. 3, pp. 1625-1628.
[39]. G. A. Bakare, U. O. Aliyu, G. K. Venayagamoorthy, and Y. K. Shu’aibu,
“Genetic algorithms based economic dispatch with application to coordination
of Nigerian thermal power plants,” Proceedings of IEEE Power Engineering
Society, General Meeting 2005, pp. 2886-2891.
[40]. Chao-Lung Chiang, “ Improved genetic algorithm for power economic
dispatch of units with valve-point effects and multiple fuels,” IEEE

146
 Transactions on Power System, vol. 20, No. 4, pp. 1690 –1699, November
2005.
[41]. K.P. Wong, and C.C. Fung, “Simulated annealing based economic dispatch
algorithm”, IEE Proc. – C Gener. Transm. Distrib., vol. 140, No. 6, pp. 509-
515, 1993.
[42]. Z. X. Liang and J. D. Glover, “A zoom feature for a programming solution to
economic dispatch including losses”, IEEE Transactions on Power System
vol. 7, No. 3, pp. 544-550, 1992.
[43]. K. P. Wong and Y. W. Wong, “Genetic and genetic/simulated annealing
approach to economic dispatch, IEE Proc. Gener. Transm. Distrib., vol. 141
No. 5, pp. 507-513, Sept. 1994.
[44]. W. Ongsakul and J. Tippayachai, “Constrained economic dispatch by micro
genetic algorithm based on migration and merit order loading solutions,”
Proceedings of International Conference on Electric Utility Deregulation and
restructuring and Power Technologies 2000, City University, London, 4-7
April 2000, pp. 510-517.
[45]. W. Ongsakul and N. Ruangpayoongsak, “Constrained dynamic economic
dispatch by simulated annealing / genetic algorithms,” Proceedings of Power
Industry Computer Application 2001 May 20 – 24, 2001, pp. 207-212.
[46]. T. Yalcinoz, H. Altun, “Power economic dispatch using a hybrid genetic
algorithm,” IEEE Power Engineering Review March 2001, pp. 59-60.
[47]. N. Kumarappan, “Neuro-Hybrid genetic algorithm based economic dispatch
for utility system,” Proceedings of International Joint Conference on Neural
Networks, July 20-24, 2003, pp. 2112-2117.
[48]. R. C. Eberhart, “Theory of evolutionary computation”. In Tutorial on Modern
Heuristic Optimization Techniques with Application to Power System, ed. by
K.Y. Lee and M.A. El-Sharkawi, IEEE Power Engineering Society
(02TP160), p. 1, 2002.
[49]. David E. Goldberg, Genetic Algorithm in Search, Optimization, and Machine
Learning, New Delhi: Pearson Education (Singapore) Pte. Ltd., Indian Branch,
2002.
[50]. D. P. Kothari and J. S. Dhillon, Power System Optimization. New Delhi:
Prentice Hall of India, 2004.

147
[51]. P. Alves da Silva, “Fundamental of Genetic Algorithms”. In Tutorial on
Modern Heuristic Optimization Techniques with Application to Power
System, ed. by K.Y. Lee and M.A. El-Sharkawi, IEEE Power Engineering
Society (02TP160), p. 28, 2002.
[52]. D. Whitley, “The GENITUR Algorithm and selection pressure-Why Rank
based-based allocation of reproduction trials is best,” Proceedings of
International Conference on Genetic Algorithms, Schaffered, Morgan
Kaufmann Publishers, Los Altos, Cal, pp.10-19.
[53]. D. E. Goldsberg and K. Deb, “A comparative study of selection schemes used
in GA,” Foundations of Genetic Algorithms I. pp.53-69, 1991.
[54]. J. Kennedy and R. C. Eberhart, “Particle swarm optimization,” Proceedings of
the IEEE International Conference on Neural Networks, Perth, Australia,
1995, Volume IV, pp. 1942-1948.
[55]. R. C. Eberhart and J. Kennedy, “A new optimizer using particle swarm
theory,” Proceedings of the Sixth International Symposium on Micro Machine
and Human Science, Nagoya, Japan, 1995, pp. 39-43.
[56]. J. Kennedy, R. C. Eberhart, and S. Yuhui, Swarm Intelligence, Morgan
Kaufmann Publishers, USA, 2001.
[57]. S. Yuhui and R. C. Eberhart, “A modified particle swarm optimizer,”
Proceedings of the 1998 IEEE International Conference on Evolutionary
Computation, Anchorage, Alaska, May 1998, pp. 69-73.
[58]. S. Yuhui and R.C. Eberhart, “Parameter selection in particle swarm
optimization,” Proceedings of the 1998 Annual Conference on Evolutionary
Computation, March 1998.
[59]. C. Maurice, “The swarm and the queen: Towards a deterministic and adaptive
particle swarm optimization,” Proceedings of Congress on Evolutionary
computation, Washington DC, 1999, pp. 1951-1957.
[60]. C. Maurice, and J. Kennedy, “The particle swarm: Explosion, stability, and
convergence in a multi-dimensional complex space,” IEEE Transaction on
Evolutionary Computations, Volume 6, pp. 58-73, 2002.
[61]. F. Bergh, “An analysis of particle swarm optimizers,” Ph.D. Dissertation,
Department of Computer Science, University of Pretoria, Pretoria, South
Africa, 2002.

148
[62]. S. Yuhui, “Particle swarm optimization,” IEEE Connections: The Newsletter
of the IEEE Neural Network Society, Volume 2, No.1, pp. 8-13, 2004.
[63]. R. C. Eberhart and S. Yuhui, “Comparing inertial weights and constriction
factors in particle swarm optimization,” Proceedings of CEC 2000, San Diego,
CA, 2000, pp. 84 -88.
[64]. D. Corne, M. Dorigo, and F. Glove, New Ideas in Optimization, McGraw Hill,
1999.
[65]. E. A. Grimaldi, F. Grimaccia, M. Mussetta, and R. E. Zich, "PSO as an
effective learning algorithm for neural network applications," Proceedings of
3rd International Conference on Computational Electromagnetic and its
Applications, 2004, pp.557-560,.
[66]. W. Sun, Y. Zhang, and F. Taoli, “The neural network model based on PSO for
short term load forecasting,” Proceedings of Fifth International Conference on
Machine Learning and Cybernetics, Dalian, 13-16 August 2006, pp. 3069-
3072.
[67]. V. G. Gudise, and G. K. Venayagamoorthy, “Comparison of particle swarm
optimization and backpropagation as training algorithms for neural networks,”
Proceedings of the IEEE Swarm Intelligence Symposium, 2003, pp. 110-117.
[68]. W. F. Tinney, and D. I Sun, “Optimal Power Flow: Research and code
development,” Technical Report: Electric Power Research Institute (EPRI),
1987.
[69]. A. Ahmad, “Optimal Power Flow solution Using Second Order Newton
Method,” Ph.D Dissertation, Department of Electrical and Electronics,
UMIST, U.K., March 1991.
[70]. A. Aftab, and T. N. Malik, “Second Order Optimal Power Flow Solution—
Essential Background,” Pakistan Council for Science and Technology
international journal Science, Technology and Development, vol. 16, No. 3,
pp. 27-34, 1997.
[71]. Optimization Toolbox for Use with MATLAB. MathWorks..
[72]. S. G. Nash, Linear and Nonlinear Programming, New York: McGraw-Hill,
1996.
[73]. P. Attaviriyanupap, “The application of Evolutionary Programming based
optimization Method to Power system Generation before and after

149
 Deregulation,” Ph.D Dissertation, Hakhaido, University, Japan, December
2002.
[74]. R. L. Haupt and S. E. Haupt, Practical Genetic Algorithms, New Jersey: John
Wiley, 2004.
[75]. E. Gamma, R. Helm, J. Ralph, and V. John, Design Patterns Elements of
Reusable Object-Oriented Software, Addison Wesley, 1995.
[76]. F. Buschmann, R. Meunier, H. Rohnert, P. Summerlad, and M. Stal, Pattern-
Oriented Software Architecture, A System of Patterns, John Wiley & Sons
Ltd., 1999.
[77]. T. N. Malik, A. Q. Abbasi, and A. Ahmad, “Computational framework for
power economic dispatch using genetic algorithm,” Proceeding of the third
International Conference on Informatics in Control, Automation and Robotics
(ICINCO 2006),Stubal, Portugal, August 1-5, 2006, pp. 191-194.
[78]. Microsoft Developer’s Network (MSDNTM ) 2003, MS Corporation.
[79]. H. Saadat, Power System Analysis, Singapore: WCB/McGraw-Hill, 1999.
[80]. www.wikipedia.org
[81]. T. N. Malik, A. Ahmad, S. Kashnood, and A. Ahmad, “Economic dispatch
using genetic algorithm based hybrid approach,” Proceedings of 14th
International conference on Nuclear Engineering, Miami, Florida, USA, July,
17-20, 2005, pp. 1-6.
[82]. WAPDA in Brief, Public Relations Division, WAPDA, Lahore.
[83]. WAPDA Annual Report 2004, Public Relations Division, WAPDA, Lahore.
[84]. Power System Data Base, National Power Plan, WAPDA, Lahore.
[85]. Power Stations Book, Load Dispatch Centre, WAPDA, Lahore.
[86]. T. N. Malik, A. Ahmad, and A. Ahmad, “Potential of Artificial Intelligence
(AI) Techniques For WAPDA Grid System in Deregulated Environment”,
Proceeding of National Conference on Emerging Technologies (NCET 2004),
organized by SZABIST, ACM, and IEEE, Karachi, Pakistan, December, 18-
19, 2004, pp. 58-62.
[87]. A. E Eiben,., R. Hinterding, and Z. Michalewicz, “Parameter Control in
Evolutionary Algorithms,” IEEE Transactions on Evolutionary Computation,
Volume 3, No. 2, pp. 124-141, July, 1999.

150
[88]. K. Zollenkopf, “Bi-Factorization – Basic Computational Algorithms and
Programming Techniques,” Conference on solution of large sparse sets of
linear equations, Academic Press 1971.
[89]. T. N. Malik, B. A. Mirza, A. Ahmad, and A. Ahmad, “Genetic Algorithm In
Economic Dispatch --- Summary & Observations”, NEW HORIZONS,
Journal of the Institution of Electrical & Electronics Engineers, Pakistan, Vol.
No. 52, pp.8-13, April to June, 2006.
[90]. T. N. Malik, A. Ahmad, and A. Ahmad, “A Review of Advances in Economic
Dispatch Using Artificial Neural Networks,” Proceeding of the first
International Conference on Informatics in Control, Automation and Robotics
(ICINCO 2004), Stubal, Portugal, August 25-28, 2004, pp.354-357.
[91]. V. G. Gudise and G. K. Venayagamoorthy, “Comparison of Particle Swarm
Optimization and Backpropagation as Training Algorithms for Neural
Networks,” Proceedings of the IEEE Swarm Intelligence Symposium 2003,
pp.110-117.
[92]. J. B. Park, K.S. lee, J. R. Shin, K. Y. Lee, “ A particle swarm optimization for
economic dispatch with non-smooth cost functions,” IEEE transactions on
Power Systems, Vol. 20, No. 1, pp. 34-42, Feb. 2005.
[93]. H. T. Young, P. C. Young, C. L. Huang, “ Evolutionary programming based
economic dispatch for units with non-smooth fuel cost functions,” IEEE
transactions on Power Systems, Vol. 11, No. 1, pp. 112-118, February 1996.
[94]. Jong-Bae Park; Yun-Won Jeong; Woo-Nam Lee; Joong-Rin Shin, “An
improved particle swarm optimization for economic dispatch problems with
non-smooth cost functions,” IEEE Power Engineering Society General
Meeting, 2006, 18-22 June 2006, Page(s):7
[95]. A. J. Wood, and B. F. Wollenberg, Power Generation Operation & Control.
New York: John Wiley, 1996.
[96]. M. A. Pai., Computer Techniques in Power System Analysis. New Delhi: Tata
McGraw-Hill, 2006.
[97]. P. P. Narayana, and K. Latha, “Evolutionary programming based economic
power dispatch solutions with independent power producers,” IEEE
international conference on Electric Utility Deregulation, Restructuring and
Power Technologies (DRPT 2004), pp. 172 – 177, April 2004.

151
[98]. Loi Lei Lai, Intelligent System Applications in Power Engineering. New York:
John Wiley & Sons, 1998.
[99]. C. T. Su., “New approach with a Hopfield modeling framework to economic
dispatch,” IEEE Transaction on Power Systems, vol. 15, No.2, pp.541-545,
May 2000.
[100]. N. Sinha, R. Charabarti, and P. K. Chattopadhyay, “Evolutionary
programming techniques for economic load dispatch,” IEEE Transactions on
Evolutionary Computation, Vol. 7, No. 1, February, pp. 83-94, 2003.

152
 APPENDIX A

Test Systems --- Description & Data

A.1 Three Machines Test Systems

A.1.1 Test System I --- 3-Machine 6-Bus System with PD = 210 MW

This system has been selected from [95]. It consists of six buses and three
machines. The quadratic cost curves of three machines have been augmented by
incorporating the coefficients ei and fi reflecting valve point effects in order to convert the
quadratic convex curve into nonconvex cost curves as shown in Figure A.1. The
generator’s cost curves and B-coefficients are shown in shown Table A.1. The Loss
coefficients have been calculated using bus-bar and line data given in [95].

Generators Cost Characteristic Curves

M-1 M-2 M-3
3000
2500
Generation Cost

2000
1500
1000
500
0
37 57 77 97 117 137 157 177 197
Power Output

Figure A-1 Generators Cost Curve with Value-Point Effect

154
Table A-1 Test System I: 3-Machines 6 Bus Test System Generator’s Cost Curves &
B-Coefficients
Bus a b c e f Pmin Pmax Qmin Qmax
1 213.1 11.669 0.00533 130 0.0635 50 200 -100 100
2 200.0 10.333 0.00889 90 0.0598 37.5 150 -100 100
3 240.0 10.833 0.00741 100 0.0685 45 180 -100 60
B-Coefficients

⎡0.0287 0.0047 - 0.0035 ⎤

⎢ 0.0047 0.0665 0.0112 ⎥
[B] = ⎢ ⎥
⎢⎣ - 0.0035 0.0112 0.0371 ⎥⎦

[B0] = [ 0.0013 0.0080 0.0029]

B00 = 0.00055719

A.1.2 Test System II --- 3-Machine 6-Bus System with PD = 850 MW

This system has been taken from [95]. The generator’s cost curves and B-
coefficients are shown in shown Table A.2. The Loss coefficients have been calculated
using bus-bar and line data given in [95].

Table A-2 Test System II: 3-Machines 6 Bus Test System Generator’s Cost Curves
& B-Coefficients
Bus a b c e f Pmin Pmax Qmin Qmax
1 561 7.92 0.001562 300.0 0.0315 100 600 10 300
2 310 7.85 0.00194 200.0 0.042 100 400 10 200
3 78 7.97 0.00482 50.0 0.063 50 200 10 100
B-Coefficients

⎡0.0274 0.0002 - 0.0045 ⎤

⎢0.0002 0.0072 - 0.0034 ⎥
[B] = ⎢ ⎥
⎣⎢ - 0.0045 - 0.0034 0.0158 ⎦⎥

[B0] = [ - 0.0004469 - 0.0005744 - 0.0008301]

B00 = 0.00031232

155
A.1.3 Test System II --- 3-Machine 5-Bus System with PD = 150 MW
This system has been taken from [79]. The generator’s cost curves and B-
coefficients are shown in shown Table A.2.
Table A-3 Test System III: 3-Machines 5 Bus Test System Generator’s Cost Curves
& B-Coefficients
Bus a B c e f Pmin Pmax Qmin Qmax
1 200 7.0 0.008 -- -- 150 85 10 50
2 180 6.3 0.009 -- -- 150 80 10 50
3 140 6.8 0.007 -- -- 100 70 10 40
B-Coefficients

⎡ 0.0218 0.0093 0.0028 ⎤

⎢ ⎥
[B] = ⎢ 0.0093 0.0228 0.0017 ⎥
⎢⎣ 0.0028 0.0017 0.0179 ⎥⎦

[B0] = [ 0.0003 0.0031 0.0015]

B00 = 0.00030532

A.2 IEEE Test Systems

Following two IEEE test systems have been selected for analysis:
a. 5-Machines 14-Bus System with PD = 259 MW
b. 6-Machines 30-Bus System PD = 283.4 MW
Systems data has been taken from [96]. Two cases have been studied for these systems.
The generators cost coefficients of machines of IEEE 30-bus system for Case-I and Case
II have been taken from [97] and [98] respectively. The generators cost curves for IEEE
14-bus system have been taken the same as that of first five machines of IEEE 30-bus
systems in both cases. The generators curves of these systems for Case-I and Case-II are
shown in Table A.4 and A.5 respectively.
The B-coefficients for the systems have been calculated using bus-bar & line data given
in [96] and their values are shown in Table A.6.

156
 Table A-4 IEEE 5-Machines 14-Bus System Generator’s Cost Curves
Test Bus a b c e f Pmin Pmax Qmin Qmax
Systems
1 150.0 2.00 0.0016 50.0 0.0630 50 200 -40 100
2 25.0 2.50 0.0100 40.0 0.0980 20 80 -40 50
Case – I

3 0.0 1.00 0.0625 0.0 0.0 15 50 0 40

6 0.0 3.25 0.00834 0.0 0.0 10 35 -6 24
8 0.0 3.00 0.025 0.0 0.0 10 30 -6 24
Bus a b c e f Pmin Pmax Qmin Qmax
1 0.0 2.00 0.00375 15.0 0.06283 50 250 -40 100
2 0.0 1.75 0.0175 10.0 0.08976 20 160 -40 50
Case – II

3 0.0 1.00 0.0625 10.0 0.14784 15 100 0 40

6 0.0 3.25 0.00834 5.0 0.20944 10 70 -6 24
8 0.0 3.00 0.025 5.0 0.25133 10 60 -6 24

Table A-5 IEEE 6-Machines 30-Bus System Generator’s Cost Curves

Bus a b c e f Pmin Pmax Qmin Qmax
1 150.0 2.00 0.0016 50.0 0.0630 50 200 -40 200
2 25.0 2.50 0.0100 40.0 0.0980 20 80 -20 100
Case – I

5 0.0 1.00 0.0625 0.0 0.0 15 50 -15 80

8 0.0 3.25 0.00834 0.0 0.0 10 35 -15 60
11 0.0 3.00 0.025 0.0 0.0 10 30 -10 50
13 0.0 3.00 0.025 0.0 0.0 12 40 -15 60
Bus a b c e f Pmin Pmax Qmin Qmax
1 0.0 2.00 0.00375 15.0 0.06283 50 250 -40 200
2 0.0 1.75 0.0175 10.0 0.08976 20 160 -20 100
Case – II

5 0.0 1.00 0.0625 10.0 0.14784 15 100 -15 80

8 0.0 3.25 0.00834 5.0 0.20944 10 70 -15 60
11 0.0 3.00 0.025 5.0 0.25133 10 60 -10 50
13 0.0 3.00 0.025 5.0 0.1848 12 80 -15 60

157
 Table A-6 Transmission Loss Coefficients IEEE Test Systems

S. No Test System B-Coefficients

⎡0.0212 0.0085 0.0069 0.0002 0.0002 ⎤
⎢0.0085 0.0188 - 0.0062 0.00510.0002 ⎥⎥
⎢
IEEE [B] = ⎢0.0069 - 0.0062 0.4817 - 0.1333- 0.1604 ⎥
14-Bus ⎢ ⎥
1 ⎢0.0002 0.0051 - 0.1333 0.2180- 0.0251 ⎥
5-Machines ⎢0.0002 0.0002
⎣ - 0.1604 - 0.0251 0.1406 ⎥⎦
System
[B0] = [-0.0002 0.0027 - 0.0056 0.0207 - 0.0076]
B00 = 0.00085346
⎡0.0223 0.0109 0.0003 - 0.0011 0.0012 0.0013 ⎤
⎢0.0109 0.0139 0.0010 - 0.0019 0.0005 0.0008 ⎥⎥
⎢
IEEE ⎢0.0003 0.0010 0.0314 - 0.0155 - 0.0050 - 0.0207 ⎥
[B] = ⎢ ⎥
30-Bus ⎢- 0.0011 - 0.0019 - 0.0155 0.0298 0.0055 0.0114 ⎥
2 ⎢ 0.0012
6-Machines 0.0005 - 0.0050 0.0055 0.0113 0.0005 ⎥
⎢ ⎥
System ⎣⎢ 0.0013 0.0008 - 0.0207 0.0114 0.0005 0.1248 ⎦⎥
[B0] = [-0.0002 0.0003 - 0.0025 0.0019 0.0010 0.0012]
B00 = 0.0011

A.3 American Power Company (APC) Test System

This system has been taken from [79]. The quadratic cost curves of six machines
have been augmented by incorporating the coefficients ei and fi reflecting valve point
effects in order to convert the quadratic convex curve into nonconvex cost curves as
shown in Figure A.2. The generators data and loss coefficients are shown in Table A.9
and Table A.10 respectively.

Table A-7 6-Machine 26-Bus Test System Generator’s Cost Curves

Bus a b c e f Pmin Pmax Qmin Qmax

1 150.0 2.00 0.0016 50.0 0.0630 50 200 -40 100

2 25.0 2.50 0.0100 40.0 0.0980 20 80 -40 50

3 0.0 1.00 0.0625 0.0 0.0 15 50 0 40
6 0.0 3.25 0.00834 0.0 0.0 10 35 -6 24
8 0.0 3.00 0.025 0.0 0.0 10 30 -6 24

158
 Table A-8 Transmission Loss Coefficients APC 6-Machines 26 Bus Systems
B-Coefficients
⎡ 0.0014 0.0015 0.0009 - 0.0001 - 0.0004 - 0.0002 ⎤
⎢0.0015 0.0043 0.0050 0.0001 - 0.0008 - 0.0003 ⎥⎥
⎢
⎢0.0009 0.0050 0.0315 - 0.0000 - 0.0020 - 0.0016 ⎥
[B] = ⎢ ⎥
⎢- 0.0001 0.0001 - 0.0000 0.0029 - 0.0006 - 0.0009 ⎥
⎢- 0.0004 - 0.0008 - 0.0020 - 0.0006 0.0085 - 0.0001⎥
⎢ ⎥
⎢⎣- 0.0002 - 0.0003 - 0.0016 - 0.0009 - 0.0001 0.0176 ⎥⎦
[B0] = [ - 0.0002 - 0.0008 0.0067 0.0001 0.0000 - 0.0012]
B00 = 0.0056

A.4 20- Machines System

The system has been taken from [99]. This is 20-Machine system with quadratic
cost curves and load demand of 2500 MW. The cost curves are given here. B-
coefficients can be referred in reference [99].
Table A-9 20-Machine Test System Generator’s Cost Curves
Bus a b c Pmin Pmax
1 1000 18.19 0.00068 150 600
2 970 19.26 0.00071 50 200
3
600 19.8 0.0065 50 200
4
700 19.1 0.005 50 200
5
420 18.1 0.00738 50 160
6
360 19.26 0.00612 20 100
7
490 17.14 0.0079 25 125
8
660 18.92 0.00813 50 150
9
765 18.27 0.00522 50 200
10
770 18.92 0.00573 30 150
11
800 16.69 0.0048 100 300
12
970 16.76 0.0031 150 500
13
900 17.36 0.0085 40 160
14
700 18.7 0.00511 20 130
15
450 18.7 0.00398 25 185
16
370 14.26 0.0712 20 80
17
480 19.14 0.0089 30 85
18
680 18.92 0.00713 30 120
19
700 18.47 0.00622 40 120
20
850 19.79 0.00773 30 100

159
A.5 13-Machines System with Nonconvex Cost Curve [100]
Table A-10 13-Machines Test System Generator’s Cost Curves
Bus a b c e f Pimin Pimax
1 550 8.1 0.00028 300 0.035 0 680
2 309 8.1 0.00056 200 0.042 0 360
3 307 8.1 0.00056 150 0.042 0 360
4 240 7.74 0.00324 150 0.063 60 180
5 240 7.74 0.00324 150 0.063 60 180
6 240 7.74 0.00324 150 0.063 60 180
7 240 7.74 0.00324 150 0.063 60 180
8 240 7.74 0.00324 150 0.063 60 180
9 240 7.74 0.00324 150 0.063 60 180
10 126 8.6 0.00284 100 0.084 40 120
11 126 8.6 0.00284 100 0.084 40 120
12 126 8.6 0.00284 100 0.084 55 120
13 126 8.6 0.00284 100 0.084 55 120

A.6 Pakistani Utility System

Table A-11 12-Machines 32-Bus NTDC Test Systems Generator’s Cost Curves
Bus a b c e f Pmin Pmax Qmin Qmax
1 117965.5955 3871.2446 3.3829 -- -- 170 210 -10 100
2 38044.3947 3930.3768 35.7800 -- -- 25 65 -20 30
3 21835.2917 1300.7907 2.0633 -- -- 230 300 -60 210
4 121270.2094 1364.9392 0.5492 -- -- 40 100 -40 70
5 27983.7444 1143.7230 10.6067 -- -- 40 120 -40 80
6 117845.1040 3073.0320 3.0347 -- -- 125 250 40 120
7 21927.9678 1631.1822 3.2305 -- -- 70 110 -10 80
8 88181.4304 964.2456 5.0629 -- -- 25 100 -30 50
9 59533.4194 1920.6626 1.6412 -- -- 100 210 40 100
10 65184.4252 1701.8310 1.8056 -- -- 110 210 40 100
11 42799.8444 2212.5550 0.0078 -- -- 240 320 -60 220
12 38044.3947 3930.3768 35.7800 -- -- 25 65 -20 30

160
 Table A-12 Transmission Loss Coefficients 12 Machines 32-Bus NTDC System
B-Coefficients
[B] =
⎡0.0072 0.0007 - 0.0018 0.0052 - 0.0012 - 0.0027 - 0.0023 - 0.0006 - 0.0025 0.0009 0.0070 0.0
⎢ 0.0007 0.0187 - 0.0028 0.0071 - 0.0022 - 0.0040 - 0.0032 - 0.0013 - 0.0035 0.0031 0.0007 0
⎢
⎢ - 0.0018 - 0.0028 0.0039 0.0005 0.0008 0.0042 - 0.0002 - 0.0003 0.0034 - 0.0019 - 0.0018 -
⎢
⎢ 0.0052 0.0071 0.0005 0.7737 - 0.0015 - 0.0299 0.0013 - 0.0008 - 0.0042 - 0.0167 0.0036 0
⎢ - 0.0012 - 0.0022 0.0008 - 0.0015 0.0028 0.0013 0.0008 0.0004 0.0003 - 0.0007 - 0.0012 - 0
⎢
⎢ - 0.0027 - 0.0040 0.0042 - 0.0299 0.0013 0.0097 - 0.0002 - 0.0003 0.0045 - 0.0001 - 0.0025 -
⎢ - 0.0023 - 0.0032 - 0.0002 0.0013 0.0008 - 0.0002 0.0166 - 0.0006 - 0.0007 - 0.0013 - 0.0022
⎢
⎢ - 0.0006 - 0.0013 - 0.0003 - 0.0008 0.0004 - 0.0003 - 0.0006 0.0017 - 0.0008 - 0.0005 - 0.0006
⎢
⎢ - 0.0025 - 0.0035 0.0034 - 0.0042 0.0003 0.0045 - 0.0007 - 0.0008 0.0081 - 0.0019 - 0.0024 -
⎢ 0.0009 0.0031 - 0.0019 - 0.0167 - 0.0007 - 0.0001 - 0.0013 - 0.0005 - 0.0019 0.0090 0.0011
⎢
⎢ 0.0070 0.0007 - 0.0018 0.0036 - 0.0012 - 0.0025 - 0.0022 - 0.0006 - 0.0024 0.0011 0.0068 0
⎢ 0.0005 0.0026 - 0.0013 0.0066 - 0.0006 - 0.0022 - 0.0017 0.0001 - 0.0020 0.0016 0.0005 0
⎣

[B0] =
[- 0.0006 - 0.0029 - 0.0002 - 0.0306 0.0014 0.0053 - 0.0000 0.0001 0.0005 0.0035 - 0.0004 - 0.0

B00 = 0.0030

161
 APPENDIX B

Quadratic Optimization

B.1 Minimizing Quadratic Functions

The stationary point of a function of one variable can be found by setting the
expression for first derivative of the function equal to zero and solving for its roots. Each
root is a stationary point which could be a minimum, maximum or an inflection point of
the function. A stationary point is minima if the second derivative of the function
evaluated at that point is positive. We start the development here by finding the minimum
of the general quadratic function:
1
f(y) = a + by + G y2 (B.1)
2
There is only one stationary point which can be found by setting the first derivative equal
to zero and solving the resulting linear equation for the optimum value of y. The quadratic
case procedure can be extended to find stationary points for higher order functions. The
major difference is that now the first derivative of the function will no longer be linear
and it will be necessary to use the iterative methods to find its roots. Newton's root
finding method among others can be used to find the roots of the non linear functions.
If function f(y) is of higher order than quadratic, the Newton's root finding
procedure will be as follows:
2 k k
d f( y ) df( y )
2
∆ y y +1 = -
dy dy (B.2)
k +1 k k +1
y = y +∆ y
where k is the iteration count. The procedure starts with estimating y as an approximation
to its minimum y*, then the first and second derivatives of f(y) are evaluated and solved
for ∆yk+1 in equation (B.2). Minimum will be verified if the following conditions are
satisfied.
df( y k + 1 )
=0 (B.3)
dy

162
 2 k +1
d f( y )
2
f0 (B.4)
d y
If equation (B.3) is not satisfied, the process is repeated increasing k by 1. If equations
(B.3) and (B.4) are both satisfied, yk+1 = y* is a true minimum. If equation (B.3) is
satisfied but (B.4) is not, there are two possibilities: 1) if the second derivative is
negative, yk+1 is a maximum point; or 2) if it is zero, yk+1 is an inflection point.

B.1.1 Minimizing Multivariable Functions

A quadratic function of more then one variable can be expressed in matrix form
as:
1 T
f(y) = a + bT y + y Gy (B.5)
2
where f(y) is a scalar function, y is an N Vector of variables, G is an (N * N) symmtric
matrix, a is a scalar. The procedure to solve for the minimum of the function (B.5) is
almost similar to that of the single variable functions except that now the partial
derivatives and matrix notation replace the simple derivatives and simple variables
notation as used in the single variable case:
∂f(y)
= ∇ f(y)= G y + b (B.6)
∂y
2
∂ f (y)
= ∇2 f(y) = G (B.7)
∂ y2
Equation (B.6) expresses the gradient of f(y). It is a vector of functions of y. Each element
of ∇f(y) is of the form ðf(y)/ðyi. Equation (B.7) shows that G, the Hessian matrix of f(y), is
a matrix of second partial derivatives of f(y). G can also be viewed as a matrix of first
partial derivatives of the elements of ∇f(y). In the quadratic case G(y) = G is a constant.
G(y) is always symmetric because the result of the partial differentiation is independent of
the order in which the derivatives are computed. A quadratic is always defined with
symmetric G. For a minimum of the quadratic to exist, G must be positive definite. For G
to be positive definite, its quadratic form yTGy must be positive for all values of y. When
G is positive definite, it can be visualized as defining an N dimensional bowl shaped
surface that opens upward. The minimum y* of the quadratic f(y) can be found directly by
solving the non singular system of linear equations obtained by setting its gradient equal
to zero.

163
 Newton's method can be used for finding the stationary points for higher order
multivariable functions. The main difference is that ∇f(y) will no longer be a vector of
linear function and ∇2f(y) = G(y) will no longer be a constant matrix. The equations for
iterative solution at iteration K for quadratic or higher order multivariable functions are:
k k +1 k
G( y ) ∆ y = - ∇f( y ) (B.8)
k +1 k k +1
y = y +∆ y (B.9)
The algorithm for determining y, the minimum of multivariable nonlinear function f(y),
by Newton's method is as follows:
1. Initialize the iteration count K = 0
2. Estimate y* , Let the estimate be yK
3. Evaluate ∇f(yK) and G(yK)
4. Test for the minimum
. If ∇f(yK) does not equal to zero, Go to Step(6)
. Else Continue
5. If ∇f(yK) = 0, and G(yk) is positive definite ; yK = y*; Normal Exit
ELSE
If ∇f(yK) = 0 and G(yK) is not positive definite, y is a saddle point; Error
Exit.
6. Solve for ∆yK+1 in Equation (B.8)
7. Update yK+1 = yK + ∆yK+1 and K = K + 1
8. Go to step(3)

164
 APPENDIX C

Output Results --- Chapter 5 & 6

Table C-1 Evolution Models & GA Parameters for Convex Economic Dispatch 3
Machines Test System Using λ – Search & RP Search
λ –Search RP-Search
3 Machines
Test System Test System Test System
Test Systems
I II III
I-II-III
Encoding
Random Random Random Random
Scheme
Selection Tournament Tournament Tournament Tournament
Model
GA

Crossover Inversion Inversion Inversion Inversion

Mutation Bit Flip Bit Flip Bit Flip Bit Flip

Population
16 20 20 20
size
Parameters

Chromosome
20 53 56 50
Length
GA

Crossover
0.8 0.75 0.75 0.75
Probability
Mutation
0.01 0.0001 0.0001 0.0001
Probability

Table C-2 Output Results Convex ED Test System–I: 3-Machine System

PD = 210 MW

GA Results After 20 Runs

Corresponding to Minimum Cost
Unit Output (MW) λ-Iteration
λ-Search RP-Search

P1 72.6377 71.1724 71.2569

P2 63.8543 65.8269 93.5659
P3 82.3439 81.9824 56.2123
Ptotal 218.8359 218.9817 221.0351
Gen Cost ($/h) 3167.1620 3167.2395 3188.6666
PL 8.8362 8.9823 11.0357
Tol 0.0001 0.0007 0.0006

165
 Iter. or Gen. 7 523 7616
Time 0.015 0.7790 4.0620
λ 13.0056 12.9631 ---
Standard Deviation --- 0.005859 0.437044

Table C-3 Output Results Convex ED Test System – II: 3-Machine System
PD = 850 MW
GA Results After 20 Runs
Unit Output (MW) λ-Iteration Corresponding to Minimum Cost
λ-Search RP-Search
P1 284.0452 278.6098 350.9856
P2 400.0000 400 354.0673
P3 195.2698 200 182.8481
Ptotal 879.3151 878.6098 887.9011
Gen Cost ($/h) 8515.1515 8514.037 8562.3129
PL 29.3151 28.6106 37.9010
Tol 0.0001 0.0008 0.0001
Iter. or Gen. 10 90 11598
Time 0.015 0.125 10.5810
λ 10.2130 10.163981 ---
Standard Deviation --- 0.004575 143.2415

Table C-4 Output Results Convex ED Test System – III: 3-Machine System
PD = 150 MW
GA Results After 20 Runs
Unit Output (MW) λ-Iteration Corresponding to Minimum Cost
λ-Search RP-Search
P1 33.4701 32.9167 44.4779
P2 64.0974 64.6903 48.9721
P3 55.1011 55.0682 59.0886
Ptotal 152.6687 152.6752 152.5385
Gen Cost ($/h) 1599.9834 1599.9887 1603.5225
PL 2.6687 2.6754 2.5381
Tol 0.0001 0.0002 0.0005
Iter. or Gen. 8 42 102
Time 0.015 0.062 0.0470
λ 7.7678 7.755193 ---
Standard Deviation --- 0.002912 0.571504

166
Table C-5 Evolution Models for Convex Economic Dispatch for IEEE Test Systems
using λ - Search
IEEE IEEE
Test Systems Test Systems
Case-I Case-II
Encoding
Random Random
Scheme
Selection Tournament Tournament
Model
GA

Crossover Inversion Inversion

Mutation Bit Flip Bit Flip

Population size 16 16
Parameters

Chromosome
20 20
Length
GA

Crossover
0.8 0.8
Probability
Mutation
0.01 0.01
Probability

Table C-6 Evolution Models for Convex Economic Dispatch for IEEE Test Systems
using Real Power - Search
IEEE Test System

Case - I Case - II
5 Machines 6 Machines 5 Machines 6 Machines
Encoding
Random Random Random Random
Scheme
Selection Tournament Tournament Tournament Tournament
Model
GA

Crossover Inversion Inversion 5-Point 5-Point

Mutation Bit Flip Bit Flip Bit Flip Bit Flip

Population
20 20 20 20
size
Parameters

Chromosome
80 95 85 102
Length
GA

Crossover
0.75 0.75 0.8 0.75
Probability
Mutation
0.0001 0.0001 0.0001 0.0001
Probability

167
Table C-7 Output Results Convex ED Case – I: IEEE 5-Machines 14 Bus Test
Systems PD = 259 MW
GA Results After 20 Runs
Unit Output (MW) λ-Iteration Corresponding to Minimum Cost
λ-Search RP-Search
P1 200.00000 200.0000 138.6709
P2 35.00031 34.4052 48.6006
P3 15.00000 15.6323 25.9793
P4 10.00000 10.0000 31.5377
P5 10.00000 10.0000 22.6438
Ptotal 270.0003 270.0376 267.4323
Gen Cost ($/h) 833.64750 833.5897 887.9356
PL 11.0003 11.0385 8.4322
Tol 0.0001 0.0010 0.0000
Iter. or Gen. 11 317 1459
Time 0.015 1.6880 2.1250
λ 3.25158 3.3997 ---
Standard Deviation --- 0.001732 21.0689

Table C-8 Output Results Convex ED Case – II: IEEE 5-Machines 14 Bus Test
Systems PD = 259 MW
GA Results After 20 Runs
Unit Output (MW) λ-Iteration Corresponding to Minimum Cost
λ-Search RP-Search
P1 173.9697 172.7829 137.6270
P2 48.0007 47.9807 63.1622
P3 17.9537 17.8401 29.5596
P4 14.2848 16.1916 22.2847
P5 14.4150 13.8079 15.7290
Ptotal 268.6239 268.6032 268.3624
Gen Cost ($/h) 720.4242 720.5030 740.7419
PL 9.62400 9.6038 9.3620
Tol 0.0001 0.0007 0.0004
Iter. or Gen. 5 1039 138
Time 0.015 4.2970 0.1870
λ 3.60908 3.5981 ---

Standard Deviation --- 0.00198 91.5675

168
Table C-9 Output Results Convex ED Case – I: IEEE 6-Machines 30 Bus Test
Systems PD = 283.3 MW
GA Results After 20 Runs
Unit Output (MW) λ-Iteration Corresponding to Minimum Cost
λ-Search RP-Search
P1 200.00000 200.0000 147.3587
P2 43.71206 36.7256 48.3067
P3 19.06197 19.9008 31.4311
P4 10.00000 15.8706 25.2707
P5 10.00000 10.0000 20.5307
P6 12.00000 12.0000 17.9936
Ptotal 294.77403 294.4971 290.8915
Gen Cost ($/h) 914.59347 914.7356 963.3991
PL 11.47400 11.0980 7.4921
Tol 0.0001 0.0009 0.0006
Iter. or Gen. 11 68 948
Time 0.015 0.5780 1.6870
λ 3.41811 3.50778 ---
Standard Deviation --- 0.0016 25.2307

Table C-10 Output Results Convex ED Case – II: IEEE 6-Machines 30 Bus Test
Systems PD = 283.3 MW
GA Results After 20 Runs
Unit Output (MW) λ-Iteration Corresponding to Minimum Cost
λ-Search RP-Search
P1 176.14258 176.0737 91.3774
P2 48.73869 48.7867 67.1836
P3 21.03060 21.0973 34.3967
P4 22.69957 22.6603 35.4713
P5 12.25559 12.3456 37.3640
P6 12.00000 12.0000 22.7387
Ptotal 292.86703 292.9636 288.5318
Gen Cost ($/h) 802.36290 802.7253 872.8807
PL 9.56702 9.5639 5.1315
Tol 0.0001 0.0003 0.0003
Iter. or Gen. 12 576 3813
Time 0.015 3.8880 4.2500
λ 3.64514 3.6464 ---
Standard Deviation --- 0.0017 53.1444

169
Table C-11 Evolution Models for Convex Economic Dispatch for American Power
Company 6 – Machines 26 Bus Test System
American Power Company
λ –Search 6 machine 26 Bus System
RP-Search
Encoding
Random Random
Scheme
Selection Tournament Tournament
GA
Model
Crossover Inversion Uniform Crossover with Mask

Mutation Bit Flip Bit Flip

Population size 20 20
Chromosome
20 107
GA Length
Parameters Crossover
0.8 0.7
Probability
Mutation
0.01 0.0001
Probability

Table C-12 Output Results Convex ED APC 6-Machines 26 Bus Test Systems
PD = 1263 MW
Unit Output (MW) GA Results during 20 Runs
λ-Iteration Corresponding to Minimum Cost
λ-Search RP-Search
P1 474.06763 472.7851 437.4999
P2 173.86752 171.9561 199.9954
P3 191.04997 193.9917 183.9772
P4 150.00000 150.0000 149.9939
P5 196.69140 197.0791 197.6379
P6 103.51813 103.7049 119.9957
Ptotal 1289.19465 1289.5168 1289.1001
Gen Cost ($/h) 15689.78647 15689.9361 15708.7556
PL 26.19466 26.5161 26.1004
Tol 0.00010 0.0008 0.0003
Iter. or Gen. 11 77 255
Time 1 0.6250 0.4840
λ 13.91158 13.8910 ----
0.0090 66.0514
Standard Deviation

170
 Table C-13 Evolution Model & GA Parameters --- Convex ED Using Hybrid
Approach
λ –Search & Hybrid Approach
Encoding
Random
Scheme
Selection Roulette Wheel
GA
Model
Crossover One Point

Mutation Bit Flip

Population size 20
Chromosome
16
Length
GA Parameters
Crossover
0.8
Probability
Mutation
0.05
Probability

Table C-14 Output Results Convex ED: Hybrid Approach 3-Machine System
λ-Iteration GA Hybrid
λ –Search Approach
P1 72.6378 70.7050 72.6378
P2 63.8544 65.6489 63.8544
P3 82.3440 81.6538 82.3440
Ptotal 218.836 218.0077 218.8362
Gen Cost ($/h) 3167.1644 3155.4267 3167.1644
PL 8.8362 8.0073 8.8362
Tol 0.0001 0.001 0.0001
Iter. or Gen. 08 6496 06
λ 13.00558 12.95436 13.00558

171
 Table C-15 Output Results Convex ED: Hybrid Approach 20-Machine System

λ-Iteration GA λ-Search Hybrid

P1 511.5277 511.5318 511.5277
P2 169.9600 169.9628 169.9600
P3 126.4621 126.4638 126.4621
P4 102.5110 102.5122 102.5110
P5 113.5521 113.5529 113.5521
P6 73.0341 73.0350 73.0341
P7 114.9055 114.9059 114.9055
P8 116.1557 116.1569 116.1557
P9 99.9913 99.9922 99.9913
P10 105.2690 105.2698 105.2690
P11 150.4304 150.4307 150.4304
P12 292.3936 292.3944 292.3936
P13 119.2701 119.2706 119.2701
P14 30.8135 30.8136 30.8135
P15 115.2019 115.2029 115.2019
P16 36.2817 36.2817 36.2817
P17 66.2426 66.2436 66.2426
P18 92.8833 92.8843 92.8833
P19 100.6586 100.6595 100.6586
P20 54.1729 54.1738 54.1729
Ptotal 2591.7169 2591.7383 2591.7169
Gen.
62453.3662 62453.7890 62453.3660
Cost ($/h)
PL 91.7170 91.7182 91.7170
Tolerance 0.0001 0.1 0.0001
Iter. or Gen. 16 909 13
λ 20.9539 20.9540 20.9539

Table C-16 Evolution Models for Convex Economic Dispatch for NTDC Test System
Using λ - Search
WAPDA WAPDA
12 & 15 25 & 34 Machine
Machine Systems
Systems
Encoding
Random Random
Scheme
Selection Tournament Tournament
Model
GA

Crossover Inversion Inversion

Mutation Bit Flip Bit Flip

172
 Population size 20 20

Parameters
Chromosome
20 30
Length

GA
Crossover
0.8 0.8
Probability
Mutation
0.01 0.01
Probability

Table C-17 Evolution Models for Convex Economic Dispatch for NTDC Test System
using Real Power – Search
WAPDA Test System

12 Machines 15 Machines 25 Machines 34 Machines

Encoding
Random Random Random Random
Scheme
Selection Tournament Tournament Tournament Tournament
Model
GA

Crossover Inversion Inversion 5-Point 5-Point

Mutation Bit Flip Bit Flip Bit Flip Bit Flip

Population
20 20 20 20
size
Parameters

Chromosome
199 247 384 502
Length
GA

Crossover
0.75 0.75 0.8 0.75
Probability
Mutation
0.0001 0.0001 0.01 0.0001
Probability

Table C-18 Output Results Convex ED 12–Machine 32–Bus NTDC System

PD = 1600 MW
GA Results After 20 Runs Corresponding to
Unit Output Minimum Cost
λ-Iteration
(MW)
λ-Search RP-Search
P1 110.0000 110.0000 95.1377
P2 100.0000 100.0000 71.5011
P3 170.0000 170.0000 173.2624
P4 25.0000 25.0000 36.9008
P5 25.0000 25.0000 34.3086
P6 249.6881 293.9320 251.1264
P7 100.0000 100.0000 96.2618
P8 61.5060 61.0584 60.3084

173
 P9 125.0000 125.0000 126.2570
P10 179.7818 150.0997 159.9279
P11 150.2432 136.5035 200.3177
P12 320.0000 320.0000 314.3273
Ptotal 1616.2191 1616.5936 1619.6372
Gen Cost ($/h) 4659491.9674 4664527.9400 4785261.2186
PL 16.2190 16.6014 19.6363
Tol 0.0001 0.0078 0.0008
Iter. or Gen. 19 47 141
λ 2441.8673 2415.4257 ---
Time 0.062 1.3440 0.3600
Standard
--- 1.3123 114251.7916
Deviation

Table C-19 Output Results Convex ED 15 – Machine NTDC System PD = 2400 MW

Unit Output GA Results After 20 Runs Corresponding to
(MW) λ-Iteration Minimum Cost
λ-Search RP-Search
P1 75.00000 75 53.3403
P2 75.00000 75 71.6788
P3 140.00000 140 139.3429
P4 140.00000 140 114.2253
P5 212.11338 212.1133 211.4241
P6 212.11338 212.1133 184.2124
P7 326.76976 326.7696 291.1291
P8 131.33449 131.3344 150.1375
P9 131.33449 131.3344 149.2716
P10 131.33449 131.3344 184.0354
P11 240.00000 240 295.6841
P12 180.00000 180 145.2863
P13 180.00000 180 147.3976
P14 125.00000 125 128.4455
P15 100.00000 100 134.3899
Ptotal 2400. 2399.999 2400.0007
Gen Cost ($/h) 5512088.5870 5512087.0870 5546108.2504
PL 0.0000 0.0000 0.0000
Tol 0.0001 0.0007 0.0007
Iter. or Gen. 3 381 647
λ 2176.0978 2176.0976 ---
Time 0.031 13.907 2.4690
Standard
--- 0.00000 60195.1588
Deviation

174
 Table C-20 Output Results Convex ED 25–Machine NTDC System PD = 2800 MW
Unit Output GA Results After 20 Runs Corresponding to
(MW) λ-Iteration Minimum Cost
λ-Search RP-Search
P1 75.0000 75.0000 70.8510
P2 75.0000 75.0000 54.1223
P3 140.0000 140.0000 137.4185
P4 140.0000 140.0000 110.6400
P5 215.4574 215.4574 204.8835
P6 215.4574 215.4574 232.7892
P7 331.6179 331.6180 301.9282
P8 135.1558 135.1558 145.2121
P9 135.1558 135.1558 113.1936
P10 135.1558 135.1558 128.1873
P11 240.0000 240.0000 275.9734
P12 180.0000 180.0000 140.4056
P13 180.0000 180.0000 168.0101
P14 125.0000 125.0000 126.6381
P15 100.0000 100.0000 167.7477
P16 100.0000 100.0000 103.8904
P17 100.0000 100.0000 107.9493
P18 30.0000 30.0000 41.2284
P19 30.0000 30.0000 48.3990
P20 15.0000 15.0000 16.8090
P21 15.0000 15.0000 18.5722
P22 15.0000 15.0000 18.3631
P23 15.0000 15.0000 16.2816
P24 15.0000 15.0000 15.2227
P25 42.0000 42.0000 35.2846
PTotal 2800 2800.0002 2800.0010
Gen Cost ($/h) 6654911.2924 6654911.7586 6712903.1928
PL 0.0000 0.0000 0.0000
Tol 0.0001 0.0002 0.0010
Iter. or Gen. 3 42 46
λ 2189.8972 2189.8973 -----
Time 0.078 21.0150 0.3130
Standard
--- 14.7640 81079.3851
Deviation

175
Table C-21 Output Results Convex ED 34–Machine NTDC System PD = 3000 MW
Unit Output GA Results After 20 Runs Corresponding
(MW) λ-Iteration to Minimum Cost
λ-Search RP-Search
P1 75.0000 75.0000 56.4287
P2 75.00000 75.0000 66.6097
P3 140.00000 140.0000 117.9322
P4 140.00000 140.0000 35.5004
P5 216.62037 216.6193 219.2273
P6 216.62037 216.6193 203.7968
P7 333.30410 333.3025 389.3330
P8 136.48481 136.4836 133.6517
P9 136.48481 136.4836 146.4309
P10 136.48481 136.4836 129.5730
P11 240.00000 240.0000 263.3811
P12 180.00000 180.0000 141.7148
P13 180.00000 180.0000 169.5106
P14 125.00000 125.0000 141.5699
P15 100.00000 100.0000 107.6160
P16 100.00000 100.0000 144.4168
P17 100.00000 100.0000 112.0521
P18 30.00000 30.0000 47.0605
P19 30.00000 30.0000 48.9245
P20 15.00000 15.0000 18.3827
P21 15.00000 15.0000 15.2579
P22 15.00000 15.0000 17.3863
P23 15.00000 15.0000 16.0383
P24 15.00000 15.0000 17.6676
P25 42.00000 42.0000 39.2849
P26 7.00000 7.0000 9.5971
P27 7.00000 7.0000 7.8073
P28 40.00000 40.0000 35.8326
P29 28.00000 28.0000 21.7561
P30 28.00000 28.0000 27.8642
P31 28.00000 28.0000 26.5535
P32 18.00000 18.0000 23.7726
P33 18.00000 18.0000 26.5948
P34 18.00000 18.0000 21.4734
PTotal 2999.9993 2999.9918 2999.9994
Gen Cost ($/h) 7073116.03408 7073164.5506 7225943.9370
PL 0.0000 0.0000 0.0000
Tol 0.001 0.0082 0.0006
Iter. or Gen. 69394 14 91
λ 2194.69650 2194.6920 ----
Time 16.3440 0.9060
Standard
--- 12.2968 38343.8
Deviation

176
 APPENDIX D

Derived Publications

[1]. T. N. Malik, A. Ahmad, and A. Ahmad, “A review of advances in economic

dispatch using artificial neural networks,” Proceeding of the first International
Conference on Informatics in Control, Automation and Robotics (ICINCO
2004) , Stubal, Portugal, August 25-28, 2004, pp.354-357.
[2]. T. N. Malik, A. Ahmad, and A. Ahmad, “Potential of artificial intelligence
(AI) techniques for WAPDA grid system in deregulated environment,”
Proceeding of National Conference on Emerging Technologies (NCET 2004),
organized by SZABIST, ACM, and IEEE, Karachi, Pakistan, December, 18-
19, 2004, pp.58-62.
[3]. T. N. Malik, A. Ahmad, and A. Ahmad, and I. Ahmad, “Power system
economic dispatch: -- An overview & analysis,” Technical Journal of U.E.T.,
Taxila, pp. 96-108, 2006.
[4]. T. N. Malik, A. Q. Abbasi, and A. Ahmad, “Computational framework for
power economic dispatch using genetic algorithm,” Proceeding of the third
International Conference on Informatics in Control, Automation and Robotics
(ICINCO 2006) , Stubal, Portugal, August 1-5, 2006, pp.191-194.
[5]. T. N. Malik, A. Ahmad, S. Kashnood, and A. Ahmad, “Economic dispatch
using genetic algorithm based hybrid approach,” Proceedings of International
Conference on Nuclear Engineering (ICONE 14), Miami, Florida, USA, July
17-20, 2006, pp. 1-7.
[6]. T. N. Malik, B. A. Mirza, A. Ahmad, and A. Ahmad, “Genetic algorithm in
economic dispatch --- Summary & Observations”, NEW HORIZONS, Journal
of the Institution of Electrical & Electronics Engineers, Pakistan, vol. No. 52,
April to June, 2006, pp.8-13.
[7]. T. N. Malik, A. Asar. Neuro-Genetic hybrid approach for the solution of
nonconvex economic dispatch problem. Mehran University Research Journal
of Engineering & Technology. Accepted for publication August 2008.

177
[8]. T. N. Malik, and A. Asar. A new hybrid approach for the solution of
nonconvex economic dispatch problem with valve-point effects. Elsevier
Journal Electric Power Systems Research. Submitted 2009.

178