Application of Multi-Objective Optimisation To MRT Systems: Founded 1905

Founded 1905
APPLICATION OF MULTI-OBJECTIVE
OPTIMISATION TO MRT SYSTEMS
BY
TIAN LIFEN
DEPARTMENT OF ELECTRICAL ENGINEERING
A THESIS SUBMITTED
IN PARTIAL FULFILLMENT OF THE REQUIREMENT FOR
THE DEGREE OF MASTER OF ENGINEERING
NATIONAL UNIVERSITY OF SINGAPORE

2004
Acknowledgements
Acknowledgements
The author genuinely appreciates the help from her supervisor, Professor C. S. Chang,
who has provided invaluable guidance to the author while the author works on the
projects and writes this thesis.
Sincere thanks and gratitude are also extended to Mr. Seow Hung Cheng of Power
System laboratory for his support throughout this research project.
This thesis is dedicated to the author’s family and friends. They always have the
unreserved understanding and great support to the author.
I
Table of Contents
Table of Contents
ACKNOWLEDGEMENTS .....................................................................................................................I
TABLE OF CONTENTS ....................................................................................................................... II
LIST OF FIGURES ................................................................................................................................ V
LIST OF TABLES ................................................................................................................................. VI
LIST OF ABBREVIATIONS ............................................................................................................. VII
LIST OF PUBLICATIONS RELATED TO THIS THESIS...........................................................VIII
SUMMARY OF THE THESIS............................................................................................................. IX
CHAPTER 1 INTRODUCTION ............................................................................................................ 1
1.1 BACKGROUND ................................................................................................................................. 1
1.2 MOTIVATION OF THE RESEARCH...................................................................................................... 2
1.2.1 Go-Circuit Optimisation ......................................................................................................... 3
1.2.2 Return-Circuit Optimisation ................................................................................................... 4
1.3 OBJECTIVE AND SCOPE OF THE RESEARCH ...................................................................................... 5
1.4 MULTI-OBJECTIVE OPTIMISATION ALGORITHMS............................................................................. 6
1.5 ORGANISATION OF THE THESIS ........................................................................................................ 8
CHAPTER 2 OUTLINE OF GO-CIRCUIT OPTIMISATION ........................................................ 11
2.1 INTRODUCTION .............................................................................................................................. 11
2.2 MATHEMATICAL MODEL ............................................................................................................... 12
2.2.1 System Model ........................................................................................................................ 12
2.2.2 Objective Functions............................................................................................................... 13
2.2.3 Impact of Operational Timetable .......................................................................................... 15
2.3 LAYOUT OF THREE-STAGE SCHEME .............................................................................................. 16
2.4 SIMULATION OUTLINE ................................................................................................................... 18
2.5 SUMMARY ..................................................................................................................................... 19
CHAPTER 3 OUTLINE OF RETURN-CIRCUIT OPTIMISATION ............................................. 21
3.1 INTRODUCTION .............................................................................................................................. 21
3.2 MATHEMATICAL MODEL ............................................................................................................... 22
3.2.1 Return-Circuit Model ............................................................................................................ 22
3.2.2 Objective Functions............................................................................................................... 23
3.2.3 Impact of Earthing and Bonding Arrangements.................................................................... 24
3.3 LAYOUT OF TWO-STAGE SCHEME ................................................................................................. 25
3.4 SIMULATION OUTLINE ................................................................................................................... 27
3.5 SUMMARY ..................................................................................................................................... 28
II
Table of Contents
CHAPTER 4 LITERATURE REVIEW OF MULTI-OBJECTIVE OPTIMISATION

APPROACHES...................................................................................................................................... 30
4.1 INTRODUCTION .............................................................................................................................. 30
4.2 MATHEMATICAL DEFINITION......................................................................................................... 30
4.3 PREFERENCE STRUCTURE .............................................................................................................. 32
4.4 REVIEW OF MULTI-OBJECTIVE OPTIMISATION METHODS ............................................................. 33
4.4.1 Traditional Approaches......................................................................................................... 34
4.4.2 Evolutionary Approaches...................................................................................................... 35

4.4.2.1 Vector Evaluated Genetic Algorithm..............................................................................................35
4.4.2.2 Pareto-Based Genetic Algorithm ....................................................................................................36
4.4.2.3 Multi-Attribute Genetic Algorithm.................................................................................................37
4.4.3 Discussions............................................................................................................................ 37
4.5 SUMMARY ..................................................................................................................................... 39
CHAPTER 5 EVOLUTIONARY MULTI-OBJECTIVE OPTIMISATION APPROACHES....... 40

5.1 INTRODUCTION .............................................................................................................................. 40
5.2 MULTI-OBJECTIVE GENETIC ALGORITHM ..................................................................................... 41
5.2.1 Selection Processing with Rank Assignment ......................................................................... 41
5.2.2 Fitness Sharing...................................................................................................................... 42
5.2.3 Variable Recombination Operators ...................................................................................... 43
5.2.4 Treatment of Preferred Priorities among Objectives ............................................................ 43
5.3 MULTI-OBJECTIVE PARTICLE SWARM ALGORITHM ....................................................................... 45
5.3.1 Search Strategy ..................................................................................................................... 45
5.3.2 Rank-Based Selection ............................................................................................................ 47
5.3.3 Weight Update....................................................................................................................... 47
5.3.4 Pareto-Optimal Set Update ................................................................................................... 48
5.4 MULTI-OBJECTIVE DIFFERENTIAL EVOLUTION ALGORITHM ......................................................... 49
5.5 SUMMARY ..................................................................................................................................... 49
CHAPTER 6 RESULTS OF GO-CIRCUIT OPTIMISATION ........................................................ 50

6.1 OPTIMAL TRACTION-SUBSTATION PLACEMENTS ........................................................................... 50
6.2 WORST-CASE SCENARIOS OF OPERATIONAL DEVIATIONS ............................................................. 53
6.3 PERFORMANCE CHECK FOR FAILURE CONDITIONS ........................................................................ 56
6.4 SUMMARY ..................................................................................................................................... 58
CHAPTER 7 RESULTS OF RETURN-CIRCUIT OPTIMISATION.............................................. 60
7.1 MULTI-OBJECTIVE OPTIMISATION FOR NORMAL CONDITIONS ...................................................... 60
7.2 PERFORMANCE CHECK FOR FAILURE CONDITIONS ........................................................................ 68
III
Table of Contents
7.3 SUMMARY ..................................................................................................................................... 69
CHAPTER 8 CONCLUSIONS............................................................................................................. 71
8.1 SUMMARY AND CONCLUSIONS ...................................................................................................... 71
8.2 SUGGESTIONS FOR FUTURE WORK ................................................................................................ 74
REFERENCES ...................................................................................................................................... 75
APPENDIX A EVOLUTIONARY ALGORITHMS .......................................................................... 79
A.1 INTRODUCTION ............................................................................................................................. 79
A.2 GENETIC ALGORITHM ................................................................................................................... 79
A.3 PARTICLE SWARM AND DIFFERENTIAL EVOLUTION ALGORITHM ................................................. 80
A.3.1 Particle Swarm Algorithm .................................................................................................... 81
A.3.2 Differential Evolution Algorithm .......................................................................................... 82
APPENDIX B FLOWCHART OF PROPOSED EVOLUTIONARY MULTI-OBJECTIVE

OPTIMISATION APPROACHES....................................................................................................... 84
B.1 FLOWCHART OF MULTI-OBJECTIVE GENETIC ALGORITHM ........................................................... 84
B.2 FLOWCHART OF MULTI-OBJECTIVE PARTICLE SWARM AND MULTI-OBJECTIVE DIFFERENTIAL

EVOLUTION ......................................................................................................................................... 85
APPENDIX C PRELIMINARY TESTING OF MULTI-OBJECTIVE OPTIMISATION

ALGORITHMS ..................................................................................................................................... 86
C.1 INTRODUCTION ............................................................................................................................. 86
C.2 CONCAVE PROBLEM ...................................................................................................................... 87
C.3 DISCONTINUOUS PROBLEM ........................................................................................................... 89
C.4 SUMMARY ..................................................................................................................................... 91
IV
List of Figures
List of Figures
Figure 1-1: Schematic of Singapore MRT system........................................................1

Figure 2-1: Sectional network representation of double-track MRT system ...............13
Figure 2-2: Three-stage scheme for go-circuit optimisation .......................................18
Figure 2-3: Flowchart for go-circuit simulation .........................................................19
Figure 3-1: Return-circuit model ...............................................................................22
Figure 3-2: Simple case study of touch voltage and stray current ...............................25
Figure 3-3: Two-stage procedure for touch voltage and stray current .........................27
Figure 3-4: Flowchart of return-circuit simulation .....................................................28
Figure 4-1: Pareto front for bi-criterion minimisation problem...................................32
Figure 4-2: Nonconvex solution boundary.................................................................35
Figure 4-3: Outline of VEGA evolution results..........................................................36
Figure 4-4: Rank assignment for Pareto-based genetic algorithm...............................37
Figure 5-1: Rank assignments for different priorities among objectives .....................44
Figure 6-1: Optimised configurations ........................................................................51
Figure 6-2: Energy consumption convergence curve..................................................52
Figure 6-3: Load sharing convergence curve .............................................................52
Figure 6-4: Pareto-optimal set for Configuration 1.....................................................54
Figure 6-5: Pareto-optimal set for Configuration 2.....................................................55
Figure 7-1: Layout of study system............................................................................61
Figure 7-2: Touch voltage distribution with different earthing arrangement ...............63
Figure 7-3: Pareto-optimal sets for Configuration 1 ...................................................65
Figure 7-4: Pareto-optimal sets for Configuration 2 ...................................................66
Figure C-1: MOPS results for Test1 ..........................................................................89
Figure C-2: MODE results for Test 1.........................................................................89
Figure C-3: MOPS results for Test 2 .........................................................................91
Figure C-4: MODE results for Test 2.........................................................................91
V
List of Tables
List of Tables
Table 6-1: Improvement of optimised energy consumption and load sharing .............53
Table 6-2: Parameter limits for bi-criterion optimisation............................................54
Table 6-3: Maximum deviation for energy consumption............................................55
Table 6-4: Maximum deviation for load sharing ........................................................55
Table 6-5: Performance check results for case 1.1 .....................................................57
Table 6-6: Performance check results for case 2.2 .....................................................58
Table 7-1: Typical arrangements of earthing and bonding..........................................62
Table 7-2: Multi-objective optimisation of earthing & bonding for configuration 1 ...67
Table 7-3: Multi-objective optimisation of earthing & bonding for configuration 2 ...68
Table 7-4: Performance check for case 2.1 ................................................................69
Table C-1: Non-dominated solution numbers for Test 1.............................................88
Table C-2: Non-dominated solution numbers for Test 2.............................................90
VI
List of Abbreviations
List of Abbreviations
TSS: Traction Substation
GA: Genetic Algorithm
DE: Differential Evolution algorithm
PS: Particle Swarm algorithm
MOGA: Multi-Objective Genetic Algorithm
MODE: Multi-Objective Differential Evolution algorithm
MOPS: Multi-Objective Particle Swarm algorithm
VII
List of Publications Related to this Thesis
List of Publications Related to this Thesis
[1] C.S. Chang and L. Tian,” Worst-case identification of touch voltage and stray
current of DC railway system using genetic algorithm”, IEE Proceedings,
Electric Power Applications, Vol. 146, No. 5, 1999
VIII
Summary of the Thesis
MRT system design can be formulated as a problem of three-stage optimisation. In the
first stage, the basic design of a MRT section is optimised by extensively searching
through a large set of design alternatives. Only the key or primary variables are
optimised in this stage. The second stage evaluates the worst-case performance of the
basic design using secondary variables arising from operational deviations and other
random variables. The need for changing the basic design to cater for both the normal
condition and failure conditions is ascertained and implemented in the third stage.
MRT supply networks can be divided into the traction substation, the go-circuit and
the return-circuit. At traction substations, AC supply voltage is stepped down and
converted to DC. Catenary wires or third rails are used in the go-circuit while running
rails and return cables are the main components of return-circuit. In this work, the go-
circuit and return-circuit are each optimised with the procedure as outlined above.
Energy consumption and load sharing are two important issues in the go-circuit.
Energy consumption calculates the total energy consumed at all the traction
substations, and load sharing measures the load distribution among all traction
substations. They are influenced by many factors and their optimisation cannot be
obtained simultaneously. In the proposed first-stage optimisation, a previously
developed algorithm is incorporated for configuring the traction placements by
optimising either energy consumption or load sharing. During operation, train
timetables deviate continually from the predefined train despatch frequency due to
variations of train headway, synchronisation delay and dwell time. This work focuses
IX
on the second-stage optimisation, which implements the bi-criterion optimisation of
energy consumption and load sharing under normal condition. The system
performance is evaluated under failure conditions in the third stage.
With running rails used as part of the traction current return-circuit, the main concerns
in the return-circuit are the touch voltage and stray current. Touch voltage is the
voltage between the running rail and the ground. Excessive instantaneous touch
voltage jeopardises safety. Stray current is the leakage current between the running rail
and the ground. The stray current is likely to be picked up by the underground
structures in the vicinity and cumulative stray current may accelerate their corrosion.
The earthing and bonding strategy within the system has a profound impact on the
control of touch voltage and stray current. Meanwhile, the improvement of touch
voltage or stray current tends to deteriorate the other. The return-circuit optimisation is
thus composed of two-stage implementation. In the first stage, different earthing and
bonding arrangements at passenger stations or traction substations are extensively
explored under normal operation. The second stage conducts performance check on
each appropriate earthing and bonding arrangement with the list of credible failure
conditions such as rectifiers and inverters out-of-service.
The technique of Pareto-optimal set is developed for the above go-circuit and return-
circuit optimisations. The solutions are often multi-objective and seldom unique, as
they usually comprise a finite set of non-dominated or Pareto-optimal points. Three
evolutionary algorithms are applied, which are namely: the Multi-Objective Genetic
Algorithm (MOGA) for discrete problems, Multi-Objective Particle Swarm (MOPS)
algorithm and Multi-Objective Differential Evolution (MODE) algorithm for
X
continuous problems. MOPS and MODE are proposed to solve MRT problems for the
first time.
The three proposed algorithms are based on the concepts of population, rank-based
selection and competitive evolution. During optimisation, a population of candidate
solutions is evolved in the feasible space to search for the Pareto-optimal set. Ranking
of the population is accomplished through Pareto ranking, where all points in the
Pareto-optimal set are successively placed on different Pareto fronts. Competitive
evolution consists of selecting subsets of points with respect to their ranks and moving
them toward the Pareto-optimal set. Test analysis of the proposed algorithms is made
on each of the go-circuit and return-circuit simulations. Numerical comparisons of
MOPS and MODE against the Multi-Attribute Genetic Algorithm (MAGA) favour the
former two algorithms.
XI
Chapter 1 Introduction
1.1 Background
In recent years, many high performance Mass Rapid Transit (MRT) and Light Rapid
Transit (LRT) systems have been built. They are playing increasingly significant roles
in public transportation. In general, the MRT power supply systems comprise three
parts [10] of traction substation, go-circuit and return circuit as shown in Figure 1-1.
Being focused on the DC transit system, this thesis mainly examines how to achieve
trade-off in the go-circuit and the return-circuit optimisation. The harmonic effects are
not discussed in this thesis, but will be addressed by other researchers [30].
AC Busbar (22kV)
Transformer Transformer
TSS 1
TSS 2
Rectifier Inverter Rectifier

Catenary or Third
rail (1500V DC)
Go-circuit Go-circuit
Train
Runnung rail
Earthing device Earthing device

Return-circuit
Figure 1-1: Schematic of Singapore MRT system
1
• Traction substation: In Singapore MRT system, several traction substations (TSSs)
are located at passenger stations due to construction convenience and cost
consideration. From the viewpoint of electrifying the MRT systems, TSSs function
as energy sources, transferring power between the AC side and the DC side. The
66 kV supply voltage is stepped down to 22 kV at intake substations via the
transformer. Rectifiers installed at TSSs convert power from the 22 kV AC into
DC so as to provide the go-circuit with traction current. While trains are usually
equipped with the regenerative braking, some TSSs are configured with inverters
to increase line receptivity. As a result, power released during regenerative braking
is either consumed by nearby accelerating trains or returned to supply through
inverters. The 22 kV AC is also stepped down to furnish service loads, such as
lighting, air-conditioning and other auxiliary devices at passenger stations.
• Go-circuit: This contains the catenary wires or third rails and positive feeders. In
most MRT systems, running rails are used as traction return paths. Hence the
traction current, distributed from the TSSs, flows along the catenary wires or third
rails to supply trains with power and then flows back to the TSSs through the
running rails.
• Return-circuit: Running rails, rail bonds and negative feeders are the main
components of the return-circuit. Running rails are often lightly insulated from the
ground so a fraction of the traction currents may leak into the earth whereas the
bulk of traction currents return to the TSSs via the negative feeders.
1.2 Motivation of the Research
2
Single-objective optimisation techniques have long been applied in MRT system.
Nevertheless, there are needs for multi-objective optimisation. Energy consumption
and load sharing in go-circuit, influenced by many aspects, cannot be optimised
simultaneously. Likewise, the mitigation of touch voltage and stray current in return-
circuit are not likely to be achieved at the same time. Practical solutions for multi-
objective optimisation are seldom unique, as they comprise a finite set of non-
dominated or Pareto-optimal points. In this thesis, the technique of Pareto-optimal set
(Section 4.2) is developed for the go-circuit and return-circuit optimisations in MRT
system. Three evolutionary algorithms are explored, namely: the Multi-Objective
Genetic Algorithm (MOGA, Section 5.2) for return-circuit optimisation, Multi-
Objective Particle Swarm (MOPS) algorithm (Section 5.3) and Multi-Objective
Differential Evolution (MODE) algorithm (Section 5.4) for go-circuit optimisation.
MOPS and MODE are proposed to solve MRT problems for the first time.
1.2.1 Go-Circuit Optimisation
In MRT systems, TSSs supply train power via the rectifiers and receive the
regenerated power via the inverter. The energy consumption, which is calculated as the
sum of the power flowing through the rectifiers and the inverters installed at each TSS,
is quite crucial to the efficiency of the MRT system. In addition to energy flows
through rectifiers and inverters, load sharing among TSSs is also of importance. As
trainloads are highly fluctuating, the TSS loads can be unevenly distributed. Some
TSSs are overloaded but others are under-loaded. In case of some traction substations
outage, the power flowing through the nearby traction substations could exceed their
3
capacity. Therefore, energy consumption and load sharing are selected as optimisation
objectives in go-circuit optimisation.
Energy consumption and load sharing are dependent on the instantaneous train
positions and control status (accelerating, coasting, dwelling or braking). These are
influenced by the service schedule (declared train despatch frequency or headway) and
operational deviations (synchronisation delay and train dwell times). Traction
substation placements also have profound impact on the energy consumption and load
sharing.
At the first stage of go-circuit optimisation, a previously developed algorithm [3] is
incorporated to configure the traction substation placements by optimising either
energy consumption or load sharing. Energy consumption and load sharing are not
consistent with each other. The same train running timetable will probably lead to their
different regulation direction. At the second stage, the effect of operational timetable is
discussed, and the competing nature of energy consumption and load sharing is
investigated. The optimisation variables, namely headway, synchronisation delay and
dwell times, vary continually during the simulation period. In order to solve this
continuous problem, MOPS and MODE algorithms are proposed and applied for the
first time to generate a variety of Pareto-optimal solutions. The compromise between
worst-case energy consumption and load sharing is then identified for performance
margin specification, which is implemented at the third stage of go-circuit
optimisation.
1.2.2 Return-Circuit Optimisation
4
Running rails are usually used as the traction current return paths. Owing to the rail-to-
ground and rail resistance, there will be a voltage difference caused by the return
current flows from between the rails and the local ground known as touch voltage.
Excessive touch voltages jeopardises safety. As running rails are often lightly
insulated, the traction current flowing back to the substations may partly take the
ground. Known as stray current, it is likely to be picked up by the underground
structures in the vicinity and through the ground to enter another structure before
returning to the TSS. Accumulative stray currents may accelerate the structures’
corrosion. Accurate evaluation and effective control of touch voltage and stray currents
are therefore the consideration factors in the return-circuit.
Simulations and field tests reveal that both touch voltage and stray current are greatly
influenced by the earthing strategy and bonding arrangement adopted in MRT system.
However, the improvement of touch voltage or stray current tends to deteriorate the
other. A two-stage scheme is thus proposed for return-circuit optimisation. At the first
stage, Multi-Objective Genetic Algorithm (MOGA) is developed to extensively
explore earthing and bonding arrangements so that the compromising mitigation of
touch voltage and stray current is obtained under normal running condition. At the
second stage, the most appropriate earthing and bonding patterns are then picked up
for performance check under credible failure conditions.
1.3 Objective and Scope of the Research
MRT systems are complex and highly integrated. The scope of this research is
confined to the electrification subsystem and focused on the DC side. In particular,
5
effects of operational parameters rather than design parameters are investigated for the
safety and efficiency of MRT system. Two sets of objectives are defined for the
problem of energy consumption and load sharing in the go-circuit, and the problem of
touch voltage and stray current in the return-circuit. Variables governing these two sets
of objectives can either be discrete or continuous. The nature of competing objectives
in these two problems is explored. Evolutionary algorithms are then applied for
attaining satisfactory trade-offs within these sets of objectives.
Although the use of rectifiers and inverters does introduce harmonic in the MRT
system, harmonic can be minimized by employing appropriate type of power
converters and by placing active harmonic filters [29]. This thesis focuses on the
application of multi-objective optimization of primary variables to MRT systems. The
harmonic effects on go-circuit and return-circuit optimisation are not addressed in this
thesis, but will be discussed by other researchers [30].
1.4 Multi-Objective Optimisation Algorithms
Multi-objective optimisation problems involve in simultaneous optimisation of
multiple non-commensurable objectives. The solution to such problems is not unique
but a family of non-dominated points (Pareto-optimal set). In traditional methods of
multi-objective optimisation, different criteria are linearly blended into a composite
scalar objective. This requires pre-establishment of the weights of different criteria. As
it is never a simple task to specify an appropriate set of weights, optimal solutions are
individually obtained for a range of weights and the so-called trade-off curves are
6
generated [5]. The computing time required for generating the trade-off curves is high
and it is also difficult to apply such techniques to non-convex problems.
The need for an improved multi-objective optimisation method to seek the Pareto-
optimal solutions is evident, and such a method should have the following
characteristics [23]:
• Efficiency: It can approximate or identify the entire global Pareto-optimal set
through a one-time solution of the multi-objective optimisation problem, avoiding
the need to solve a multitude of single-objective optimisation problems.
• Objectivity: It does not require the priori assessment of preferences to objective
functions in the generation of the Pareto-optimal set.
• Reliability: It can facilitate the evaluation of candidate non-dominated solutions in
a quick and reliable fashion.
The search processes of evolutionary algorithms, using a population of candidates,
suggest their application to multi-objective optimisation problems for finding a number
of Pareto-optimal solutions in parallel. In this thesis, Genetic Algorithm (GA), Particle
Swarm (PS) algorithm and Differential Evolution (DE) algorithm are appropriately
refined for multi-objective optimisation.
The performance of GA is mostly determined by its selection operation. In order to
deal with multiple conflicting objectives, a degree of control should be exerted over
the selection process. With the employment of rank assignment method [16], MOGA
selects individuals for survival according to their mutual dominance as well as their
fitness values. It attempts to trace all the non-dominance individuals in the present
7
population as far as possible and each objective function is utilised separately rather
than collectively. When all the non-dominated individuals in the current generation are
picked, the recombination operators are then applied to produce the next generation.
The above procedure is iterated to locate the Pareto-optimal points and produce
subsequent populations until convergence is met. At the end of MOGA evolution, the
final non-dominated set represents the collection of compromising solutions among all
the objectives.
Encoding brittleness in GA has degraded its performance in continuous problem
domain. Contrary to GA, MOPS and MODE avoid parameter encoding and work
directly on real-value parameter vectors to search for the optimal solutions. The
concept of scalarising function is also introduced to replace the global and individual
best solutions with compromise solutions respectively. The behaviour of each
individual in the population is influenced not only by its own historical performance
but also by its peers. Moreover, rank-based selection is applied in MOPS and MODE
to determine the individual’s likelihood of producing offspring and distribute the
population towards the promising areas.
By modifying the individuals’ ranks, the specification of preferences among objectives
is allowed in MOGA, MOPS and MODE. The technique of fitness sharing is adopted
to maintain the population diversity. With the treatment of dominance properties
among individuals, these evolutionary algorithms possess a more exploitative
mechanism to obtain the Pareto-optimal set.
1.5 Organisation of the Thesis
8
This thesis is organised as follows:
• Chapter 1 describes briefly the background and the objective of the research. Two
kinds of multi-objective problems, i.e., touch voltage versus stray current (discrete
problem) and energy consumption versus load sharing (continuous problem) are
introduced by outlining their corresponding optimisation approaches.
• Chapter 2 defines the energy consumption and load sharing in go-circuit. After
presenting the DC-powered two-track MRT system model, this chapter formulates
a three-stage scheme for optimising energy consumption and load sharing and
checking system performance.
• Chapter 3 defines the touch voltage and stray current in return-circuit. The
concepts of return-circuit modelling and load referral solution method are
explained. A two-stage scheme is also put forward to attain the trade-off between
touch voltage and stray current, and to evaluate the system performance under
failure conditions.
• Chapter 4 categorises multi-objective optimisation methods and identifies their
specific advantages and disadvantages.
• Chapter 5 develops a Pareto-based multi-objective genetic algorithm. A particle
swarm algorithm based as well as a differential evolution based multi-objective
optimisation approach is also proposed for the first time to solve the continuous
problems.
• Chapter 6 discusses mainly the impact of actual operational timetable on energy
consumption and load sharing calculation. The traction substation placements are
9
first configured by obtaining either minimal load sharing or equal load sharing.
Then the proposed continuous multi-objective optimisation methods are applied to
explore the compromise nature of energy consumption and load sharing against
varying operational timetable. Simulation results indicate the effectiveness of the
proposed three-stage scheme for fulfilling the design objective.
• Chapter 7 discusses the influence of earthing and bonding policies, and, employs
the proposed discrete multi-objective algorithm to achieve trade-off between the
touch voltage and the stray current. The priority assigned to different objectives
allows the decision-maker’s preference to be attained during optimisation.
Simulation results for different earthing and bonding arrangements are compared
and the promising solutions are further checked under failure conditions.
• Chapter 8 summarises the research with final conclusion and some suggestions for
future work.
10
Chapter 2 Outline of Go-Circuit Optimisation
2.1 Introduction
Energy consumption by TSSs and load sharing among TSSs in go-circuit are
influenced by diverse factors such as station spacing, service schedule, firing angles of
rectifiers and inverters, and transformer tap positions. The impact of traction substation
placements and firing angles of rectifiers and inverters were investigated in [3,5]. In
[3], the energy consumption and load sharing were treated as two separate objective
functions, and Tabu Search (TS) algorithm was applied to explore different
combinations of traction substation placements. For simplicity of optimisation, the
energy consumption and load sharing were combined linearly in [5] to form a single
objective. The conventional Genetic Algorithm (GA) was then employed to examine
the effect of firing angles of rectifiers and inverters. The performance curves were
plotted against the objective weights reflecting the relative significance among
objectives, and the so-called trade-off was made between energy consumption and load
sharing.
The competing nature between energy consumption and load sharing is not fully
discussed in the above work. The traction substation placements and train operational
timetable have great influence on the MRT system, which runs under normal
conditions or failure conditions. Therefore, the go-circuit optimisation in this thesis
presents normal and failure conditions, and is carried out in three stages. Stage I uses
the algorithm of [3] to optimise the traction substation placements, so that minimum
11
energy consumption or equal load sharing is attained. At Stage II, the train operational
timetable varies continuously. Two novel multi-objective optimisation methods
(Multi-Objective Particle Swarm and Multi-Objective Differential Evolution
algorithm) are developed for the first time in this work to obtain the Pareto-optimal set,
and achieve the trade-off between energy consumption and load sharing under normal
running conditions. At Stage III, the worst-case solutions are picked up from Pareto-
optimal set and the performance check is implemented under failure conditions.
2.2 Mathematical Model
2.2.1 System Model
Figure 2-1 schematically shows a two-track DC-powered MRT system. The
components are modelled in terms of resistances and current sources, and the
interconnection of these components constitutes a nodal electrical circuit.
From the viewpoint of circuit theory, the DC-fed MRT system can be divided into go-
circuit and return-circuit. In go-circuit, the traction power provided by TSS flows
through the catenary wire or third rail, supplies trains and passes back to TSS via
running rails. TSS is represented by a Thevenin equivalent voltage source (constant
voltage in series with a resistance) or a Norton equivalent current source (constant
current in parallel with a resistance). Trains are modelled as a voltage-dependent
resistance and the running rail is represented by lumped parameters in go-circuit.
The return-circuit deals with the touch voltage and leakage current. Its simulation is
explained in Chapter 3.
12
Negative Busbar
Substation
Positive Busbar
Feeders
Up Track
Catenary or
Train Third Rail
Rails
Earth
Crossbond
Down Track
Figure 2-1: Sectional network representation of double-track MRT system
2.2.2 Objective Functions
Energy consumption and load sharing are two objective functions in the go-circuit
optimisation.
In order to increase the system receptivity, some of the traction substations (TSSs) are
fitted with inverter to provide stable regenerative braking and transfer the surplus
braking energy to the AC side. That is, TSS supplies trains with nominal DC voltage
(1500 V in Singapore MRT system) via rectifiers, and recovers energy by the use of
inverter. Although the presence of rectifiers and inverters introduces harmonic to MRT
13
systems, the harmonic effects can be minimized by employing appropriate types of
power converters and by placing active harmonic filters [29]. Therefore, the harmonic
effects are not addressed in this thesis and will be covered by other researchers [30].
The energy consumption here is defined as the sum of traction energy contributed by
all the TSSs, i.e. the total power flows through the rectifiers and the inverters installed
at the TSSs.
t0 + headway N tss
 
EC = ∫  ∑ Ptss (t )  ⋅ dt / headway
i
t0  i =1  (2-1)
i
where EC represents the energy consumption. Ptss (t ) stands for the power flow through
the i-th TSS at time t. Ntss is the total number of TSS. t0 is the simulation start time.
Considering the respective rectifier loading and inverter loading, load sharing is
formulated as the sum of the deviation power of each rectifier and inverter from their
mean values.
t0 + headway
 N rec j N inv 
LS = ∫  ∑ Prec (t ) − Prec,mean (t ) + ∑ Pinvk
(t ) − Pinv ,mean (t )  ⋅ dt / headway
t0  j =1 k =1 
(2-2)
j
where LS represents the load sharing. Prec (t ) stands for the power flow through the j-
k
th rectifier and Pinv (t ) stands for the power flow through the k-th inverter. Nrec and Ninv
are the number of rectifiers and inverters installed at TSSs respectively. The mean
rectifying power Prec ,mean (t ) and the mean inverting power Pinv ,mean (t ) are computed as
14
N rec
Prec.mean (t ) = ∑ Prec
i
(t ) / N rec
i =1
N inv
(2-3)
Pinv. mean (t ) = ∑ P (t ) / N inv
i
inv
i =1
2.2.3 Impact of Operational Timetable
Power network of MRT system is a complex combination of substations, feeding trains
with dynamically varying demand. Energy consumption and load sharing, which are
functions of the instantaneous net power drawn from TSSs, largely depend on the
operational timetable being used.
Operational timetable variations can be described with three time variables in units of
second: headway, synchronisation delay and dwell time. Headway and synchronisation
delay variables determine the distance between two adjacent trains running in the same
direction and in the opposite (up- or down-) directions respectively. Dwell time, also
known as the station-waiting period, leads to deviation of each train schedule from the
prescribed timetable.
Under ideal operating condition, trains travel along the track in accordance with
established train movement profile and schedule. Trains in the same direction are
dispatched at a prescribed headway and the corresponding trains in opposite directions
are run with certain synchronisation delay. Headway varies with the time of the day,
being shorter during rush hours and longer during late-evening and early-morning
hours. When trains stopover at stations, regardless of how passengers are crowded,
trains are ideally assumed to leave for the next station after fixed dwell time. Due to
15
repeatable nature of MRT systems, each train follows ideally the same movement
pattern.
Under actual operating conditions, however, headway, synchronisation delay and train
dwell time will deviate from the prescribed pattern due to variations in passenger load
and the state of the track ahead. Therefore, the net power drawn from the TSSs, which
is governed by the relative positions and demands of trains both in the same direction
and in the opposite direction, will be inevitably affected. Energy consumption and load
sharing are in turn influenced by the varying operational timetable.
2.3 Layout of Three-Stage Scheme
Considering the impacts of substation placements and system operational timetable,
the go-circuit optimisation is divided into three stages as Figure 2-2.
At Stage I, Tabu Search (TS) algorithm is applied to explore different possible
combinations of traction substation placements, and optimise energy consumption or
load sharing under predefined operational timetable. By guiding the local search
descent method to avoid bad local optima, TS leads to the appropriate traction
substation placements for either minimum energy consumption or equal load sharing.
In practical operation conditions, the actual operational timetable may deviate
continuously from the predefined value due to unpredictable emergency such as an
abruptly increased number of passengers. At Stage II, the operational timetable is
taken as the optimisation variable, and the compromise worst cases between energy
consumption and load sharing are identified. Multi-Objective Particle Swarm (MOPS)
algorithm and Multi-Objective Differential Evolution (MODE) algorithm proposed in
16
Section 5.3 are applied to exploit the competing nature of energy consumption and
load sharing with varying headway, synchronisation delay and dwell times. The
Pareto-optimal set is obtained to reflect the trade-off between energy consumption and
load sharing.
Start
Select specific operational timetable (headway,

synchronisation delay and dwell time)
Stage I
Obtain optimal traction substation
placements by applying TS
Change operational timetable for

worst-case identification
Stage II
Obtain Pareto-optimal Set by
using MOPS or MODE
Select one solution from the

Pareto-optimal Set
Performance check with the list

of credible failure conditions
Relocate the
traction substation
placements Performance No
accepted ? Stage III
Add to the forbidden

Yes timetable Set
No All solutions
selected?
Yes
Yes
All solutions fail?
No
Stop
17
Figure 2-2: Three-stage scheme for go-circuit optimisation
In the failure conditions, some rectifiers and inverters at the TSSs or even the TSS are
out-of-service. The MRT system should sustain from such credible failure scenarios.
Stage III conducts the performance check for solutions picked up from the Pareto-
optimal set. If the system performance is undesirable, the corresponding operational
timetable should be avoided. A list of avoidable operational timetable is generated
after all the Pareto-optimal solutions are checked. In case that all the worst cases fail in
the performance check, the traction substation placements derived from Stage I need
some adjustments to ensure the reliability of system operation.
2.4 Simulation Outline
The power flow in MRT system is affected by numerous factors, which include track
topology, electrical equipment characteristics and operational timetable.
In the go-circuit simulation as illustrated in Figure 2-3, given certain track topology
and train movement control strategy, the train movement module is first executed and
the train current-distance and speed-distance profiles are obtained. Then, according to
the specified operational timetable, the locations of trains are computed as functions of
simulation time. The system nodal admittance matrix is also formed given the traction
substation placements. As the moving trainloads cause the structure of nodal
admittance matrix to fluctuate, the load referral method [4] is utilised to eliminate the
train nodes by referring them to the nearest traction substations. The formed network
nodal admittance matrix thus has fixed dimension. After calculating the TSS injection
currents and the train consumption currents, the TSS and train voltages are computed.
18
In this way, the power flow in go-circuit is deduced, and the energy consumption and
load sharing are obtained.
Start
Set track topology and train

control strategy
Run train movement module to obtain

train movement profile
Specify operation timetable
Implement load referral to form the

network nodal admittance matrix
Solve eqn. [V]=[Y][I] to derive power flow
Return-circuit simulation
Calculate energy consumption and load

sharing
End
Figure 2-3: Flowchart for go-circuit simulation
2.5 Summary
Based on the DC-powered MRT system model, this chapter outlines the go-circuit
optimisation procedure and clarifies the influence of operational timetable on energy
consumption and load sharing. The optimisation of energy consumption and load
sharing are conflicting and affected by traction substation placements and operational
19
timetable deviation. A three-stage scheme is accordingly put forward to implement bi-
criterion optimisation of energy consumption and load sharing under normal condition
and evaluate the system performance under failure conditions.
20
Chapter 3 Outline of Return-Circuit Optimisation
3.1 Introduction
In MRT systems, running rails are used as the traction current return conductor, which
causes concern for touch voltage and stray current. In order to realise the computer-
based simulation of touch voltage and stray current, the finite-cell idea in power
system was introduced for return-circuit modelling. By dividing the track into a
number of equal-length cells, a multi-conductor shunt energised model was
accordingly set up to calculate the touch voltage [7]. The follow-on work [8] then
examined and discussed the relative merits of four earthing schemes, namely: float
earth, float earth with rail potential control device, direct earth and diode earth, on the
rail potential and stray current. The control of touch voltage and stray current were also
addressed in [11,12,13] from the viewpoint of engineering rather than from computer
simulation.
The monitoring and mitigation of touch voltage and stray current are receiving
increasing attention while their computer simulation is well done. The problem of
reducing the touch voltage and stray current in DC railways is multi-objective and
conflicting. It is affected by many factors such as the earthing and bonding design as
well as the normal and failure operating conditions. Therefore, the return-circuit
optimisation in this thesis is performed in two stages. At Stage I, a multi-objective
optimisation method, Multi-Objective Genetic Algorithm, is applied to search among
various earthing and bonding arrangements to obtain Pareto-optimal set for touch
21
voltage and stray current. The worst cases are then picked up for performance check at
Stage II.
3.2 Mathematical Model
3.2.1 Return-Circuit Model
In the return-circuit, trains inject traction currents into the running rails, and, TSSs
absorb the return currents from the running rails. All these currents are actually
injected currents with appropriate (positive or negative) signs. Since the running rail is
represented as a single conductor with distributed leakage resistance, i.e., a
transmission line in Figure 3-1(a), the return-circuit can be modelled as a transmission
line under shunt energisation at multiple points [7].
I1 I2 ... Ik ... In-1 In Injection Current
Running Rails
(a) A transmission line under shunt energisation
Nodes ... k-1 k k+1 ...

1
:
Conductors
j
:
m
(b) A multi-conductor system
Figure 3-1: Return-circuit model
22
When the four running rails in double track MRT systems are not bonded together, the
four running rails cannot be treated as a single conductor. Instead, a multi-conductor
system model as in Figure 3-1(b) will be used. This model is based on the multi-
conductor theory given in [1].
3.2.2 Objective Functions
The touch voltage and stray current are the objective functions in return-circuit
optimisation.
The touch voltage in this thesis refers to the voltage difference between the rails and
the local ground as below.
Vtouch = Vr − V g (3-1)
where Vtouch is the touch voltage, Vr is the running rail voltage and Vg is the local
ground surface potential.
Touch voltage takes instantaneous value and brings about immediate harm to
personnel, which is not accumulated with the time. On the contrary, as stray current
causes accumulative corrosion to nearby underground structures, which is proportional
to the product of the magnitude of stray current and the time duration, stray current
used in this thesis is represented by the stray-current integral collected from all
earthing points. The stray-current integral, independent of the simulation duration, is
defined as [8]:
N (t )
I str = ( ∫tt12 ∑ i re (t , k ) ⋅ dt ) /(t 2 − t1 ) (3-2)
k =1
23
where i re (t , k ) is the current flow from running rail to the earth at electrical node k and
at time t, and N(t) is the total number of electrical nodes at time t. The simulation spans
from time t1 to t2.
3.2.3 Impact of Earthing and Bonding Arrangements
Touch voltages and stray current are influenced under normal condition by factors,
such as the traction substation spacing, rail resistance and rail conditions, headway
interval, synchronisation delay, weather condition, soil resistivity, and earthing and
bonding design. Among these factors, a close TSS spacing decreases the go-circuit
voltage drops but the spacing is determined by other considerations such as the ease of
construction and commercial benefit. Once the MRT system is designed, the rail
resistance is fixed and cannot be changed easily. The headway depends on the
passenger traffic flow. It usually takes constant values during certain periods, e.g., rush
hours, normal hours and evening hours. The synchronisation delay varies during the
actual train run from zero to the headway specified at that time. In addition, the touch
voltage and stray current are affected by operational abnormality (or failure), which
can be results of track paralleling switched off, substations out of service etc.
To briefly estimate the impact of earthing and bonding arrangement on the return-
circuit, a simplified single-TSS and single-train model for DC transit system is shown
in Figure 3-2. The impacts of harmonics are not shown in the model, as they can be
minimized by employing appropriate power converters and by placing active harmonic
filters. The harmonic effects on MRT systems are discussed in [30]. The stray current
of this model is:
24
RR I T
Is = (3-3)
RT + RR + RS
where I T is the train current. RR is the running rail resistance. RS is the earth
resistance at the TSS and RT is the earth resistance as seen at the train.
Catenary wire
+
Substation Train
IT
- IR
RR
IS Running rail IS
RS RT
Figure 3-2: Simple case study of touch voltage and stray current
In Equation (3-3), the train current IT is largely affected by the train schedule and
passenger flow. As the running rail resistance RR is decided at the design stage of the
MRT system, RS and RT are the only variables to be investigated. In general, low RS
and RT give rise to a high stray current but a relatively low rail potential. Regular
bonding of the rails equalises the rail-to-earth potentials of all rails along the return-
circuit. It also reduces the return resistance because of parallel rail paths constituting
the return-circuit. The earthing and bonding arrangements hence play a significant role
in determining the touch voltages and stray current.
3.3 Layout of Two-Stage Scheme
25
To provide accurate modelling of stray current, the peak loading condition should be
studied. On the other hand, touch voltage tends to be higher under failure conditions.
Meanwhile, improvement of the stray current or touch voltage tends to deteriorate the
other. For instance, high rail-to-ground insulation is liable to present large touch
voltages but small stray currents. Therefore, a two-stage scheme is proposed as in
Figure 3-3 to represent the normal and failure conditions.
At Stage I of return-circuit optimisation, the method of Pareto-optimal set is developed
to best improve the objectives of touch voltages and stray current for normal running
condition. The Multi-Objective Genetic Algorithm (MOGA) described in Section 5.2
is applied to obtain a set of Pareto-optimal solutions. In each solution, the earthing and
bonding design is optimised and therefore represents a different degree of trade-off
between touch voltages and stray current. The Pareto-optimal set from Stage I contains
a large collection of optimal designs. This is also a source of reference for different
design variations. Because the specific times at which initiating events that cause
failure conditions are unpredictable, any optimal design taken from Stage I must be
operated at all times, in such a way that the system will not be endangered, should any
credible failures occur.
At Stage II, the decision-maker picks one optimal design from the Pareto-optimal set,
and performs performance-check with the list of credible failure conditions. The worst-
case touch voltage and stray-current integral is identified during checking. The
decision-maker prepares his/her own list of credible failures from predefined events
such as substation out-of-service, broken or deteriorating bonding and joints etc.
Should the worst case be unsatisfactory, the decision-maker is prompted to pick
another optimal design from the Pareto-optimal set for further performance-check.
26
Start
Select service timetable for

the study
Stage I
Obtain Pareto Optimal Set by using
MOGA for normal running condition
Preference of
decision-maker
Select one solution from the Set
Performance check with the list of Stage II

credible failure conditions
Performance No
Accepted ?
Yes
Stop
Figure 3-3: Two-stage procedure for touch voltage and stray current
3.4 Simulation Outline
The flowchart of the return-circuit simulation is illustrated in Figure 3-4. After
specifying the load configuration, the train movement module is first run to obtain the
speed-distance profile and the current-distance profile. The AC/DC load flow is then
employed to calculate the train currents and TSS currents in the go-circuit. Using these
currents, a second network solution is applied to the return-circuit to calculate the
touch voltage and stray current.
27
Start
Set load configuration
Train movement module
Create current-distance and

velocity-distance profiles
Initialise simulation time t=0
No
t < Tmax ?
Yes
Calculate load flow to produce Search the minimum
system operation parameters touch voltage and
calculate the stray
current integral
Apply Gauss elimination to
form π circuit
Output results
Solve eqn. [I]=[Y][V] to obtain
instantaneous values of touch
voltage and stray current End
t = t+1
Figure 3-4: Flowchart of return-circuit simulation
3.5 Summary
After introducing the return-circuit modelling, this chapter discusses the impact of
earthing and bonding arrangements. The control of touch voltage and stray current are
conflicting and affected by normal and abnormal conditions. A two-stage scheme is
thus formulated to carry out multi-objective optimisation for different earthing and
28
bonding arrangement under normal condition, and implement performance check
under failure conditions.
29
Chapter 4 Literature Review of Multi-Objective Optimisation Approaches
Chapter 4 Literature Review of Multi-Objective
Optimisation Approaches
4.1 Introduction
Problems with multiple criteria can be broken down into two broad categories:
• Multiple-objective programming problems (i.e., multi-objective optimisation
problems), which have a very large number of feasible alternatives described
with relation to the decision variables in a deterministic environment.
• Multiple-attribute problems, which have a relatively small number of alternatives
represented in terms of attributes in an uncertain environment.
Multi-objective optimisation theory was founded in welfare theories by Pareto during
18th century. Since then, multi-objective optimisation has been introduced to several
engineering disciplines such as water resource planning, bridge construction, and
chemical plant design. This thesis discusses the application of multi-objective
optimisation to MRT systems.
4.2 Mathematical Definition
Given a set of objective functions, the multiple-objective optimisation aims to find a
solution parameter vector so that all the objectives can be optimised in a compromise
30
way. Generally, the vector of F(x) in multi-objective optimisation contains a group of
objective functions fi(x), which is to be minimised (or maximised) individually:
min F ( x) = [ f 1( x), f 2( x ), Λ , fn( x)]

x∈C (4-1)
subject to C = {x : h( x) = 0, g ( x) ≤ 0}
where n (≥2) defines the number of objective functions in the vector F(x). x is
presented as the decision variable vector and C denotes the feasible set of x that is
subjected to equality and inequality constraints.
Observing that the values f1(x),…, fn(x) represent incommensurable quantities that
cannot be minimised simultaneously, there is no unique solution to this problem.
Instead, characteristic of the multi-objective optimisation problem is a very large set of
acceptable solutions. These solutions are not optimal at the viewpoint of any single
objective, and they cannot be improved with respect to one criterion without
deteriorating at least one other criterion. But they are superior to the rest of the
solutions in the search space when all the objectives are considered. Elements with
such characteristic are known as Pareto-optimal solutions or non-dominated solutions.
Since none of the solutions in a Pareto-optimal set is absolutely better than any other,
any one of them might be of interest and an acceptable solution.
Definition: A point x*∈C is qualified as a Pareto-optimal or an efficient solution or a
non-dominated point for multi-objective optimisation problem, if and
only if there is no x ′ ∈C such that fi( x ′ )≤fi(x*) for all i∈{1,2,…n} with at
least one strict inequality.
31
non-dominated
solutions
f2
dominated
solutions
f1
Figure 4-1: Pareto front for bi-criterion minimisation problem
Figure 4-1 shows all the possible (dominated and non-dominated) solutions of a bi-
criterion optimisation problem for f1(x) and f2(x), where all non-dominated solutions
are situated on a circular arc known as a Pareto-optimal front. Point A and point B are
non-dominated solutions as moving from A to B yields an increase in f1 but a decrease
in f2. Solution points that are not situated at the front are dominated solutions and are
discarded during the optimisation.
4.3 Preference Structure
The scalar concept of optimality in single-objective optimisation does not apply
directly in the multi-objective problem because the objective vector F(x) lacks a
natural ordering whenever F(x) is vector-valued. Usually, the decision-makers know
some of the trade-offs being associated with the problem at hand and determine the
ordering. In this sense, it is more meaningful to define multi-objective optimisation as
finding a feasible alternative, x, that yields the most preferred vector of objective
function values, F(x), based on the decision-maker’s true preferences.
32
In order to effectively discriminate between two candidate solutions x0 and x1, some
type of preference structure is imposed on F(x), which defines the relevance of each
objective function. Three preference structures for multi-objective optimisation
problems are introduced in [14].
• Value function preference: x0 is said to dominate x1, in the case that a function g
is defined on the objective space F such that g(x0)<g(x1) if and only if x0 is
preferred to x1.
• Pareto preference: The Pareto-optimal set of a multi-objective optimisation
problem is the set of all non-dominated points.
• Lexicographic preference: The solution x0 is said to dominate x1 if fk(x0)<fk(x1)
and fi(x0)=fi(x1) for i=1,…,k-1.
The Pareto preference structure is most widely used in multi-objective optimisation. It
is also applied in this thesis.
4.4 Review of Multi-Objective Optimisation Methods
While stating the multi-objective optimisation problem as in Equation (4-1) is simple,
solving it to identify the Pareto-optimal front, a rarely finite set, is not an easy task.
Depending on the preference assessed to the different objectives involved, multi-
objective approaches can be summarised into three categories, e.g., methods that
require a priori, a posteriori and interactive articulation of preference among objectives
[23]. However, according to their development, multi-objective methods can also be
classified as traditional approaches and evolutionary approaches.
33
4.4.1 Traditional Approaches
The traditional methods always solve multi-objective optimisation problems by
scalarising the multiple objectives into a single objective whose solution is a Pareto-
optimal point for the original multi-objective problem. In this way, the theory of single
objective optimisation can be applied to multi-objective problems
The weighing objective method, one of the typical traditional approaches, transforms
the multi-objective problem into a single-objective optimisation by weighing each
objective with a positive constant and aggregating them to obtain the overall objective
function:
k
H ( x ) = ∑ wi * f i ( x ) (4-2)
i =1
where k is the total number of the objectives and fi(x) is the i-th objective function. The
weights wi are non-negative numbers assumed beforehand but not all zero, that is, for
some i≤k, wi>0.
The solution to this combined objective is Pareto-optimal and each selection of
weights yields a particular point within the Pareto-optimal front. By assigning different
weight vectors for the objectives, this weighing algorithm will generate and converge
to various efficient solutions. Since the chosen weights sometimes do not reflect the
relative importance of each original objective appropriately, some degree of
uncertainty is introduced. In addition, the relationship between the vector of weights
and the Pareto front is such that a uniform spread of weight parameters rarely produces
a uniform spread of points on the Pareto front. [21] Often, all the points found are
clustered in certain parts of the Pareto front with no points in the interesting area. A
34
complete set of Pareto-optimal solutions can be obtained for convex problems by
weighing method. The concave portions on the Pareto front as the segment between
point A and B shown in Figure 4-2, however, cannot be found for concave problems.
f2
f1
Figure 4-2: Non-convex solution boundary
4.4.2 Evolutionary Approaches
Evolutionary algorithms include genetic algorithm, differential evolution algorithm,
particle swarm algorithm and their variants. So far there is no literature on the multi-
objective differential evolution or multi-objective particle swarm algorithm. Only a
review of multi-objective genetic algorithm is presented here.
4.4.2.1 Vector Evaluated Genetic Algorithm
The Vector Evaluated Genetic Algorithm (VEGA) [16], which treats the non-
commensurable objectives individually, was the pioneer work for applying
evolutionary methods to solve multi-objective problems. The population in the current
generation is divided into sub-populations based on each of the objectives to reproduce
the next generations. The selection procedure is executed independently for each
objective, but crossover is performed across sub-population boundaries. Despite
35
simple implementation of this scheme, the independent selection of champions in each
objective has the risk for bias against the middle individuals. Those middle points that
have average values on all objectives are discarded during selection procedure. Hence,
the generated candidate solutions tend to satisfactory results at only one objective but
perform poorly at the other objectives as shown in Figure 4-3. This problem is even
more severe for a concave Pareto front.
f2
f1
Figure 4-3: Outline of VEGA evolution results
4.4.2.2 Pareto-Based Genetic Algorithm
Different from the VEGA, the Pareto-based approach is to process selection and
reproduction based on not only the objective values themselves but also their
dominance properties. All the locally non-dominated points achieve equal reproductive
potential through a non-dominated sorting procedure. The Pareto-based genetic
algorithm assigns rank to each individual in the population with regard to its own
fitness value and mutual dominance feature. In a rank assignment scheme [16], the
non-dominated individuals in the population are ranked as 1 and marked. Then the
non-dominated individuals among the unmarked points are ranked as 2 and marked.
The procedure is repeated until the entire population is ranked or a certain termination
36
condition is met. Figure 4-4 provides an example of rank assignment to six points. This
rank assignment scheme is also explained in Section 5.2.1.
f2 e(2)
d(2)
b(1)
f(3)
c(2)
a(1)
f1
Figure 4-4: Rank assignment for Pareto-based genetic algorithm
4.4.2.3 Multi-Attribute Genetic Algorithm
In Multi-Attribute Genetic Algorithm (MAGA) [22], each individual in a population
has an additional feature, i.e. attribute. The number of attributes is as many as that of
the objectives. In the crossover process, the individuals from each attribute are
recombined to generate offspring, which represents intermediate solutions that are not
totally optimal with respect to any single objective. During the execution of the
algorithm, the set of non-dominated solutions is updated until the iteration is
terminated
4.4.3 Discussions
Usually there are two methods to apply genetic algorithm for multi-objective
optimisation problems. The most common method, such as Pareto-based GA, is to
exert selection pressure during recombination process with respect to all the multiple
objectives other than one single objective. The other method, like multi-attribute GA,
37
embeds the information of multiple objectives into the GA parent strings and then
adopts suitable recombination operators to generate offspring.
Among the reviewed evolutionary algorithms, the performance of VEGA is somewhat
similar to that of weighing method. VEGA ignores the mutual dominance between
objective values during selection process, and it never selects candidates according to
trade-offs among objectives. At the end of the evolution, this approach ultimately
converges to a few points on the Pareto front and there is usually only one objective
gets extremely optimised.
The representative of different attributes in MAGA tries to fulfil different objectives so
that the offspring has a chance to be non-dominant. Since MAGA exerts very little
control on the evolution process, it is only used for performance comparison in this
thesis.
In Pareto-based GA, the solutions on the Pareto front are assigned the same rank. The
other solutions are assigned a lower rank with relation to the population density of the
region of solutions that dominate them. The candidate fitness is also represented by
ranks to determine selection probabilities for reproduction. By combining Pareto
dominance with partial preference information in the form of a goal vector, this
algorithm can perform local search. Moreover, with the technique of fitness sharing to
maintain the population diversity, the Pareto-based GA seems suitable to solve the
multi-objective optimisation problems. Actually, the Multi-Objective Genetic
Algorithm (MOGA) in this thesis is developed based on the Pareto-based GA (see
Section 5.2 for detail).
38
4.5 Summary
In this chapter, the mathematical definition of multi-objective optimisation problem is
presented and the preference structures posed on objective spaces are introduced. An
extensive review of the literature for multi-objective optimisation is also made. At the
end, the advantages and disadvantages of applying traditional approaches and
evolutionary approaches are discussed.
39
Chapter 5 Evolutionary Multi-Objective Optimisation Approaches
Chapter 5 Evolutionary Multi-Objective Optimisation
Approaches
5.1 Introduction
Population-based evolutionary algorithms can obtain multiple Pareto-optimal solutions
simultaneously for multi-objective optimisation problems. The literature review in
Section 4.4.2 shows that Pareto-Based GA performs selection based on not only the
objective values but also their mutual dominance, which distributes the population and
redirects the search towards the apparently most profitable areas. Three evolutionary
algorithms are applied in this thesis, which are namely: the Multi-Objective Genetic
Algorithm (MOGA), Multi-Objective Particle Swarm (MOPS) algorithm and Multi-
Objective Differential Evolution (MODE) algorithm. They utilise the same rank
assignment method as Pareto-Based GA to select the parents for evolution, so that the
obtained Pareto-optimal solutions spread uniformly along the Pareto-optimal front and
reflect the trade-offs among objectives. The fitness sharing technique is also adopted to
maintain the population diversity.
Despite the success of evolutionary algorithms in application to a variety of practical
problems, there remains a general concern as to whether a single evolutionary
algorithm is appropriate for any optimisation problem. For instance, GA achieves
much of its breadth, but it ignores the information of the trade-off, which can be
exploited to formulate effective selection and reproduction operators to generate the
Pareto-optimal set. Because GA also requires encoding to map from genotype
40
(problem domain) into phenotype (string representation), it suffers from the
discretisation of the search space for continuous problems. On the contrary, PS and DE
work directly on the continuous search space, which makes them suitable to solve
continuous problems. In order to improve the performance of different multi-objective
optimisation methods, MOGA is applied for go-circuit optimisation (discrete
problem), and MOPS and MODE are proposed for return-circuit optimisation
(continuous problem) in this thesis.
5.2 Multi-Objective Genetic Algorithm
5.2.1 Selection Processing with Rank Assignment
While retaining the same concept of crossover and mutation as the conventional GA,
MOGA, which is derived from the Pareto-Based GA, exerts a degree of control over
the selection process by applying the rank assignment method as developed in [16] for
the multi-objective problem. During the present evolution, candidates in the current
generation are compared by considering the following Pareto optimality conditions and
ranked to decide on their chances for survival in the next generation. First, all the non-
dominated candidates in the current generation are identified as rank 1 and marked.
Then the non-dominated candidates among the unmarked solutions are ranked as 2 and
marked. The solutions of rank 2 are of course inferior to those of rank 1. The above
rank assignment procedure is repeated recursively until the current rank number
reaches a pre-specified value or the size of the unmarked population becomes smaller
than a certain value. Afterwards, those candidates with a favourable rank (say, less
than 6) survive whereas candidates with a rank higher than the threshold (greater than
41
6) are discarded. The preserved candidates are then selected and recombined to
produce the subsequent generations. Meanwhile the non-dominated candidates among
the preserved ones are recorded in the Pareto-optimal set. Each time a non-dominated
solution is generated, the set of existing Pareto solutions will be updated. If any
member in the set of Pareto solutions does not dominate the new non-dominated
candidate, this candidate will be added to the set. On the other hand, any solution in
the set that is dominated by the added point will be eliminated from the set. This
generation-and-elimination procedure is iterated to locate the Pareto optimal points and
produce subsequent populations until the predefined maximum iteration or rank
number is reached. The final Pareto-optimal set comprises only non-dominated
candidates generated during the whole evolution procedure.
In conventional GA, if there are a few extraordinary individuals in a population and the
proportional selection scheme is applied, these individuals would take over a
significant portion of the finite population in a generation and undesirably result in
premature convergence. MOGA however uses rank-based selection method to avert
premature convergence caused by unusually strong individuals and speed up final
convergence when the candidate values tend to be the same.
5.2.2 Fitness Sharing
Conventional GA suffers from one inherent drawback of genetic drift [16], which
forces all candidate solutions into a few clusters rather than uniform scatter along the
Pareto front. In order to maintain appropriate diversity, the technique of fitness sharing
is used at the phenotypic level. Fitness sharing requires points that are close together to
share fitness and avoids an excessive number of individuals into one area. The multiple
42
sub-populations that arise along the Pareto front are thus stabilised and the obtained
multi-objective solutions are uniformly converged onto the Pareto front.
5.2.3 Variable Recombination Operators
In a typical GA, the crossover and mutation operators are the most important
components that influence the GA’s efficiency to give maximum exploration of the
search space and to ensure convergence towards global optima. To attain the best GA
performance, the crossover probability in MOGA is assigned a large value at the
beginning of the optimisation and linearly reduced in subsequently generations. The
mutation probability is varied in the opposite direction. The equations for changing
these probability values are:
Pt = P0 − (P0 − PT )*t /T
crossover crossover crossover crossover
(5-1)
Pt = P0 + ( PT − P0 )*t /T
mutation mutation mutation mutation
where T is the total number of iterations and t is the current iteration number.
Superscripts 0 and T denote probability values at the start and the end of search
procedure.
5.2.4 Treatment of Preferred Priorities among Objectives
In practice, the decision-maker may want to make preference to optimising certain
objectives. MOGA allows the decision-maker to exercise such preference. The rank
assignment method in Section 5.2.1 is modified as in the bi-criterion minimisation
example of Figure 5-1(b). Figure 5-1(a) shows the rank assignment for 6 candidate
solutions, where equal priorities are placed between minimising objective f1 and
43
minimising objective f2. Both solutions a and b are ranked as level 1 since f1(a)>f1(b)
but f2(a)<f2(b). Likewise, solutions c, d and e are ranked as level 2 and solution f is
ranked as level 3. Should the decision-maker wish to give a higher priority to
minimising the objective f2 than minimising the objective f1, candidate solutions a, c
and f are ranked at a higher level (more successful) than candidate solutions b, d and e.
For instance, candidate solution b is ranked as level 4 in Figure 5-1(b) instead of level
1 in Figure 5-1(a) because b has a larger f2 value than c and f. Candidate solutions a, c
and f will thus be selected for reproduction, and will have a higher probability of
survival in subsequent generations. At the end of the evolution, more solutions will
gather in the region of lower f2 values along the Pareto-optimal front, which means that
the decision-maker’s preference is satisfied and reflected to the distribution of the
obtained Pareto-optimal solutions.
f2 e(2) f2 e(6)
d(2) d(5)
b(1) b(4)
f(3) f(3)
c(2) c(2)
a(1) a(1)
f1 f1
(a) f2 has the same priority as f1 (b) f2 has greater priority than f1
Figure 5-1: Rank assignments for different priorities among objectives
44
5.3 Multi-Objective Particle Swarm Algorithm
5.3.1 Search Strategy
The Particle Swarm (PS) algorithm is another population-based method for optimising
hard numeric functions. The underlying motivation for its development is the metaphor
of individuals refining their knowledge by interacting with one another [27]. The
potential solution to the optimisation problem is represented as a particle moving in a
hyperspace. Each particle adjusts its search direction according to its previous best
position and the global best location, that is, the search direction is guided by the best
local and global solutions. Therefore, the key task of applying PS to multi-objective
optimisation is to determine the best solutions. The main difference between single-
and multi-objective optimisation rests on the fact that objectives are not guaranteed to
be comparable in the latter case. It seems reasonable to search for compromise
solutions in multi-objective problem rather than the best solution.
In a general multi-objective optimisation, the vector of F(x) contains a set of objective
functions fi(x), which is uniquely defined by the set of variable x. F(x) is to be
minimised (or maximised) individually:
min F ( x) = [ f 1( x), f 2( x), Λ , fn( x)]

x∈C (5-2)
subject to x∈D
where D denotes the decision space.
The utopian solution for Equation (5-2) is
45
~ ~ ~ ~
F = ( f1 , f 2 ,Κ , f n ) (5-3)
~
where f i = inf{ f i ( x)} .
For any solution Fˆ , the distance between Fˆ and the utopian point F~ can be
approximated by the metrics S(y,q,w), according to some specified norm lq-norm and a
weight vector w=(w1,w2,…,wn), w≥0 [25].
n
~ ~
S ( Fˆ , q, w) = Fˆ − F = [∑ wiq ( Fˆi − Fi ) q ]1 / q (5-4)
q ,w
i =1
The weight vector w reflects the different degrees of importance among the objectives.
The value of q chosen determines the specific way in which the distance between the
particle and the utopian point is measured. Frequently used measures for q are 1,2 and
∞. In this way, the compromise solution, which is nearest to the utopian solution, can
be obtained by minimising S(f(x),q,w) over the feasible region.
Once the global compromise solution is determined, the individual best solutions can
be captured. Similarly, the individual best solutions are denoted in the sense of
compromise. From all the locations visited by the particle so far, the position, which
has the smallest weighed distance from the global compromise solution, is set as the
individual compromise solution.
Different from the original version of PS, MOPS replaces the previous global best
solutions with the current global compromise solution. At each generation, the current
positions of all the particles are compared to the utopian solution and the compromise
location is calculated. The previous locations visited by the particle influence the
search for individual compromise solution instead of global compromise solution.
46
5.3.2 Rank-Based Selection
The rank-based selection in MOPS utilises the particle’s rank in the population to
determine its likelihood of producing offspring. First, the objective vector values of
each particle in the population are calculated. According to their relative dominance,
each particle is then assigned a corresponding rank. In a population of p individuals,
the assigned ranks are 1,2,3,…, Rmax, where Rmax≤p. The highest-ranked particles are
the closest to the Pareto-optimal set while the lowest-ranked ones are the farthest.
Given that the selection probability pb is proportional to rank, rank-based selection
favours points closest to the Pareto Set. [23]:
Rmax − rk + 1 Rmax − rk + 1
pbk = p
= p
(5-5)
∑ ( Rmax − r j + 1) p × ( Rmax + 1) − ∑ r j
j =1 j =1
where rk is the rank value of the k-th individual, and Rmax is the maximum rank value
among the whole population. The population size is represented as p.
The selected particles keep records of their current positions and velocities, as well as
their personal best points to affect their next position. With the addition of rank-based
selection, MOPS possesses a more exploitative search mechanism.
MOPS algorithm also allows the decision-maker to specify preference among
objectives. The rank assignment method is the same as in Section 5.2.5.
5.3.3 Weight Update
The objectives in the multi-objective optimisation are often non-commensurable. Since
the compromise solution in MOPS is scale-dependent, the weight vector w is added to
47
calculate the distance metrics. The weight value wi associated with the i-th objective
function consists of gradation element γ i and changing element µ i [25]:
wi = γ i ⋅ µ i
n
γ i = [ Ri ∑ 1 / R j ]−1 (5-6)
j =1
µ i = randomi
where Ri is the difference between the maximum and minimum values of the i-th
objective function and randomi is the random number between 0 and 1. n is the total
number of objective functions.
The gradation element γ i standardises the ranges for the different objective function
values. Applying the scalarising function S ( Fˆ , q, w) to determine a compromise
solution leads to explore efficient solutions only in the direction of the utopian
solution. Nevertheless, the changing element µ i may help to browse the border of the
Pareto-optimal set at the objective space level [25].
5.3.4 Pareto-Optimal Set Update
During population evolution, all the generated non-dominated candidates are checked
against the existing Pareto-optimal set. If any solution in the set does not dominate the
new generated non-dominated candidate, then the candidate is added to the set. In the
same way, all the solutions in the set that are Pareto dominated by the newly generated
candidates are eliminated.
48
5.4 Multi-Objective Differential Evolution Algorithm
Differential Evolution (DE) is another population-based algorithm efficient to solve
real-valued problems. The thoughts of extending DE to multi-objective optimisation
are the same as those embedded in MOPS. However, the Multi-Objective Differential
Evolution (MODE) algorithm works on the basis of the DE principles.
The optimisation parameters as well as the distribution of initial population influence
the performance of evolutionary algorithms. In this thesis, MOPS and MODE are
applied for go-circuit optimisation to demonstrate the effect of the optimisation
parameters.
5.5 Summary
In this chapter, descriptions of Pareto-based multi-objective optimisation algorithms,
namely MOGA, MOPS and MODE, are given. These algorithms employ Pareto
ranking to sort the individuals with respect to their mutual dominance as well as their
values. Pareto-based selection incorporates the decision-maker’s preference and
favours the individuals with acceptable ranks to survive throughout the evolution. The
use of fitness sharing further helps the non-dominated solutions distribute evenly along
the Pareto-optimal front.
Contrary to GA discretises the search space, MOPS and MODE search directly over
continuous feasible space. They introduce the concept of scalarising function to
replace best solutions with compromise solutions. With the employment of rank-based
selection, MOPS and MODE explore the Pareto-optimal front effectively.
49
Chapter 6 Results of Go-Circuit Optimisation
6.1 Optimal Traction-Substation Placements
At Stage I of the go-circuit optimisation (Section 2.3), the traction-substation
placements are optimised for a given operational timetable. A well-developed
combinatorial algorithm using Tabu Search [3] is applied to select 7 traction-
substations (5 of them with inverters) from the 16 passenger stations in the test MRT
system. The tabu tenure is set to be 5 and the optimisation runs for 100 iterations. The
time step of simulation is taken as one second. The headway for the UP- and DOWN
directions can be different, but are assumed identical in the study cases (= 180
seconds). The optimisation period is set equal to the same value, since the train
movement pattern repeats in each headway interval [3]. The synchronisation delay is
set to zero for convenience.
The algorithm will have a choice of two objectives as defined in Section 2.2, namely:
the overall energy consumption or load sharing among traction-substations. Two study
cases are formed:
Configuration 1: by optimising only the overall energy consumption
Configuration 2: by optimising only the load sharing among traction-substations
50
Configuration 1
Configuration 2
Initial Configuration
St10
St11
St12
St13
St14
St15
St0
St1
St2
St3
St4
St5
St6
St7
St8
St9
Passenger station Traction substation (Rectifier only)
Traction substation (Rectifier and Inverter)
Figure 6-1: Optimised configurations
The optimised traction-substation placements of the 2 study cases are given in Figure
6-1. In order to compare the nature of energy consumption and load sharing
optimisation, Figures 6-2 and 6-3 display the convergence plots of the optimised
objectives, along with the non-optimised objectives. During obtaining Configuration 1,
where only energy consumption is optimised, the energy consumption drops from
12.41 MW to 12.39 MW but the load sharing value occasionally increases and form a
glitch in Figure 6-3. On the other hand, to obtain Configuration 2, where only load
sharing is optimised, the load sharing decreases dramatically from 6.69 MW to 5.71
MW whereas the energy consumption keeps rising in Figure 6-2. It is also revealed
from Figure 6-2 and 6-3 that energy consumption and load sharing are unlikely to be
optimised at the same time.
51
12.47
Energy Consumption (MW)

12.46 Configuration 2
12.45
12.44
12.43
12.42
12.41
12.40
Configuration 1
12.39
12.38
0 20 40 60 80 100
Iteration
Figure 6-2: Energy consumption convergence curve
7.00
6.80
Load Sharing (MW)
6.60
6.20
6.00
5.60
0 20 40 60 80 100
Iteration
Figure 6-3: Load sharing convergence curve
As expected, the optimised objectives, energy consumption for Configuration 1 and
load sharing for Configuration 2, receive more improvement than the non-optimised
objectives. The improvement of each objective during optimisation is expressed by:
Initial value −Optimised value

improvement = × 100% (6-1)
Initial value
52
The final improvements of the two study cases (Configuration 1 and Configuration 2)
are illustrated in Table 6-1.
Objective function Initial value Optimised value Improvement
Energy consumption (MW) 12.41 12.39 1.6 %
Load sharing (MW) 6.69 5.71 14.7 %
Table 6-1: Improvement of optimised energy consumption and load sharing
6.2 Worst-Case Scenarios of Operational Deviations
During operation, operational parameters such as the headway, synchronisation delay
and dwell time deviate continuously from their nominal values. Stage II of the go-
circuit optimisation assesses the impact of these deviations, and evaluates the
performance deviations from those of the optimal traction-substation placements of
Stage I. The worst-case scenarios are thus identified. In contrast to Stage I, which is
single-objective optimisation, the optimisation here operates on both objectives of
Stage I (overall energy consumption and load sharing among traction-substations).
The % deterioration is expressed by:
Worst − case objective−Optimised objective from Stage I

Deterioration = × 100%
Optimised objective from Stage I
(6-2)
The two multi-objective evolutionary algorithms, MODE1 and MOPS, as outlined in
Chapter 5, are used for the bi-objective optimisation. The optimisation program is
written in Visual C++ and takes about 30 minutes with a population of 300 and a
generation of 20. As there are 14 intermediate passenger stations on each UP- or
53
DOWN-track, the two algorithms each takes 30 deviation variables (namely: headway
+ synchronisation delay + 14 dwell times x 2 tracks). Parameter limits that restrict
these variables during optimisation are listed in Table 6-2. The non-dominated
solutions for the worst-case operational scenarios for the two study cases
(Configuration 1 and Configuration 2) are summarised in Figures 6-4 and 6-5.
Optimisation variables Value range (second) for Velocity limit for

MODE1 and MOPS MOPS
Headway (90,420) 30
Synchronisation delay (0,headway) 20
Dwell times (10,120) 10
Table 6-2: Parameter limits for bi-criterion optimisation
16.00
15.00 MODE
Load sharing (MW)
14.00 MOPS
13.00
12.00
11.00
10.00
9.00
8.00
17.50 18.00 18.50 19.00 19.50 20.00
Energy consumption (MW)
Figure 6-4: Pareto-optimal set for Configuration 1
54
14.00
MODE
13.00
MOPS
Load sharing (MW)

12.00
11.00
10.00
9.00
8.00
7.00
6.00
16.50 17.00 17.50 18.00 18.50
Energy consumption (MW)
Figure 6-5: Pareto-optimal set for Configuration 2
Tables 6-3 and 6-4 capture the maximum deviations of the two objectives for the non-
dominated solutions as in Figures 6-4 and 6-5. These are expressed in percentage
according to Equation (6-2).
Case Maximum energy consumption Deterioration for energy

(MW) consumption
1.1 19.48 57.2 %
2.1 18.34 48.0 %
Table 6-3: Maximum deviation for energy consumption
Case Maximum load sharing (MW) Deterioration for load sharing
1.2 13.32 133.3 %
2.2 11.15 95.3 %
Table 6-4: Maximum deviation for load sharing
Due to their different search strategies, MODE1 and MOPS each produces solutions
exhibiting a different distribution. This point is demonstrated in Figure 6-4, where the
55
solutions generated by MODE are close together in one area whereas the solutions
produced by MOPS uniformly scatter along the Pareto front. In this sense, the
performance of MOPS is better than that of MODE. As seen from Figures 6-4 and 6-5,
the solutions produced for Configuration 1 have a wider spread than Configuration 2
for both energy consumption and load sharing.
6.3 Performance Check for Failure Conditions
Occasionally some of the rectifiers and inverters, or even the traction-substation may
be out of service. The MRT system should maintain acceptable operation from such
failure conditions. To verify the non-dominated solutions (Section 6.2) obtained for
Configuration 1 and Configuration 2, performance check is carried out at Stage III of
go-circuit optimisation.
A list of 16 credible failure conditions is conducted on the test MRT system. Table 6-5
summarises the performance results for case 1.1 (the non-dominated solution with
maximum energy consumption under Configuration 1). Under abnormal conditions,
the lowest train voltage is required to be above 900V. Hence, the design requirement
for case 1.1 is acceptable, as the lowest train voltage of all the 16 testing scenarios is
about 1000V.
The performance check results for case 2.2 (the non-dominated solution with minimum
load sharing under Configuration 2) are shown in Table 6-6. When substation St0 is
out of service, the lowest train voltage is about 800V. In this case, either the
corresponding operational timetable for Case 2.2 should be avoided during system
56
operation, or the design parameters (e.g. traction-substation placements) need to be
revised.
Failure Substation Energy consumption Load sharing Lowest train

case outage (MW) (MW) voltage (V)
1 None 17.67 13.32 1301.8
2 St0 1Rec 17.67 13.33 1301.1
3 St3 1Rec 17.67 13.33 1299.7
4 St6 1Rec 17.63 13.17 1287.7
5 St7 1Rec 17.68 13.36 1269.4
6 St10 1Rec 17.69 13.21 1237.7
7 St14 1Rec 17.23 12.90 1295.9
8 St15 1Rec 17.67 13.35 1299.0
9 St0 Inv 18.08 13.04 1295.9
10 St0 17.27 11.90 1293.2
11 St3 17.25 10.25 1174.1
12 St6 15.66 11.14 1257.5
13 St7 17.85 10.17 1069.7
14 St10 17.17 11.94 1146.6
15 St14 16.31 12.41 1271.4
16 St15 17.05 11.87 1292.6
Table 6-5: Performance check results for case 1.1
57
Failure Substation Energy consumption Load sharing Lowest train

outage (MW) (MW) voltage (V)
1 None 16.72 11.15 1209.0
2 St0 1Rec 16.73 11.37 1132.0
3 St5 1Rec 16.81 10.99 1116.9
4 St6 1Rec 16.83 11.17 1140.4
5 St9 1Rec 16.48 10.97 1140.9
6 St11 1Rec 16.76 11.59 1206.4
7 St12 1Rec 16.71 11.33 1208.1
8 St15 1Rec 16.55 11.29 1208.7
9 St0 Inv 17.40 10.57 1209.0
10 St0 18.39 12.11 850.34
11 St5 17.12 11.21 1042.4
12 St6 16.84 11.71 1018.9
13 St9 15.74 10.53 1163.0
14 St11 16.46 10.96 1182.6
15 St12 16.45 10.19 1206.1
16 St15 15.99 8.97 1208.0
Table 6-6: Performance check results for case 2.2
6.4 Summary
In the three-stage scheme of go-circuit optimisation described in Section 2.3, a
previously reported TS algorithm is first employed to optimise traction-substation
placements, given a fixed operational timetable, so that either minimal energy
58
consumption or equal load sharing is achieved. The multi-objective evolutionary
algorithms, MOPS and MODE1 proposed in Section 5.3, are then applied to attain
trade-off between energy consumption and load sharing in the go-circuit. Operational
timetable, namely headway, synchronisation delay and dwell time, is varied to identify
the worst-case scenarios of operational deviations. From the obtained Pareto-optimal
set, the decision-maker picks up some solutions for performance check. If the selected
solutions fail the performance check, the corresponding operational timetable should
be avoided, or Stage I should be re-run to obtain other optimal traction-substation
placements. In this way, the safety and efficiency of system operation are ensured.
Simulation results show the effectiveness of MOPS and MODE1 to identify the
Pareto-optimal solutions, and the robustness of the presented three-stage scheme. As
MOPS produces non-dominated solutions evenly spread along the Pareto-optimal front
and the solutions generated by MODE1 gather in one area, MOPS outperforms
MODE1 on the go-circuit optimisation.
59
Chapter 7 Results of Return-Circuit Optimisation
7.1 Multi-Objective Optimisation for Normal Conditions
Different from the go-circuit optimisation, which configures traction substation
locations to obtain bi-criterion optimisation of energy consumption and load sharing,
the return-circuit optimisation explores all possible earthing and bonding arrangements
at the fixed passenger stations and traction substations to seek compromise solutions
between touch voltage and stray current. In this sense, the return-circuit and go-circuit
can be each optimised independently. Moreover, unlike single-objective optimisation,
none of the solutions for multi-objective optimisation is optimal for each objective.
Instead, the non-dominated solutions are superior to the rest of solutions in the search
space with all the objectives considered. Different solution is selected from the Pareto-
optimal set corresponding to the user’s preference. Based on such considerations, the
initial system configuration in Figure 6-1 is also used in this chapter for return-circuit
optimisation.
The test MRT system in this chapter has two tracks as in Figure 7-1. There are 16
passenger stations, 7 of which are chosen as TSSs. The passenger stations are located
at an interval of 1∼2 kms. At each passenger station, the pair of running rails carrying
each train is bonded (known as cross bonding) to reduce the impedance of the return-
circuit. However the pair of running rails has the option of being bonded to the other
pair of running rails in the opposite direction (interbound bonding). Since all trains are
dispatched with the same pre-specified headway and synchronisation delay, it is
60
sufficient to use the headway interval as the period of the optimisation [3]. The time
step of simulation is set to be one second. It is also assumed that the stray current path
resistance, which changes with many practical factors, is known and can be varied later
for scenario studies.
St10
St11
St12
St13
St14
St15
St0
St1
St2
St3
St4
St5
St6
St7
St8
St9
Passenger station Traction substation (Rectifier only)
Traction substation (Rectifier and Inverter)
Figure 7-1: Layout of study system
As Section 3.3 shows, earthing and bonding strategies have great impact on the touch
voltage and stray current in the return-circuit. To gain further insight, a simulation
program is written in Visual C++ to carry out simulations for a variety of earthing and
bonding arrangements at passenger stations and traction substations. The
corresponding touch voltage and stray current listed in Table 7-1 have exhibited
certain patterns of variations for the given earthed and unearthed arrangement.
Moreover, Figure 7-2 [8] reveals that floating the return pole reduces the touch voltage
to roughly half as compared to a solidly earthed arrangement. The floating
arrangement however makes the low-level stray current difficult to detect and clear.
The bonding of rails provides adequate path for return traction current, decreases the
mutual resistance and thus increases the traction current in return path.
61
Simu- Simulation condition Positive Negative Stray

lation touch touch current
No. voltage (V) voltage (V) integral
(Ah)
1 As default 74.13 -80.75 41.62
2 All PSs direct earth, bonding as 67.04 -78.83 156.94
default
3 All PSs diode earth, bonding as 98.34 -58.53 83.10
default
4 All PSs interbound bond, 74.11 -80.64 41.84
earthing as default
5 All TSSs direct earth, bonding as 77.49 -76.55 82.78
default
6 All TSSs diode earth, bonding as 93.91 -61.66 67.31
default
7 All TSSs interbound bond, 74.14 -80.59 41.84
earthing as default
8 All PSs direct earth & 78.03 -65.74 202.85
interbound bond
9 All PSs diode earth & 98.20 -57.82 83.02
interbound bond
10 All TSSs direct earth & 87.79 -63.94 107.55
interbound bond
11 All TSSs diode earth & 93.81 -57.82 67.36
interbound bond
12 All TSSs direct earth and all PSs 85.55 -70.05 104.01
diode earth, bonding as default
13 All TSSs diode earth and all PSs 74.61 -70.01 150.34
direct earth, bonding as default
14 All TSSs interbound bond and 64.80 -83.36 117.56
all PSs direct earth
15 All TSSs interbound bond and 90.97 -69.47 62.78
all PSs diode earth
16 All TSSs direct earth and all PSs 77.41 -76.72 82.89
interbound bond
17 All TSSs diode earth and all PSs 93.83 -60.54 67.36
interbound bond
Note: As default, all TSSs and PSs are not provided with interbound bonding and are
floating.
Table 7-1: Typical arrangements of earthing and bonding
62
Catenary wire
+
TSS -
Train
Running rail
2Vmax
Earthed System
Vmax
Floating System
0V
-Vmax
Figure 7-2: Touch voltage distribution with different earthing arrangement
Generally speaking, both the touch voltage and the average stray current tend to
increase when the headway is decreased. The headway used in this study is 180s,
which is the headway most adopted during the peak hours.
MRT systems have been provided with total floating, direct earth or diode earth
arrangement [8]. The return-circuit simulation also allows a mix of different earthing
arrangements at either the TSSs or passenger stations. To best improve the touch
voltage and stray current, Multi-Objective Genetic Algorithm (MOGA) explained in
Section 5.2 is executed to optimise the three objectives: the maximum positive touch
voltage (f1), the minimum negative touch voltage (f2) and the total integral stray current
collected from all earthing points (f3). The function f2 is chosen due to the
unidirectional characteristic of the diode earth.
63
The proposed multi-objective optimisation is carried out for two configurations:
Configuration 1: earthing and bonding taken at all TSSs only
Configuration 2: earthing and bonding taken at all passenger stations (PSs)
Since the study MRT system has 16 passenger stations, a GA string of 48 bits (16 x 3
bits substrings) is formed for Configuration 2 to represent the earthing and interbound
bonding at all 16 passenger stations. It is assumed that all passenger stations are cross-
bonded. Each substring consists of 3 binary bits. The first 2 bits are used for
identifying the earthing strategy, which can be floating, directly earthed or diode
earthed. The last bit is set to either one or zero for providing either or not interbound
bonding at each passenger station. Likewise, the GA uses a string of 21 bits for the 7
TSSs in Configuration 1. The computer program for the multi-objective optimisation
of touch voltage and stray current is written in Visual C++. The program takes
typically 50 minutes for the optimisation with a generation number of 5 and a
population of 100.
64
140
stray current integral 120
100
80
60
40 -50
50 -60
60 e
tag
posi 70 vol
-70
tive 80 ch
touc
h vo 90
-80
e tou
ltage ativ
neg
100 -90
(a) Equal priority assigned to minimum touch voltage and stray-current integral
o positive touch voltage

has highest priority
x negative touch voltage
120 + stray current integral
100
stray current integral
80
60
40 -40
50
60 -60 e
70 tag
posi -80 h vol
80 c
tive
touc tou
h vo 90
ative
ltage 100 -100 neg
(b) Different priority assigned to minimum touch voltage and stray-current integral
Figure 7-3: Pareto-optimal sets for Configuration 1
65
stray current integral 80
70
60
50
40 -60
60
70 -70
ge
posi 80 -80 volta
tive uch
touc
h vo
90
tive to
ltage 100 -90 a
neg
(a) Equal priority assigned to minimum touch voltage and stray-current integral
o positive touch voltage

x negative touch voltage
+ stray current integral
90
80
stray current integral
70
60
50
40 -60
50
60 -70
ge
posi 70
volta
tive -80 ch
touc 80 tou
h vo
ltage
90
ative
100 -90 neg
(b) Different priority assigned to minimum touch voltage and stray-current integral
Figure 7-4: Pareto-optimal sets for Configuration 2
66
As shown in Figure 7-3 and Figure 7-4, the decision-maker is presented with four
Pareto-optimal sets, which contains the solutions with equal and different priorities for
Configuration 1 and Configuration 2. The solutions contained in each Set represent a
different trade-off among the three objective functions. If the positive touch voltage
(f1) takes priority over the other two objectives, then more solutions (“ο” in Figures 7-
3(b) and 7-4(b)) gather in the region of large positive touch voltage values. Similar
observation is made if priority is assigned to the other two objectives. From Figures 7-
3(b) and 7-4(b), three extreme solutions are picked from the vertices of the plots for
Configuration 1 and Configuration 2. Tables 7-2 and Table 7-3 each contains the three
extreme solutions together with a well-balanced solution obtained in the case of equal
priority on the three objectives.
Solution Priorities on f1, f2 Lowest Lowest Lowest stray

and f3 positive touch negative touch current integral
voltage (V) voltage (V) (Ah)
1.1 Same priority for all 64.83 -60.26 42.89
1.2 Highest priority on f1 59.52 -64.79 41.80
Table 7-2: Multi-objective optimisation of earthing & bonding for configuration 1
67
Solution Priorities on f1, f2 Lowest Lowest Lowest stray

and f3 positive touch negative touch current
voltage(V) voltage(V) integral(Ah)
2.1 Same priority for all 53.28 -57.57 51.38
Table 7-3: Multi-objective optimisation of earthing & bonding for configuration 2
By comparing the four solutions in Table 7-2 with the four solutions in Table 7-3,
Configuration 2 appears to carry higher stray current (f3) than Configuration 1. The
former has however outperformed the latter from the point of view of both f1 and f2.
This appears reasonable as the former provides more earthing and bonding locations.
Since both Configuration 1 and Configuration 2 are optimised, the solutions in Tables
7-2 and 7-3 have out-performed the typical cases as in Table 7-1.
7.2 Performance Check for Failure Conditions
After obtaining the Pareto-optimal sets, the performances of non-dominated solutions
are further verified under some credible failure conditions. Table 7-4 shows the
performance checks on a well-balanced solution of Configuration 2 as listed in the first
row of Table 7-3. The list of ten credible failure conditions is given in the second
column. Almost all the three objectives have deteriorated drastically due to outage of a
TSS or a rectifier at TSS. However, the optimisation solution under test yields the
lowest train voltage of more than 1000V among the ten failure cases. As expected, the
68
return-circuit has little impact on train voltage because only leakage current is involved
in the return-circuit. The train voltages are required to be above 900V under abnormal
conditions, so the selected solution performs well and the corresponding earthing and
bonding arrangements should be applicable for system operation.
Failure Substation Positive touch Negative touch Stray-current Lowest train

Cases Outage voltage (V) voltage (V) integral (Ah) voltage (V)
1 None 53.28 -89.41 86.92 1213.0
2 St0 1Rec 55.53 -91.18 87.49 1212.0
3 St2 1Rec 66.89 -95.60 90.72 1153.3
4 St5 1Rec 62.32 -103.46 95.04 1151.1
5 St8 1Rec 63.54 -104.10 102.38 1128.1
6 St11 1Rec 53.23 -94.06 92.05 1163.7
7 St13 1Rec 55.51 -82.89 88.71 1205.2
8 St15 1Rec 53.28 -89.41 87.39 1211.2
9 St0 2Rec 53.28 -83.74 92.91 1216.2
10 St0 95.35 -107.78 94.65 1080.4
Table 7-4: Performance check for case 2.1
7.3 Summary
The two-stage scheme of return-circuit optimisation outlined in Section 3.4 is
implemented in this chapter to attain trade-off between the touch voltage and stray
current in MRT systems. After further discussing the influence of earthing and
bonding arrangements on touch voltage and stray current, MOGA described in Section
69
5.2 is applied, and the method of Pareto-optimal set is developed to best improve the
touch voltage and stray current for normal condition. Care is taken to ensure
comprehensive and uniform spread of optimal solutions within the set. The decision-
maker is then given a powerful tool for picking the most appropriate earthing and
bonding design from the set and for identifying the worst-case performance from the
list of credible failure conditions. Simulation results indicate that MOGA successfully
traces the Pareto-optimal solutions. The process of performance check helps maintain
the reliability of system operation.
70
Chapter 8 Conclusions
8.1 Summary and Conclusions
In MRT system design and operation, several competing objectives are always
required for optimisation in the context of different set of alternatives. A three-stage
scheme is accordingly put forward in this thesis to implement go-circuit and return-
circuit optimisation. At the first stage, the basic deign is obtained by optimising the
primary variables. At the second stage, multi-objective evolutionary algorithms are
applied and the trade-off among conflicting objectives is attained. As the varying
running parameters influence the system performance, the second stage takes into
account the operational deviation and identifies the worst-case of the basic system
design derived from the first stage. The need for changing the basic design to cater for
both the normal condition and failure conditions is further ascertained by performance
check at the third stage.
Energy consumption and load sharing have been adopted for measuring the cost-
effectiveness of MRT operation. However, minimum energy consumption and equal
load sharing cannot be achieved simultaneously. Energy consumption and load sharing
are greatly influenced by traction substation placements and system operational
timetable. Therefore, at Stage I of go-circuit optimisation, traction substation
placements are configured for predefined operational timetable by employing Tabu
Search (TS) algorithm, so that minimum energy consumption or equal load sharing is
achieved. At Stage II, Multi-Objective Particle Swarm (MOPS) and Multi-Objective
71
Differential Evolution (MODE) algorithms are used to determine the effect of
operational timetable, i.e., the headway, synchronisation delay and dwell times. The
worst-case energy consumption and load sharing are then identified for the
performance check of Stage III.
As running rails are used as traction current return conductors, there are increasing
concerns on touch voltage and stray current in the return-circuit. The mitigation of
both touch voltage and stray current is one kind of conflicting multi-objective. As seen
from the return-circuit model, earthing and bonding strategies have a profound impact
on touch voltage and stray current. At Stage I of return-circuit optimisation, Multi-
Objective Genetic Algorithm (MOGA) is employed to best improve touch voltage and
stray current under normal condition for a variety of earthing and bonding
arrangements at passenger stations and/or traction substations. At Stage II, the most
appropriate earthing and bonding policies are picked up for performance check from
the list of credible failure conditions.
Simulation results for the above go-circuit optimisation and return-circuit optimisation
indicate that the proposed multi-objective evolutionary algorithms take full advantage
of the problem nature and are very effective to provide the decision-maker with a
comprehensive Pareto-optimal set to choose from. In terms of the uniform spread of
the non-dominated solutions along the Pareto-optimal front, MOPS outperforms
MODE1 on the go-circuit optimisation.
Characteristic of multi-objective optimisation is seldom unique but a finite set of non-
dominated or Pareto-optimal solutions. Traditional multi-objective optimisation
techniques, such as the weighing methods, cannot capture the non-dominated solutions
simultaneously. By working with a population, evolutionary algorithms have a built-in
72
advantage of being able to deal with multi-objective optimisation. Despite GA’s great
success on optimising a variety of problems, GA requires encoding to map from
problem domain to string presentation and thus suffers from the discretisation of
search space for continuous problem. On the contrary, PS and DE work directly on the
continuous search space. In order to exploit the merits of different algorithms, three
evolutionary algorithms are used in this thesis for different problems, which are
namely: MOGA for return-circuit optimisation (discrete problem), MOPS and MODE
for go-circuit optimisation (continuous problem). Especially, MOPS and MODE are
proposed to solve the MRT problems for the first time. These presented algorithms are
based on three concepts of population, rank-based selection and competitive evolution.
At the heart of MOPS and MODE is the replacement of the best solution with
compromise solution, as multiple objectives are usually non-commensurable. Also of
importance to the proposed algorithms is the adoption of Pareto ranking to successfully
sort candidates according to the values of several objectives and even to the decision-
makers’ preference. The individuals with favoured ranks have higher probability to
survive throughout the evolution procedure. Based on Pareto ranking and a
competitive evolution scheme, namely rank-based selection, the proposed algorithms
determine a direction of improvement and evolve the entire population towards the
Pareto-optimal set.
In the development stage, MOPS and MODE are compared to another type of
algorithm, Multi-Attribute Genetic Algorithm (MAGA), in that all of them use a
population of points and competitive evolution to direct the population towards the
optima. By properly adjusting the algorithm parameters, MOPS and MODE perform
quite well on the test numerical problems. Not only the concave trade-off curve but
73
also the discontinuous trade-off curve is successfully reached with a relatively small
population and generation number.
8.2 Suggestions for Future Work
The generation of Pareto-optimal set and the selection of preferred solutions from the
Pareto-optimal set are two critical issues for multi-objective optimisation problems.
The first issue has been well addressed in literatures and discussed in detail in this
thesis. The second issue, however, needs more studies since the decision-maker is still
left with a challenge to pick up the preferred solutions from so large a Pareto-optimal
set.
To end this thesis, the author would like to mention two interesting as well as
important topics for future research:
• How should the Pareto-optimal set be reduced to a manageable size, so as to aid
the decision-maker in making a choice from a number of Pareto-optimal solutions?
• How should the solutions generated by the multi-objective evolutionary algorithms
best approximate the Pareto-optimal front?
74
References
References
[1] E.D. Sunde, "Earth conduction effects in transmission systems", Dover

publication Inc., 1968
[2] S. D. Jacimovic, "Maximum permissible values of step and touch voltages with
special consideration to electrified railroads", IEEE Trans. Industry Application,
Vol. IA-20, No. 4, 1984
[3] C.S. Chang, J.S. Low and D. Srinivasan, “Application of Tabu Search in Optimal
Design and Operation of MRT Power Supply System”, IEE Proceedings,
Electric Power Applications, No.1, 1999
[4] C.S. Chang, T.T. Chan and K.K. Lee, "Network switching and voltage
evaluation using an expert system in AC railway systems", IEE Proceedings,
Electric Power Applications, Vol.139, No.1, 1992
[5] C.S. Chang, W. Wang. A.C. Liew and F.S. Wen, “Bicriterion optimisation for
traction substations in mass rapid transit systems using genetic algorithm”, IEE
Proceedings, Electric Power Applications, Vol.145, No.1, 1998
[6] James B. Forsythe, “Light Rail/Rapid Transit: New Approaches for the
Evaluation of Energy Savings, Part I-Life-Cycle Cost from Synthetic
Routes/Operational Models”, IEEE trans. Industry Applications, Vol. IA-16,
No.5, 1980
[7] J.G. Yu and C.J. Goodman, "Modelling of rail potential rise and leakage current
in DC rail transit systems", IEE colloquium on stray current effects of DC
railways and tramways, 1990
[8] J.G. Yu, "The effects of earthing strategies on rail potential and stray currents in
DC transit railways", International conference on developments in mass transit
systems, 1998
75
References
[9] Y. Cai, M.R. Irving, S.H. Case, "Modelling and numerical solution of
multibranched DC rail traction power systems”, IEE Proceedings, Electric Power
Applications, Vol.142, No.5, 1995
[10] C.S. Chang and L. Tian,” Worst-case identification of touch voltage and stray
current of DC railway system using genetic algorithm”, IEE Proceedings,
Electric Power Applications, Vol. 146, No. 5, 1999
[11] K.S. Bahra and P. G.Batty, "Earthing and bonding of electrified railways",
International conference on developments in mass transit systems, 1998
[12] W. Sidoriak, "Rail isolation on the Baltimore central light rail line", Materials
Performance, Vol.32, No.7, 1993
[13] S.R. Farnsworth and N. J. Moriber, "MAX: Porland's light rail system-control of
stray currents at the source", Materials Performance, Vol.29, No.7, 1990
[14] P.L. Yu, “Multiple criteria decision making: five basic concepts”, Handbooks in
Operations Research and Management Science, Vol. 1 (Optimisation), 1989
[15] G. Colson and C.de Bruyn, Models and methods in multiple criteria decision
making, 1989
[16] D.E. Goldberg, Genetic algorithms in search, optimisation, and machine

learning, Addison-Wesley, 1989
[17] H. Tamaki, H. Kita and S. Kobayashi, "Multi-objective optimisation by genetic

algorithms: a review", IEEE international conference on Evolutionary
computation, Japan, 1996
[18] C.M. Fouseca and P.J. Fleming, “Genetic algorithms for multi-objective
optimisation: formulation, discussion and generalisation”, Proceedings of the
fifth international conference on genetic algorithms, 1993
[19] K. DeJong, "Learning with genetic algorithm: an overview", Machine Learning,

Kluwer academic, 1988
76
References
[20] K.C. Tan and Y. Li, "Multi-objective genetic algorithm based time and
frequency domain design unification of linear control systems", Technical
Report CSC-97007, University of Glasgow, 1997
[21] G. Ambrose, D.R. Hansen and L. Duckstein, Multiobjective decision analysis

with engineering and business applications, Wiley, 1982
[22] Joanna Lis and A.E.Eiben, “A multi-sexual genetic algorithm for multi-objective
optimisation”, Proceedings of 1997 IEEE international Conference on
Evolutionary Computation
[23] Patrice O. Yapo, A multiobjective global optimisation algorithm with application

to calibration of hydrologic models, Ph.D Thesis, University of Arizona, 1996
[24] Srinivas, N. and Deb, K., “Multiobjective optimisation using non-dominated

sorting in genetic algorithms”, Evolutionary Computation, Vol.2, No.3, 1994
[25] X. Gandibleux, N. Mezdaoui, A. Freville, “A Tabu search procedure to solve

multiobjective combinatorial optimisation problems”, Lecture notes in
Economics and Mathematical System, Vol. 455
[26] Storn R. and Proce K., “Minimising the real functions of the ICEC’96 contest by
differential evolution”, International Conference on Evolutionary Computation,
Japan, 1996
[27] J. Kennedy, “The particle swarm: social adaptation of knowledge”, Proceedings

of 1997 IEEE international Conference on Evolutionary Computation
[28] R. C. Eberhart and Y. Shi, “Comparison between genetic algorithms and particle
swarm optimisation”, Proceedings of 1998 International Conference on
Evolutionary Programming, 1998
[29] H. Fujita and H. Akagi, "An approach to harmonic current-free AC/DC power
conversion for large industrial loads: the integration of a series active filters with
a double-series diode rectifier", IEEE Transactions on Industry Applications,
Vol.33, 1997
77
References
[30] Fan Wang, “Genetic algorithm based harmonic evaluation and filter placement
for a rapid transit system”, Master’s Thesis, National University of Singapore,
1999
78
Appendix A Evolutionary Algorithms
A.1 Introduction
Evolutionary Algorithms are a class of stochastic search and optimisation methods that
include Genetic Algorithm (GA), evolutionary programming, Particle Swarm (PS)
algorithm, Differential Evolution (DE) algorithm, and their variants. They seldom
require much auxiliary information to search for better candidates, and are robust and
suitable for finding optima effectively with a small probability of falling in the local
optima. Increasing number of engineering optimisation problems have been
successfully solved by evolutionary algorithms.
Evolutionary algorithms work on optimisation problems by maintaining a population
of individuals and incorporating random variation and selection for iterations. Each
individual represents a potential solution to the problem at hand. In each generation, a
fitness value is first assigned to each offspring. Depending on its fitness, each
population member is then given a specific survival probability. The evolution
converges after some number of generations. In most cases, the best individual
represents a near-optimum (reasonable) solution.
A.2 Genetic Algorithm
The Genetic Algorithm (GA) was motivated by ideas from natural genetics. GA starts
with a population of chromosomes, the abstract representations of candidate solutions.
79
Evaluations of chromosomes in the current generation are based on the problem-
dependent fitness function. Those chromosomes with higher fitness are selectively
picked for reproduction. By employing crossover and mutation, a low probability
operator and a high probability operator respectively, information encoded in these
selected chromosomes is recombined. Successive populations are generated to form
the subsequent generation. In this way, GA attempts to find all the optima in the search
space and realise the Darwinian notion of competitive evolution.
The key feature of the GA is its ability to exploit accumulating information about an
initially unknown search space so as to bias subsequent search into the useful
subspace. Moreover, GA operates on several solutions at the same time, gathering
information from the current points to direct subsequent search. Its merit to maintain
multiple solutions concurrently makes GA less susceptible to local optima or noises
[19].
A.3 Particle Swarm and Differential Evolution Algorithm
Particle Swarm (PS) and Differential Evolution (DE), as novel population-based
evolutionary computation techniques, are efficient to optimise continuous problems.
Instead of using genetic operators, PS and DE embody explicitly or incorporate
implicitly three key operations of GA: crossover, mutation and fitness evaluation. The
fitness evaluation merely depends on the problem structure, but the crossover and
mutation processes greatly influence the effectiveness of optimisation algorithms.
Both PS and DE replace GA’s traditional bit-inversion mutation scheme with a method
that perturbs real-valued vectors with population-derived noise and make mutation an
80
adaptive procedure [28]. Besides their good convergence properties and suitability for
parallelism, PS and DE are simple to operate. They work on only a few control
variables that remain constant throughout the entire optimisation procedure.
A.3.1 Particle Swarm Algorithm
The PS algorithm, introduced by Eberhart and Kennedy [27], was inspired from the
simulation of social behaviour. The particles in PS are represented as multidimensional
points. During the optimisation procedure, they evolve towards the global and local
best positions. That is, each particle’s searching trajectory is influenced by its own and
the others’ exploring experiences.
Each particle in the population keeps a record of its current position, velocity, and its
best position found so far. Specifically, a particle is manipulated using the following
equation:
vi = w * vi + α * ( p i − xi ) + β * ( p g − xi )
(A-1)
xi = xi + vi
where xi is the position of the ith member of the population and vi is the current
velocity of the individual. The local best for the ith individual is denoted by pi and
p g is the global best position. α and β are random numbers in the range of [0,1].
They reflect the degree of global best and local best to guide the particle’s subsequent
search. The inertia weight w is used here to control the impact of the previous history
of velocities on the current velocity, thereby influencing the trade-off between global
exploration and local exploration ability of the particle.
81
At the beginning of the optimisation procedure, a population of particles is initialised
with random position x and velocity v. Each particle is then evaluated at the time step,
i.e., each position vector xi is calculated to obtain the fitness value for the particle. The
best position vector for each particle is accordingly adjusted, and the fitness value of
each particle is compared with the best fitness found so far in the population to identify
the best-performing particle.
Particle swarm optimisation does not have crossover and mutation operation, but the
concepts are characterised. Unlike GA, an individual (particle) in the population does
not explicitly exchange genetic information between a few randomly selected
individuals. There is no parts (chromosomes in GA) divided in each particle and its
evolution is performed as a whole. PS uses a highly directional mutation operation,
which changes the velocity of the particles between the local best and global best. In
this way, any point in the search space is eventually accessed if there is enough
iteration or sufficiently large velocity limit [28].
A.3.2 Differential Evolution Algorithm
Differential Evolution (DE) algorithm developed by Stron and Price is the best genetic
type of algorithm for solving the real-valued test function suite of the first International
Contest on Evolutionary Computation [26].
The performance of DE is dependent on three variables: the population size NP, the
mutation-scaling factor F and the crossover constant CR. A new scheme is used in DE
to generate variation vector in the population, which is totally different from other
evolutionary algorithms except PS. The two or four individuals in the population are
82
randomly selected and their difference vector is taken as the variation vector. In this
way, the direction and distance information is extracted from the population to
generate random deviations, which results in the excellent convergence of the
population. For the ith individual in the population, the variation vector ∆xi is
produced as in the form:
∆xi = x j + F * [( x p − x q ) + ( x m − x n )] (A-2)
where j, p, q, m, n∈[0,NP-1] are randomly chosen integer and mutually different from
the running index i. F is a scaling factor∈[0,2] which controls the amplification of the
variation vector.
In order to increase the potential diversity of the population, the crossover operation is
also introduced to DE. Assuming the ith individual takes the vector form of
(xi1,xi2,…,xiD), the value of its jth element is then calculated as below after crossover.
x if a random number > CR

xij =  ij (A-3)
∆xij otherwise
where the crossover factor CR∈[0,1] is a control variable that effectively determines
when a parameter should be mutated and thus helps the algorithm converge.
83
Appendix B Flowchart of Proposed Evolutionary Multi-Objective Optimisation Approaches
Appendix B Flowchart of Proposed Evolutionary Multi-
Objective Optimisation Approaches
B.1 Flowchart of Multi-Objective Genetic Algorithm
The pseudo description of the proposed Multi-Objective Genetic Algorithm is outlined
as below.
Generate initial population P(0);

Evaluate P(0);
t:=0;
repeat
Generate P(t+1) using P(t) as follows
{ Assign ranks to the individuals in the P(t) as follows
{ rank_value:=1;
repeat
Find all the non-dominated ones among all the unvisited individuals;
Assign their ranks to be rank_value;
Set visit flag to the assigned individuals;
rank_value:= rank_value+1;
until the rank_value is equal to the set value or all the individuals are visited
}
if (different priorities among objectives)
Modify the ranks in favour of the objective with preference;
Select individuals for reproduction on basis of rank;
Recombine the selected individuals employing crossover and mutation operators;
}
84
Appendix B Flowchart of Proposed Evolutionary Multi-Objective Optimisation Approaches
Evaluate P(t+1);
t:=t+1;
until termination condition is met
B.2 Flowchart of Multi-Objective Particle Swarm and Multi-
Objective Differential Evolution
The main steps of Multi-Objective Particle Swarm and Multi-Objective Differential
Evolution are outlined as following.
Generate initial population P(0);

Evaluate P(0);
t:=0;
repeat
Generate P(t+1) using P(t) as follows
{ Assign ranks to the individuals in the P(t);
Identify the compromise solutions using the predefined scalarising function;
Update Pareto-optimal set;
if(different priorities among objectives)
Modify the ranks in favour of the objective with preference;
Select individuals for reproduction on basis of rank;
Recombine the selected individuals employing PS or DE;
}
Evaluate P(t+1);
t:=t+1;
until termination condition is met
85
Appendix C Preliminary Testing of Multi-objective Optimisation Algorithms
Appendix C Preliminary Testing of Multi-objective
Optimisation Algorithms
C.1 Introduction
Two numerical multi-objective optimisation problems are used here to test the
performance of MOPS and MODE. With the reference of Multi-Attribute Genetic
Algorithm (MAGA) [22], the parameters of MOPS and MODE are varied to observe
how these values affect the algorithm efficiency, measured in terms of the number of
non-dominated solutions obtained and the tendency of the resultant trade-off curve.
There are several variants of DE that generate variation vectors during evolution
procedure. In order to explain MOPS and MODE in a uniform way, two variants of
DE that make use of the global best individual are employed, namely MODE1 and
MODE2.
In MODE1, the variation vector places a perturbation at a location between a randomly
chosen population member and the best population member:
∆xi = x r1 + F * ( xbest + x r 2 − x r 3 − x r 4 ) (C-1)
In MODE2, the variation vector is generated based on the best member, with a
perturbation located among randomly chosen members.
∆xi = xbest + F * ( x r1 + x r 2 − x r 3 − x r 4 ) (C-2)
86
In equation (C-1) and (C-2), r1, r2, r3 and r4 are random integer numbers that differ
from each other. The subscript best represents the index of the best individual in the
population. The search direction of each individual in the population of MODE is
affected by randomly selected individuals as well as by the global best.
The velocity limit of each dimension Vmax in MOPS determines how large steps
through the feasible space each particle is allowed to take. When these steps are
constrained to be too small, individuals may be unable to jump out of poor regions.
When the steps are set to be big, individuals often speed past the target region but
discover even better positions they set out for. Further, individuals are able to escape
from local optima with sufficiently large steps. The mathematical experiments below
demonstrate the effect of the steps each particle takes on the search procedure. For
simplicity, the inertia weight w for PS is set to be 0.9 at the start of the run and
gradually reduced to 0.4 at the end in these two experiments.
C.2 Concave Problem
The first problem is to minimise
 f 1 ( x , y ) = ( x 2 + y 2 )1 / 8

 f 2 ( x, y ) = (( x − 0.5) + ( y − 0.5) )
2 2 1/ 4
where x∈(-5,10)
The trade-off surface of this example is concave, which leads to potential difficulty for
conventional multi-objective optimisation approaches. In order to compare the
algorithms meaningfully, the population of MOPS and MODE is set to 30 and allowed
87
to evolve for 100 generations. The MAGA population is set to 30 and its iteration
number is 197. During the evolution procedure, individuals that violate the constraint
are treated as lethal ones and are discarded. Table C-1 lists the number of non-
dominated solutions found by each algorithm at the end of evolution. The results of
different Vmax for MOPS in terms of f1 and f2 are shown in Figure C-1 whereas the
performance of two variants of MODE (MODE1 and MODE2) is displayed in Figure
C-2.
Algorithm parameters for Test 1 Number of obtained non-

dominated solutions
Vxmax=10.0, Vymax=10.0 5
MOPS Vxmax=5.0, Vymax=0.5 23
MODE MODE1 166
MODE2 512
MAGA 129
Table C-1: Non-dominated solution numbers for Test 1
88
0.9
0.8
0.7 Vxmax=2.0
0.6 Vymax=0.5
0.5 Vxmax=0.5
f2 0.4 Vymax=0.1
0.3 Vxmax=0.2
0.2 Vymax=0.05
0.1
0
0.2 0.4 0.6 0.8 1.0
f1
Figure C-1: MOPS results for Test1
0.9
0.8
0.7
0.6 MODE1
0.5
f2
0.4 MODE2
0.3
0.2
0.1
0
0.2 0.4 0.6 0.8 1.0
f1
Figure C-2: MODE results for Test 1
C.3 Discontinuous Problem
The second test problem is to minimise
89
 − x x ≤1
 
 f ( x ) =  x − 2 1< x ≤ 3
 4 − x 3< x ≤ 4
1
  x − 4 x>4

 f 2 ( x ) = ( x − 5) 2
The trade-off curve for this problem is not continuous over the problem domain. The
population size of MOPS, MODE and MAGA is set to 100. The total iteration number
is 200 for the first two algorithms and 631 for the last algorithm. Similarly, the number
of non-dominated solutions obtained for MOPS, MODE and MAGA is tabulated in
Table C-2 and the performances of MOPS and MODE are shown in Figure C-3 and
Figure C-4.
Algorithm parameters for Test 2 Number of obtained non-

dominated solutions
Vxmax=40.0 191
MOPS Vxmax=20.0 338
Vxmax=10.0 652
Vxmax=5.0 1142
MODE MODE1 9129
MODE2 1431
MAGA 494
Table C-2: Non-dominated solution numbers for Test 2
90
18.0
16.0
14.0 Vxmax=40.0
12.0
Vxmax=20.0
2
10.0
f Vxmax=10.0
8.0
6.0 Vxmax=5.0
4.0
2.0
0.0
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5
f1
Figure C-3: MOPS results for Test 2
18
16
14
12
MODE1
10
f2
8
MODE2
6
4
2
0
-1.5 -1 -0.5 0 0.5 1 1.5
f1
Figure C-4: MODE results for Test 2
C.4 Summary
As seen from the above optimisation results, Vmax does influence the performance of
MOPS to a great extent. With small Vmax, MOPS is liable to be trapped into local
optima but can perform a fine grain search to improve upon the quality of obtained
91
non-dominated solutions, which is obvious from Test 1. On the other hand, large value
of Vmax accelerates MOPS to search the whole feasible space but the interested area is
likely to be ignored.
In fact, any evolutionary algorithm faces with the same challenge of determining
appropriate parameter values to solve specific problem. As far as the two test problems
in this section concerned, MODE outperforms MOPS since the distribution rather than
the aggregation of non-dominated solutions is emphasised in multi-objective
optimisation. Nevertheless, compared with MAGA, experiment results favour MOPS
and MODE. In MAGA, the population evolution procedure is controlled by crossover
and mutation operation, which may result in premature convergence. The search
procedures of MOPS and MODE, however, are guided by the compromise solutions so
the exploration and exploit abilities are fully exerted.
92

Application of Multi-Objective Optimisation To MRT Systems: Founded 1905

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Application of Multi-Objective Optimisation To MRT Systems: Founded 1905

Uploaded by

Copyright:

Available Formats

Founded 1905

NATIONAL UNIVERSITY OF SINGAPORE

projects and writes this thesis.

System laboratory for his support throughout this research project.

unreserved understanding and great support to the author.

TABLE OF CONTENTS ....................................................................................................................... II

LIST OF FIGURES ................................................................................................................................ V

LIST OF TABLES ................................................................................................................................. VI

LIST OF ABBREVIATIONS ............................................................................................................. VII

LIST OF PUBLICATIONS RELATED TO THIS THESIS...........................................................VIII

SUMMARY OF THE THESIS............................................................................................................. IX

CHAPTER 1 INTRODUCTION ............................................................................................................ 1

1.1 BACKGROUND ................................................................................................................................. 1

1.2 MOTIVATION OF THE RESEARCH...................................................................................................... 2

1.2.1 Go-Circuit Optimisation ......................................................................................................... 3

1.2.2 Return-Circuit Optimisation ................................................................................................... 4

1.3 OBJECTIVE AND SCOPE OF THE RESEARCH ...................................................................................... 5

1.4 MULTI-OBJECTIVE OPTIMISATION ALGORITHMS............................................................................. 6

1.5 ORGANISATION OF THE THESIS ........................................................................................................ 8

CHAPTER 2 OUTLINE OF GO-CIRCUIT OPTIMISATION ........................................................ 11

2.1 INTRODUCTION .............................................................................................................................. 11

2.2 MATHEMATICAL MODEL ............................................................................................................... 12

2.2.1 System Model ........................................................................................................................ 12

2.2.2 Objective Functions............................................................................................................... 13

2.2.3 Impact of Operational Timetable .......................................................................................... 15

2.3 LAYOUT OF THREE-STAGE SCHEME .............................................................................................. 16

2.4 SIMULATION OUTLINE ................................................................................................................... 18

2.5 SUMMARY ..................................................................................................................................... 19

CHAPTER 3 OUTLINE OF RETURN-CIRCUIT OPTIMISATION ............................................. 21

3.1 INTRODUCTION .............................................................................................................................. 21

3.2 MATHEMATICAL MODEL ............................................................................................................... 22

3.2.1 Return-Circuit Model ............................................................................................................ 22

3.2.2 Objective Functions............................................................................................................... 23

3.2.3 Impact of Earthing and Bonding Arrangements.................................................................... 24

3.3 LAYOUT OF TWO-STAGE SCHEME ................................................................................................. 25

3.4 SIMULATION OUTLINE ................................................................................................................... 27

3.5 SUMMARY ..................................................................................................................................... 28

CHAPTER 4 LITERATURE REVIEW OF MULTI-OBJECTIVE OPTIMISATION

4.1 INTRODUCTION .............................................................................................................................. 30

4.2 MATHEMATICAL DEFINITION......................................................................................................... 30

4.3 PREFERENCE STRUCTURE .............................................................................................................. 32

4.4 REVIEW OF MULTI-OBJECTIVE OPTIMISATION METHODS ............................................................. 33

4.4.1 Traditional Approaches......................................................................................................... 34

4.4.2 Evolutionary Approaches...................................................................................................... 35

4.4.2.2 Pareto-Based Genetic Algorithm ....................................................................................................36

4.4.2.3 Multi-Attribute Genetic Algorithm.................................................................................................37

4.5 SUMMARY ..................................................................................................................................... 39

CHAPTER 5 EVOLUTIONARY MULTI-OBJECTIVE OPTIMISATION APPROACHES....... 40

5.2 MULTI-OBJECTIVE GENETIC ALGORITHM ..................................................................................... 41

5.2.1 Selection Processing with Rank Assignment ......................................................................... 41

5.2.2 Fitness Sharing...................................................................................................................... 42

5.2.3 Variable Recombination Operators ...................................................................................... 43

5.2.4 Treatment of Preferred Priorities among Objectives ............................................................ 43

5.3 MULTI-OBJECTIVE PARTICLE SWARM ALGORITHM ....................................................................... 45

5.3.1 Search Strategy ..................................................................................................................... 45

5.3.2 Rank-Based Selection ............................................................................................................ 47

5.3.3 Weight Update....................................................................................................................... 47

5.3.4 Pareto-Optimal Set Update ................................................................................................... 48

5.4 MULTI-OBJECTIVE DIFFERENTIAL EVOLUTION ALGORITHM ......................................................... 49

5.5 SUMMARY ..................................................................................................................................... 49

CHAPTER 6 RESULTS OF GO-CIRCUIT OPTIMISATION ........................................................ 50

6.2 WORST-CASE SCENARIOS OF OPERATIONAL DEVIATIONS ............................................................. 53