T HE U NIVERSITY OF S YDNEY

D OCTORAL T HESIS

Power system stability scanning

and security assessment using
machine learning

Author: Supervisor:
Ruidong L IU Dr. Gregor Verbič

A thesis submitted in fulfillment of the requirements

for the degree of Doctor of Philosophy

in the

The Centre for Future Energy Networks

Electrical & Information Engineering
ii

November 8, 2018
 iii

Declaration of Authorship
I, Ruidong L IU, declare that this thesis titled, “Power system stability scan-
ning and security assessment using machine learning” and the work pre-
sented in it are my own. I confirm that:

• This work was done wholly or mainly while in candidature for a re-
search degree at this University.

• Where any part of this thesis has previously been submitted for a de-
gree or any other qualification at this University or any other institu-
tion, this has been clearly stated.

• Where I have consulted the published work of others, this is always

clearly attributed.

• Where I have quoted from the work of others, the source is always
given. With the exception of such quotations, this thesis is entirely my
own work.

• I have acknowledged all main sources of help.

• Where the thesis is based on work done by myself jointly with others,
I have made clear exactly what was done by others and what I have
contributed myself.

Signed:

Date:
 v

The University of Sydney

Abstract
Engineering & Information Technologies
Electrical & Information Engineering

Doctor of Philosophy

Power system stability scanning and security assessment using machine

learning
by Ruidong L IU
vi

Power systems are undergoing a major transformation driven by the in-

creasing uptake of renewable energy sources, DC power transmission, and
the decentralization of electric power supply, like rooftop PV, energy stor-
age, home energy management systems, and electric vehicles. How future
grids will look like, however, is still uncertain as the evolution depends not
only on technological development but also on the regulatory environment.
Future grids planning thus requires a major departure from conventional
power system planning, where only a handful of the most critical scenar-
ios is analyzed. To account for a wide range of possible future evolutions,
scenario analysis has been proposed in many industries. As opposed to the
conventional power system planning, where the aim is to find an optimal
transmission and/or generation expansion plan for an existing grid, the aim
in future grids scenario analysis is to analyze possible evolution pathways to
inform power system planning and policy making. Therefore, future grids’
planning may involve large amount of scenarios and the existing planning
tools may no longer suitable.
Other than the raised future grids’ planning issues, operation of future
grids using conventional tools is also challenged by the new features of fu-
ture grids such as intermittent generation, demand response and fast re-
sponding power electronic plants which lead to much more diverse oper-
ation conditions compared to the existing networks. Among all operation
issues, monitoring stability as well as security of a power system and action
with deliberated preventive or remedial adjustment is of vital important. On-
line Dynamic Security Assessment (DSA) can evaluate security of a power
system almost instantly when current or imminent operation conditions are
supplied. DSA is proved to be more efficient than traditional extensive off-
line security studies and is widely utilized by utilities in power systems oper-
ation. However, given the new features of future grids, the existing dynamic
security assessment tools also need to be examined and refurbished to suit
future grids.
The focus of this dissertation are, for future grid planning, to develop a
framework using Machine Learning (ML) to effectively assess the security of
future grids by analyzing a large amount of the scenarios; for future grids op-
eration, to propose approaches to address technique issues brought by future
grids’ diverse operation conditions using ML techniques. Unsupervised learn-
ing, supervised learning and semi-supervised learning techniques are utilized in
a set of proposed planning and operation security assessment tools.
I was involved in an Australia Commonwealth Scientific and Industrial
 vii

Research Organisationa (CSIRO) project during my candidature and my con-

tribution in this project is given in the first chapter of the dissertation. To
the best of our knowledge, the Future Grid Research Program funded by
the CSIRO, is the first to propose a comprehensive modeling framework
for future grid scenario analysis. The project aims to provide a modeling
framework for the future Australian electricity grid out to 2050 and also
analyze power flow, stability, security and resilience to changing technolo-
gies beyond energy balancing. Diverse scenarios for levels and placement of
renewable generation, different transmission and topologies, different load
management strategies and storage technologies are considered. The second
chapter includes scenarios design for future grids and transient stability as-
sessment using time-domain simulation method based on power flow study
of the scenarios. The study work carried out in the project showed very high
computational burden of the conventional time-domain method and which
leads to the research focus on machine learning methods to overcome the
disadvantages of the existing methods in future grids planning.
One of the main contributions of the dissertation is a framework for fast
stability scanning using ML for future power system planning. The ML tech-
niques used in the third chapter (unsupervised learning) have been used in
the past in many other power engineering applications [1, 2, 3, 4]. Imple-
menting them directly to the problem of fast stability scanning, however, is
infeasible because the number of operating conditions and the number of
features required to describe an operating condition are much larger than
in the existing applications. Unsupervised clustering technique is the first
machine learning approach utilized in this dissertation to deal with future
power system planning stability assessment.
DSA provides power system operators with security information of power
systems for current or imminent operating conditions considering various
system topologies and contingencies. The security information is basis for
preventive or emergency control to prevent systems0 insecurity.
Transient Stability Assessment (TSA) is one of the most important tasks in
DSA. To account for the diverse operation topologies of future grids, in the
fourth chapter a reliable and high accuracy TSA model (supervised learn-
ing) is proposed considering various system topologies. The innovation of
this model are three-fold: (i) first, from feature selection algorithm aspect, a
hybrid filter-wrapper feature selection method is proposed; (ii) then, from In-
telligent System (IS) training algorithm aspect, a boosting learning algorithm
viii

is used during IS training process; (iii) last, from stability decision-making al-
gorithm aspect, an Extreme Learning Machine (ELM) 1 -based ensemble with
a new decision making rule based on weighted outputs of ELMs is proposed
to achieve 100% predictive accuracy.
The core of a TSA is called classifier which is able to give out straight
away transient stability of a power system once an operating condition is
fed. A key feature of a classifier is its generalization ability, which refers
to the ability of the classifier to give reliable and accurate predictions using
previously unseen operating conditions. The generalization ability depends
on the classifier’s structure, the learning algorithm used, the training set size
and its quality [9]. In supervised learning, a training set consists of a group of
operating conditions and the corresponding stability labels (security indexes
in DSA). The labels are normally obtained using computationally expensive
time-domain simulations.
To accommodate fast changing future grids’ operation conditions, DSA
classifier must be updated regularly to ensure its robustness, which requires
generating new training samples and retraining. Clearly, evaluating a large
amount of operating conditions in order to cover a wide range of diverse
operating conditions in a training set quickly becomes computationally pro-
hibitive. Reducing the number of samples in the training set is also not an
option as this would reduce the classifier’s generalization ability [9].
As presented in the fifth chapter, an alternative is to use semi-supervised
learning [10], which uses both labeled and unlabeled operating conditions. In
this dissertation, the last contribution is a new DSA framework based on a
combination of semi-supervised learning and data editing. To improve the
generalization ability of a classifier, a large number of unlabeled operating
conditions are used, which can be computed efficiently by power flow study.
As a result, the proposed DSA framework requires significantly less labeled
operating conditions to achieve a high generalization ability which satisfies
future grids’ operation needs.

1
The ELM used in [5, 6, 7] is essentially a randomized neural network with omission of
bias, first proposed in [8].
 ix

Acknowledgements
I would like to express my deepest thanks to my supervisor, Dr. Gregor Ver-
bič, for his tremendous academic guidance and encouragement throughout
this research work. Thanks also go to Dr. Jin Ma for his help on publications
and valuable academic suggestions.
I am also grateful to Professor David Hill for his guidance and support
during my candidature in the Centre for Future Energy Networks. Special
thanks go to the members of the Centre and Dr. Archie Chapman for their
valuable suggestions and comments.
Finally, I wish to express my utmost gratitude to my parents, my dear
wife and our kids for their care, support and love.. . .
 xi

Publications

Publications during my candidature and my contribution to each publication

in percentage are given below.

R. Liu, G. Verbič and J. Ma, "A machine learning approach for fast future
grid small-signal stability scanning," 2016 IEEE International Conference on
Power System Technology (POWERCON), Wollongong, NSW, 2016, pp. 1-6.

Contributor Modeling Results Analysis Paper writing

Ruidong Liu 90% 90% 80%
Others 10% 10% 20%

R. Liu, G. Verbič and Y. Xu, "A new reliability-driven intelligent system for
power system dynamic security assessment," 2017 Australasian Universities
Power Engineering Conference (AUPEC), Melbourne, VIC, 2017, pp. 1-6.

Contributor Modeling Results Analysis Paper writing

Ruidong Liu 90% 90% 80%
Others 10% 10% 20%

R. Liu, G. Verbič, J. Ma and D. J. Hill, "Fast Stability Scanning for Future Grid
Scenario Analysis," in IEEE Transactions on Power Systems, vol. 33, no. 1,
pp. 514-524, Jan. 2018.

Contributor Modeling Results Analysis Paper writing

Ruidong Liu 90% 90% 80%
Others 10% 10% 20%

R. Liu, G. Verbič and J. Ma, "A New Dynamic Security Assessment Frame-
work Based on Semi-supervised Learning and Data Editing," Submitted to
journal of Electric Power Systems Research, May 2018.

Contributor Modeling Results Analysis Paper writing

Ruidong Liu 90% 90% 80%
Others 10% 10% 20%
xii

H. Marzooghi, M. Gramroodi, A.S. Ahmadyar, R. Liu, G. Verbič, D. J. Hill

"Scenario and Sensitivity Based Stability Analysis of the Australian Future
Grid," Submitted to IEEE Transactions on Power Systems, 2018.

Contributor Modeling Results Analysis Paper writing

Ruidong Liu 25% 25% 20%
Others 75% 75% 80%
 xiii

Contents

Declaration of Authorship iii

Abstract vi

Acknowledgements ix

Publications xi

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 History of Electricity . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Electricity Networks Around the World . . . . . . . . . 1
1.1.3 Electricity in Our Life . . . . . . . . . . . . . . . . . . . 2
1.1.4 Reliable Electricity Supply . . . . . . . . . . . . . . . . . 2
1.2 Security Assessment in Planning and Operation. . . . . . . . . 3
1.2.1 Concepts of Security Assessment . . . . . . . . . . . . . 3
1.2.2 Existing Techniques and Methods . . . . . . . . . . . . 5
1.2.2.1 Time-domain Simulation . . . . . . . . . . . . 5
1.2.2.2 Deterministic Techniques . . . . . . . . . . . . 6
1.2.2.3 Probabilistic Techniques . . . . . . . . . . . . 10
1.2.2.4 Intelligent System Techniques . . . . . . . . . 11
1.3 Challenges in Future Grids Planing and Operation . . . . . . . 16
1.4 Recent Studies on Modeling of Future Grids . . . . . . . . . . 17
1.5 Focus of the Presented Work . . . . . . . . . . . . . . . . . . . . 19
1.6 Structure of the Thesis . . . . . . . . . . . . . . . . . . . . . . . 20
1.7 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2 The CSIRO Future Grid Study Project 25

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2 Project Background . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.2.1 Existing Future Grid Studies . . . . . . . . . . . . . . . 26
2.2.2 Limitations of Existing Studies . . . . . . . . . . . . . . 27
2.3 Future Grids Security Assessment Framework . . . . . . . . . 28
xiv

2.3.1 Test System . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.3.2 Simulation Preparation and Procedures . . . . . . . . . 28
2.3.3 Assumptions and Scenarios Description . . . . . . . . . 31
2.4 Future Grid Transient Stability Study . . . . . . . . . . . . . . . 35
2.5 Transient Stability and Stability Indices . . . . . . . . . . . . . 36
2.6 Extended Equal-Area Criterion Brief . . . . . . . . . . . . . . . 37
2.7 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.8 Security Assessment Simulation Burden . . . . . . . . . . . . . 40

3 Un-supervised machine Learning Method for Fast Stability Scan-

ning 41
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2.1 Existing Studies . . . . . . . . . . . . . . . . . . . . . . . 42
3.2.2 Motivation Behind the Study . . . . . . . . . . . . . . . 43
3.3 Contribution of This Study . . . . . . . . . . . . . . . . . . . . . 44
3.4 Simulation Platform . . . . . . . . . . . . . . . . . . . . . . . . 44
3.4.1 Test System . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.4.2 Scenario Description (Line 1 in Algorithm 1) . . . . . . 46
3.4.3 Time-series Analysis (Lines 2-5 in Algorithm 1) . . . . 46
3.4.3.1 Market Model (Line 3 in Algorithm 1) . . . . 46
3.4.3.2 Load-flow Analysis (Line 4 in Algorithm 1) . 47
3.4.4 Stability Analysis (Lines 6-8 in Algorithm 1) . . . . . . 47
3.4.4.1 Modal Analysis . . . . . . . . . . . . . . . . . 47
3.4.4.2 Steady-state Voltage Stability . . . . . . . . . . 48
3.5 Clustering and Feature Selection . . . . . . . . . . . . . . . . . 48
3.5.1 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.5.1.1 k-means Algorithm . . . . . . . . . . . . . . . 49
3.5.1.2 Particle Swarm Optimization (PSO) . . . . . . 50
3.5.2 Feature Selection . . . . . . . . . . . . . . . . . . . . . . 51
3.5.2.1 Relief Algorithm . . . . . . . . . . . . . . . . . 51
3.6 A Novel Fast Stability Scanning Framework . . . . . . . . . . . 53
3.6.1 Feature Selection for Weighted Clustering (Lines 1-5 in
Algorithm 3) . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.6.2 Self-adaptive PSO-k-means Clustering (Line 6 in Algo-
rithm 3) . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.6.3 Stability Scanning (Lines 7-11 in Algorithm 3) . . . . . 56
3.7 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . 57
 xv

3.7.1 Feature Selection . . . . . . . . . . . . . . . . . . . . . . 57

3.7.2 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.7.3 Small-signal Stability . . . . . . . . . . . . . . . . . . . . 59
3.7.4 Voltage Stability . . . . . . . . . . . . . . . . . . . . . . . 60
3.7.5 Worst Case Operating Point Shift . . . . . . . . . . . . . 62
3.7.6 Simulation Burden of Stability Scanning . . . . . . . . . 64
3.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

4 Supervised machine Learning Method for TSA 65

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2.1 Existing Studies . . . . . . . . . . . . . . . . . . . . . . . 66
4.2.2 Motivation Behind the Study . . . . . . . . . . . . . . . 66
4.3 Contribution of the Study . . . . . . . . . . . . . . . . . . . . . 67
4.4 Extreme Learning Machine Ensemble Learning and Feature
Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.4.1 Extreme Learning Machine Theory . . . . . . . . . . . . 67
4.4.2 Feature Selection for IS . . . . . . . . . . . . . . . . . . . 68
4.4.2.1 Filter and Wrapper Feature Selection . . . . . 69
4.4.2.2 Hybrid Filter-Wrapper Feature Selection . . . 69
4.4.3 Ensemble Learning and Rule for Classification . . . . . 71
4.5 Simulation Platform and Stability Database . . . . . . . . . . . 73
4.5.1 The 39-bus Test System . . . . . . . . . . . . . . . . . . 73
4.5.2 Stability Database . . . . . . . . . . . . . . . . . . . . . . 73
4.6 Results and Analysis . . . . . . . . . . . . . . . . . . . . . . . . 74
4.6.1 Results of Feature Selection . . . . . . . . . . . . . . . . 74
4.6.2 Evaluation with Different System Topologies and Con-
tingencies . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.6.3 Evaluation with Hybrid Feature Selection and Boosting
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.6.4 Decision-making for 100% Predictive Accuracy . . . . 78
4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

5 Semi-supervised machine Learning Method for TSA 83

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.2.1 Existing Studies . . . . . . . . . . . . . . . . . . . . . . . 84
5.2.2 Motivation Behind the Study . . . . . . . . . . . . . . . 84
5.3 Contribution of the Study . . . . . . . . . . . . . . . . . . . . . 85
xvi

5.4 Review of Pertinent Machine Learning Techniques . . . . . . . 86

5.4.1 Training Set Size and Generalization performance . . . 86
5.4.2 Supervised, Unsupervised and Semi-supervised Learn-
ing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.4.3 Tri-training . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.4.4 Data Noise and Data Editing . . . . . . . . . . . . . . . 90
5.5 A New Dynamic Security Assessment Framework . . . . . . . 91
5.5.1 Training Set Generation . . . . . . . . . . . . . . . . . . 91
5.5.2 Extreme Learning Machine Classifier . . . . . . . . . . 92
5.5.3 Feature Selection . . . . . . . . . . . . . . . . . . . . . . 93
5.5.4 Combined Tri-training and Data Editing Algorithm . . 93
5.5.4.1 Initialization (lines 1-4 in Algorithm 8) . . . . 93
5.5.4.2 Accuracy Evaluation (line 9 in Algorithm 8) . 94
5.5.4.3 Pseudo Labeling and Data Editing (lines 11-
22 in Algorithm 8) . . . . . . . . . . . . . . . . 94
5.5.4.4 Tri-training Criteria (lines 23-33 in Algorithm 8) 95
5.5.4.5 Classifier Update (lines 36-41 in Algorithm 8) 95
5.6 Simulation Platform and Stability Database . . . . . . . . . . . 95
5.6.1 The 39-bus Test System . . . . . . . . . . . . . . . . . . 95
5.6.2 Training Set . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.7 Case Study Results and Analysis . . . . . . . . . . . . . . . . . 98
5.7.1 Results of the Feature Selection . . . . . . . . . . . . . . 98
5.7.2 Impact of the Training Set Size . . . . . . . . . . . . . . 99
5.7.3 Impact of the Increased Penetration of Renewables . . 100
5.7.4 Semi-supervised Learning Improves Online TSA Per-
formance by Using Unlabeled Samples . . . . . . . . . 101
5.7.5 Data Editing . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.7.6 Comparison of TSA Classifiers . . . . . . . . . . . . . . 102
5.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

6 Conclusion 105
6.1 Conclusion of the Presented Works . . . . . . . . . . . . . . . . 105
6.2 Suggestion for Future Work . . . . . . . . . . . . . . . . . . . . 106

Bibliography 109
 xvii

List of Figures

1.1 Stability Classification. . . . . . . . . . . . . . . . . . . . . . . . 6

2.1 Simplified 14-generator model. . . . . . . . . . . . . . . . . . . 29

2.2 NEM 16-zones by AEMO. . . . . . . . . . . . . . . . . . . . . . 30
2.3 Hourly transient stability indices of all studied scenarios. . . . 38
2.4 Projection of transient stability indices on all buses of all stud-
ied scenarios. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3.1 14-generator test system. . . . . . . . . . . . . . . . . . . . . . . 45

3.2 Comparison of the clustering results: conventional k-means
vs. the proposed self-adaptive PSO-k-means. . . . . . . . . . . 59
3.3 SSA critical damping ratio fast scanning results: (a) time series,
(b) error distribution using PSO-k-means and k-means. . . . . 60
3.4 Convergence of feature selection: (a) SSS, (b) VS. . . . . . . . . 61
3.5 VSA loading margin fast scanning results: (a) time series, (b)
error distribution using PSO-k-means and k-means. . . . . . . 62
3.6 SSA: Critical mode damping ratio vs. demand. . . . . . . . . . 63
3.7 VSA: Loading margin vs. demand. . . . . . . . . . . . . . . . . 63

4.1 Ranked weights of candidate features as result of the RRelief-F

feature selection. . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.2 Classification accuracy improvement during the SFFS process. 76
4.3 Classification accuracy affected by topology change. . . . . . . 77
4.4 Classification accuracy improved by SFFS and boosting learning. 78
4.5 Prediction results of all misclassified OCs for faults on the buses. 79
4.6 Ensemble classification using weighted outputs improves class
separation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

5.1 Block diagram of the proposed DSA framework. . . . . . . . . 92

5.2 IEEE New-England 39 Bus System. . . . . . . . . . . . . . . . . 97
5.3 Top 100 features as a result of feature selection. . . . . . . . . . 99
5.4 Classification accuracy vs the training set size. . . . . . . . . . 100
xviii

5.5 Bus voltage phase angle variance for the conventional and the
renewable scenario. . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.6 Comparison of the classification results for the conventional
and the renewable scenarios. . . . . . . . . . . . . . . . . . . . 102
5.7 Comparison of the number of noisy samples per iteration in
tri-training with and without data editing. . . . . . . . . . . . . 103
5.8 Compare prediction results of supervised, semi-supervised and
semi-supervised with active learning algorithms. . . . . . . . . 104
 xix

List of Tables

1.1 Notable Power Outages . . . . . . . . . . . . . . . . . . . . . . 3

2.1 Abbreviations Used in Study Scenarios Description . . . . . . 35

2.2 Comparison of Generation Portfolio . . . . . . . . . . . . . . . 35

3.1 Features and weights for SSS . . . . . . . . . . . . . . . . . . . 57

3.2 Features and weights for VS . . . . . . . . . . . . . . . . . . . . 58

4.1 Test System Topologies . . . . . . . . . . . . . . . . . . . . . . . 73

5.1 Test System Topologies . . . . . . . . . . . . . . . . . . . . . . . 98

 xxi

List of Abbreviations

CSIRO Commonwealth Scientific Industrial and Research Organisation

DSA Dynamic Security Assessment
ML Machine Learning
TSA Transient Stability Assessment
IS Intelligent System
AUPEC Australasian Universities Power Engineering Conference
DAE Differential Algebraic Equations
DG Distributed Generation
PDF Probabilistic Density Function
ELM Extreme Learning Machine
ANN Artificial Neural Network
SVM Support Vector Machines
DT Decision Tree
PMUs Phasor Measurement Units
VMV Voltage Magnitude Violation
TV Thermal Limit Violation
VS Voltage Stability
TS Transient Stability
OCs Operating Conditions
TEF Transient Energy Function
CVM Core Vector Machine
NEM National Electricity Market
RES Renewable Energy Source
FACTS Flexible Alternating Current Transmission System
CST Concentrating Solar Thermal
SH Snowy Hydro
NSW New South Wales
VIC Victoria
QLD Queensland
SA South Australia
AEMO Australian Energy Market Operator
AVR Automatic Voltage Regulation
xxii

PSS Power System Stabiliser

UC Unit Commitment
PSAT Powerflow and Short-Circut Analysis Tool
DSATools Dynamic Security Assessment Software
SSS Small Signal Stability
EEAC Extended Equal Area Criteria
TSAT Transient Security Assessment Tool
TSI TS Index
VSAT Voltage Security Analysis Tool
FS Frequency Stability
RoCoF Rate of Change of Frequency
NTNDP National Transmission Network Development Plan
FiTs Feed in Tariffs
BS Battery System
IPBS Integrated PV-battery Systems
SoC State of Charge
WF Wind Farms
PSO Particle Swarm Optimization
SMSE Sum of the Mean Squared Error
SDB Stability Database
BP Back-Propagation
SLFN Feedforward Neural Network
SFFS Sequential Floating Forward Selection
LRS Plus-L and Minus-R Selection
PAC Probably Approximately Correct learning
KNN K-Nearest Neighbor
 xxiii

For All Members of My Family. . .

 1

Chapter 1

Introduction

1.1 Background

1.1.1 History of Electricity

Electricity has been in use for a very long time - electric fish and eels were
using it a long time before it was ever harnessed by humans. In the 18th
century, Benjamin Franklin conducted an extensive famous research in elec-
tricity in which he flew a kite with a dampened string and attached a metal
key to the string in order to route electricity from a thunderstorm. Michael
Faraday made the generation of electricity a practical possibility. Michael
Faraday, in 1831, created the world’s first generator by moving copper coils
through a magnetic field. The conceptual generator model was further re-
fined by Thomas Edison years later. Edison’s generator is the first in mankind
history to power electric streetlights in New York City in 1881. The work of
Edison, Tesla, Westinghouse and many other inventors and engineers in elec-
trical engineering in the 20th century helped set the foundations for today,
shaping electricity as an essential tool of modern life and a driving force for
human societies development.

1.1.2 Electricity Networks Around the World

Nowadays, electricity networks on different continents cover vast areas
and are normally interconnected to achieve reliability and economic opera-
tion.
The mainland North American electricity network consists with the East-
ern Interconnection, the Western Interconnection, the Texas Interconnection,
the Quebec Interconnection, and the Alaska Interconnection. The regions are
2 Chapter 1. Introduction

not directly connected or synchronized to each other, but there are HVDC in-
terconnections connect all the regional networks. This immense North Amer-
ican network of power lines, generation facilities, and related communica-
tions systems is often referred to as "the world’s largest machine" and supply
electricity continuously to one of the world’s most developed regions.
The interconnected network of Continental Europe is the largest synchronous
electricity network in the world. the network is run by 43 electricity trans-
mission system operators; supplies over 500 million customers in 36 coun-
tries, including most of the European Union. In 2016, 1136 GW of production
capacity was connected to the grid in which 506 GW is contributed by re-
newable generation including hydro [11]. As the largest developing country,
China’s power industry is characterized by fast growth and an enormous
installed base. In 2014, it had the largest installed electricity generation ca-
pacity in the world with 1505 GW and generated 5583 TWh [12, 13]. China
also has the largest thermal power capacity, the largest hydropower capacity,
the largest wind power capacity and the largest solar capacity in the world.
Despite an expected rapid increase in installed capacity scheduled in 2014 for
both wind and solar, and expected increase to 60 GW in nuclear by 2020, coal
will still account between 65% and 75% of capacity in 2020 [14].

1.1.3 Electricity in Our Life

One hallmark of our modern societies is electricity usage and electricity is
an essential element contributed to our economic development. Not to men-
tion the developed countries, in recent years, people in many developing
countries from India, China to many African countries have experienced ris-
ing living standards, as more people are able to access electricity to keep their
home appliance working. Many devices facilitating our life and impacting a
big portion of our living standard are powered by electricity. Computers,
TVs, mobile phones, white goods such as washing machine, dishwashing
machine which free us from daily work require electricity. Electricity is now
a basic element of our life.

1.1.4 Reliable Electricity Supply

Our societies are heavily relying on the electricity and our day to day life
and economy can be greatly affected if we cannot have a reliable electricity
supply.
1.2. Security Assessment in Planning and Operation. 3

During a power blackout, traffic lights go off, creating chaos on the af-
fected roads. Airports and railways cannot be properly operated. Commu-
nication system covering almost every corner of our societies are no longer
functioning. People cannot find out what is happening around since many
are relying on heavily internet to gain information. Factories and offices are
shut down. One can imagine the vast impact of a large scale blackout on our
life and the economy. The power outage during USA California’s capacity
crisis in 2000 affected 1.5 million people, had effects on California’s GDP (it
was cut by 0.7-1.5%) and are thought to have cost around $40 billion [15]. Ac-
cording to the Royal Academy of Engineers, the cost of an electricity shortfall
in the UK would run into billions of pounds. In South Africa, power shortage
has had devastating economic consequences too. Table 1.1 lists some notable
wide-scale power outages around the world [16].

TABLE 1.1: Notable Power Outages

Events People Affected Location Date

1 620 million India 30/07/2012 31/07/2012
2 230 million India 2/01/2001
3 150 million Bangladesh 1/11/2014
4 140 million Pakistan 26/01/2015
5 100 million Indonesia 18/08/2005

1.2 Security Assessment in Planning and Opera-

tion.
To design and maintain a reliable power system is of important since our
societies and economy are heavily relying on a continuous power supply.
Power system reliability is therefore the overall objective in power system
planning and operation.

1.2.1 Concepts of Security Assessment

The conceptual definition and relationship of reliability, security and sta-
bility are given below [17]:
Reliability of a power system refers to the probability of its satisfactory operation
over the long run. It denotes the ability to supply adequate electric service on a nearly
continuous basis, with few interruptions over an extended time period.
4 Chapter 1. Introduction

Security of a power system refers to the degree of risk in its ability to

survive imminent disturbances (contingencies) without interruption of cus-
tomer service. It relates to robustness of the system to imminent disturbances
and, hence, depends on the system operating condition as well as the contin-
gent probability of disturbances.
Stability of a power system, refers to the continuance of intact operation
following a disturbance. It depends on the operating condition and the na-
ture of the physical disturbance.
The essential differences among the three aspects of power system per-
formance are also given in [17]:

• Reliability is the overall objective in power system design and opera-

tion. To be reliable, the power system must be secure most of the time.
To be secure, the system must be stable but must also be secure against
other contingencies that would not be classified as stability problems
e.g., damage to equipment such as an explosive failure of a cable, fall of
transmission towers due to ice loading or sabotage. As well, a system
may be stable following a contingency, yet insecure due to post-fault
system conditions resulting in equipment overloads or voltage viola-
tions.

• System security may be further distinguished from stability in terms of

the resulting consequences. For example, two systems may both be sta-
ble with equal stability margins, but one may be relatively more secure
because the consequences of instability are less severe.

• Security and stability are time-varying attributes which can be judged

by studying the performance of the power system under a particular
set of conditions. Reliability, on the other hand, is a function of the
time-average performance of the power system; it can only be judged
by consideration of the system’s behavior over an appreciable period
of time.

Reliability of a power system depends on proper system planning, oper-

ation and management and is the overall target. Security and stability as-
sessment are required in power system planning and operation in order to
achieve a higher reliability performance.
Security assessment includes two important components: static security
analysis involves steady-state analysis of post-disturbance system conditions
1.2. Security Assessment in Planning and Operation. 5

to verify that no equipment ratings and voltage constraints are violated; dy-
namic security analysis involves examining different categories of system
stability performances such as angle stability, voltage stability and frequency
stability. Stability of a power system refers to the continuance of intact oper-
ation following a disturbance and is thus the most important integral com-
ponent of system security assessment. Security assessment of power systems
is crucial for both systems planning and systems operation.

1.2.2 Existing Techniques and Methods

Security analytical techniques are well developed for conventional power
systems in the past decades; some typical studies are revisited including de-
terministic, probabilistic techniques and intelligent system techniques.

1.2.2.1 Time-domain Simulation

Time-domain simulation solves a large amount of Differential Algebraic

Equations (DAE) representing the system under study by using integration
methods in a step-by-step process for a transient time frame (usually up to
10 seconds after the disturbance). Simulation softwares check interested sta-
bility indexes in the integration process and signal when the indexes excess
a certain limit, such as generators’ angle difference in a network for transient
stability study. Conventionally, in power systems planning and operation
security assessment, only limited number of scenarios and contingencies are
investigated to ensure expansion of the system or a particular operation con-
dition does not jeopardize the system security. Thus, time-domain simula-
tion is widely used as an offline application since long simulation time is not
yet an issue.
In order to assess power systems security in a fast manner, many other
methods including deterministic and probabilistic techniques have been well
studied and implemented in power system’s planning and operation in the
past decades. Compared to the time-domain simulation, these techniques are
much faster and therefore can be used in power systems operation for on-line
security assessment. This section revisits some of the major techniques which
have been published.
6 Chapter 1. Introduction

1.2.2.2 Deterministic Techniques

Stability analysis is one of the cores involved in power system security

assessment, the area covers steady-state and dynamic stabilities. Voltage sta-
bility is widely accepted as both steady-state and dynamic phenomenon. The
angle stability is studied under small disturbances or large disturbances, cor-
respond to damping/oscillation study and transient study. Frequency stabil-
ity is another aspect of the power system stability issue which might draw
more attention due to much less inertia in the future grid. Traditionally and
has been widely used approach in stability studies is so called deterministic
method, which requires a specific power system configuration, fixed param-
eter sets and predefined disturbances.
In 2004, a definition and classification of power system stability report
is published by IEEE/CIGRE Joint Task Force on Stability Terms and Defini-
tions [17]. This report addresses the issue of stability definition and classifica-
tion in power systems from a fundamental viewpoint and closely examines
the practical ramifications. Fig. 1.1 gives the overall picture of the power
system stability problem, identifying its categories and subcategories.

F IGURE 1.1: Stability Classification.

Voltage Stability Analysis

1.2. Security Assessment in Planning and Operation. 7

Voltage stability refers to the ability of a power system to maintain steady

voltages at all buses in the system after being subjected to a disturbance from
a given initial operating condition [17].
Voltage stability concerns are not new and in the past decades, researchers
have developed different approaches to investigate if a system is near or close
to voltage instability, these approaches can be classified into dynamic and
steady state analysis.
Among those voltage steady-state stability analysis methods, there is one
so called voltage collapse proximity indicator method. Based on the power
flow Jacobian and measures of singularity, different indices in terms of power
flow solutions have been developed. The singular value method is based
on calculating the reduced Jacobian matrix minimum singular value, which
gives indices to voltage stability proximity. The singular value method is re-
garded by some researchers to reflect Jacobian singularity of the power sys-
tem better [18, 19].
In [18], the authors presented a fast method to compute the minimum sin-
gular value of a power flow Jacobian matrix, together with the correspond-
ing right and left singular vectors. The method presented is based on con-
sideration of amplification in direction defined by the singular vectors. The
performed case studies results show that the minimum singular value of the
power flow Jacobian and its sub-matrix are good indicator s of the proxim-
ity to the static voltage stability limit. The right singular vector corresponds
to sensitive voltages (angles) and the left singular vector indicates the most
sensitive directions for changes of active and reactive power injections [18].
Another method is linear sensitivity analysis, which is based on load-
flow linear first-order sensitivities. By using them, it is straightforward to
determine what changes in parameters (control variables) would be most ef-
fective to produce desired changes in dependent variables (state variables) of
the power system [20].
Simulation approach is another important tool to analyse steady-state
voltage stability issue. Conventional methods used for voltage stability as-
sessment are the PV and QV curves and the modal analysis technique [21].
The V-Q curves analysis was developed from the difficulties of convergence
of power flow program for cases stressed near the maximum power trans-
fer on a path [22]. Other than the aforementioned methods, so-called direct
methods for finding voltage collapse points in the power system have been
developed and can be used as an alternative for the singular value method.
8 Chapter 1. Introduction

Voltage stability assessment in power system also uses Decision Tree Tech-
nique [23] and Artificial Intelligence techniques like Fuzzy networks [24] and
Artificial Neural Networks [25].
Computational efficiency (speed) is the key element for online stability
monitoring. This can be achieved through increasing the power of computa-
tional devices or by using different stability indices to reduce computational
complexity.
Damping Analysis
Small-disturbance (or small-signal) rotor angle stability is concerned with
the ability of the power system to maintain synchronism under small distur-
bances. The disturbances are considered to be sufficiently small that lineari-
sation of system equations is permissible for purposes of analysis [17].
Conventionally time domain simulation and eigenvalue analysis of lin-
earised system models are widely used to evaluate the small signal stability
of power systems[26]. These methods have been used for decades. Other
than the numerical simulation and eigenvalues analysis approach, there are
small signal stability region method [27], non-linear analysis method [28] and
probabilistic approach [29]. The eigenvalue method is based on Lyapunov’s
first method, and includes QR methods [30] and partial eigenvalue method
[31].
Although the traditional eigenvalue analysis of the state matrix A which
is derived from linearisation around an operating point of the differential al-
gebraic equations had been widely performed on power systems since the
1960’s, application of bifurcation theory to power system stability analysis
was started in the 1990’s [32]. From the point of view of bifurcation theory,
local bifurcations and hence system stability are studied through the determi-
nation of a series of system eigenvalues associated with the gradual evolution
of certain system parameters (e.g., demand changes) [33]. Local bifurcations
having a significant effect on the stability of the system and most studied in
the literature are the saddle-node (one of the eigenvalues becomes zero) and
the Hopf bifurcations (a pair of complex eigenvalues cross the imaginary axes
of the complex plane).
Probabilistic analysis method has also been studied which is discussed
later in this proposal.
Frequency Stability Analysis
Frequency stability refers to the ability of a power system to maintain
steady frequency following a severe system upset resulting in a significant
1.2. Security Assessment in Planning and Operation. 9

imbalance between generation and load [17]. Frequency stability issue is not
involved in this research work.
Transient Stability Analysis
Large-disturbance rotor angle stability or transient stability, as it is com-
monly referred to, is concerned with the ability of the power system to main-
tain synchronism when subjected to a severe disturbance, such as a short
circuit on a transmission line. [17].
Transient stability analysis conducted by power system planning and op-
eration engineers to evaluate the response of the system to various severe
disturbances. It has been in practice to investigate transient stability by nu-
merical simulation in utility planning. But the numerical integration method
requires intensive and time-consuming computational effort and tradition-
ally precluded from online security analysis. This necessitated the devel-
opment of fast simulation and direct methods of transient stability analysis
which is mostly based on Lyapunov second method [26].
Direct methods was likely first proposed by Magnusson [34] in the late
1940’s, and pursued in the 1950’s by Aylett [35], and in the 1960’s by El-Abiad
and Nagappan [36]. In contrast to the time-domain approach, direct methods
determine system stability directly based on energy functions. These meth-
ods determine whether or not the system will remain stable once the fault
is cleared by comparing the system energy (when the fault is cleared) to a
critical energy value.
Although direct stability methods have been in use since 1960’s, methods
featuring energy functions have been the more preferred ones. However,
direct methods must overcome several challenges (modeling, function and
reliability) and limitations (scenario, condition and accuracy) before they can
become a widely accepted practical tool [37].
In [38], the authors offered a systematic procedure of constructing energy
functions for both network-reduction and network-preserving power system
models. An advanced method, called the BCU method, of computing the
controlling unstable equilibrium point is presented along with its theoret-
ical foundation. Numerical solution algorithms capable of supporting on-
line applications of direct methods are provided. Practical demonstrations
of using direct methods and the BCU method for online transient stability
assessments on two power systems are described. Another important direct
method called Extended Equal Area Criterion is widely used in commercial
power system study software [39].
10 Chapter 1. Introduction

1.2.2.3 Probabilistic Techniques

In contrast to those above revisited deterministic analysis approaches which

are widely used by researchers and power engineers in planning and opera-
tion security assessment work, probabilistic approaches provide an alterna-
tive to study varies stability issues in power system. Deterministic methods,
which require a specific power system configuration, fixed parameter sets
and predefined disturbances. However the configuration of the network af-
ter contingencies is unpredictable, the system parameters are not constant
but vary around rated values slightly, the intermittent output of RES gener-
ators and the contingencies themselves are also unpredictable, etc. To inves-
tigate the security level of a power system thoroughly, one needs to consider
all possible scenarios, which are the combination of the aforementioned un-
certainties. Even with different direct analytical methods, it’s still quite dif-
ficult or impossible to investigate the massive scenarios, not even mention
the much more time consuming but more accurate and reliable simulation
method.
The subject of planning and reliability measures is moving away from de-
terministic criteria toward hybrid deterministic-probabilistic planning meth-
ods to provide a quantified risk assessment using performance indices which
are sensitive to factors that affect reliability [40].
In [41], the authors applied Two Point Estimate Method in static volt-
age probabilistic stability analysis, utilized power margin to calculate the
stability probability without construction of complex relationship between
the variables and the stability index. This avoids model reduction approach
which is used in analytical studies. Much less simulation work is required to
find the statistical stability profile of the voltage. The research only focused
on the uncertainties of generation and demand.
In [42], the authors performed large amount of time-domain simulation
to investigate the impact of possible installed Distributed Generation (DG)
in distribution network on the voltage profiles of the system. The simulation
is based on a model which allocated the DGs in a probabilistic fashion to
account for the uncertainties of future installations. In this paper, the Gibbs
sampler algorithm is used to generate three key parameters for the allocation
of the non-deterministic DGs: type, size, and location. The Gibbs sampler
algorithm is one of the Markov Chain Monte Carlo methods.
In [29], the authors proposed an analytical method of probabilistic analy-
sis to investigate the impact of stochastic uncertainty of grid-connected wind
generation on power system small-signal stability. The method can directly
1.2. Security Assessment in Planning and Operation. 11

calculate the Probabilistic Density Function (PDF) of critical eigenvalues of

a large-scale power system from the PDF of wind power generation. It is
shown that the stochastic variation of grid-connected wind generation can
cause the system to lose stability even though the system is stable determin-
istically.
In [43], the authors discussed the impact of uncertain power injections
in the grid on the load margin. Two common analysis of voltage stability:
closest saddle node bifurcation and prefixed direction of load and production
increase are covered in the paper. The loading margin is interpreted as a
stochastic variable itself, this allows to interpret load margins at different
levels of probability of voltage collapse with or without corrective actions
undertaken. The probabilistic margin is assessed with a minimum number
of samples by use of a stochastic response surface method implementation.
Above mentioned studies justified the advantages of probabilistic analy-
sis in the future grid security studies. Both indirect (simulation) and direct
(analytical) methods are proved to be successful in voltage and small signal
stability assessment. The advantage of calculation speed of direct methods
and accuracy of indirect methods can be combined to achieve the target of
online security assessment for the future grid.

1.2.2.4 Intelligent System Techniques

Other than the above mentioned techniques, intelligent systems are also
used in power system security assessment to overcome disadvantages of the
conventional techniques, such as time-domain simulation method. Time-
domain simulation is an accurate, flexible, and reliable security assessment
method. However, since the early ’90s, the power industry has drastically
changed. Open power markets, renewable energy, and the current shift to-
wards Smart-Grids significantly complicate the planning and operating sys-
tems. In power systems operation, conventional practice based on offline
studies is inadequate and costly.
In power systems operation, as summarized in [44], reasons to shift from
offline to on-line security assessment are as follows:

• Offline methods tend to be conservative when determining the avail-

able power-transfer ability, and this results in added cost.

• Electricity load growth outpaces the growth of the infrastructure, push-

ing the system to operate closer to its stability boundary.
12 Chapter 1. Introduction

• Market activity drives the system into an unpredictable, erratic power

transfer pattern.

• The offline method can’t capture the realistic operating conditions of

renewable energy (particularly wind and solar power), and the system
becomes unreliable.

An alternative to achieve fast and reliable power system security assess-

ment is the intelligent system. As summarized in [45], There are significant op-
portunities for the introduction of intelligent systems for use in on-line DSA. Many
on-line DSA systems in use today may be referred to as deterministic systems since
they rely largely on enumerated analytical solutions.
An intelligent system is a computer-based system that can represent, rea-
son about, and interpret data. In doing so it can learn about the structure of
the data, analyze the data to extract patterns and meaning, derive new in-
formation, and identify strategies and behaviors to act on the results of its
analysis. Intelligent systems come in many forms and have many applica-
tions, from processing huge data sets to controlling robots and drones. Ma-
chine learning is one of many application of IS based around the idea that a
machine - computer based - can learn from large amount of data to extract
desired information by itself.
Contrast to deterministic or probabilistic techniques which aim to con-
struct a mathematical relationship between a given power system condition
(with interested disturbance applied) and the system security status, an IS
can be used to map the same condition to the security status straightaway
when it is properly trained before hand.
Intelligent systems are seen to have four features which can bring benefits
to the real-time environment,

• Intelligent systems can be very fast. Although distributed computation

is now commonplace, full simulation methods require minutes of time
to reach a conclusion. For large power systems in which many con-
tingencies must be assessed, even with multiple-CPU computing, this
time may be of concern and for on-line analysis, time is critical; partic-
ularly if a system is entering an insecure state and decisions must be
made quickly.

• Intelligent systems are learning systems. Deterministic systems will

conduct the same computations every cycle even if some of the calcula-
tions could be deemed inconsequential or if conditions arise rendering
1.2. Security Assessment in Planning and Operation. 13

the computations less accurate. IS systems have the ability to establish

if a system condition has been seen previously and predict the solution
accordingly.

• Intelligent system can provide a high degree of discovery. Discovery

refers to the ability to uncover salient, but previously unknown, char-
acteristics of, or relationships in, a system.

• Intelligent systems have the ability to synthesize large volumes of data

into manageable and meaningful information. Considering the poten-
tially massive volume of data provided from system measurements and
simulation results, the ability to sift through such data and “collate” the
useful results is critical for DSA systems.

Many works have been published using IS for on-line power system se-
curity assessment.
In [46], the authors provided a broad overview of online power system
security analysis, with the intent of identifying areas needing additional re-
search and development. Major components which are involved in the secu-
rity analysis are identified, including data preparation (measurements, filter-
ing, state estimation), online load flow study, contingency selection and secu-
rity evaluation. Within the paper, data estimation and contingency analysis
are discussed in detail, also optimization of preventive and corrective actions
are covered. The procedure and approaches to carry out dynamic stability as-
sessment is provided. The study indicated that artificial intelligence or expert
systems have proven to be appropriate solutions to other power system op-
erations problems and speculated that these technologies will play a major
role in dynamic stability assessment. The paper also briefly discussed pat-
tern recognition methods and probabilistic methods likelihood application
in dynamic security assessment in the areas of contingency screening and in
quantifying the probability of the next state of the system. The study gives
clear scope related to online security assessment, the key issues need to be
tackled and procedures traditionally been used.
In [44], the authors provided an overview of possibility and advantages
of application of IS in online stability assessment. The key factors involved in
implementation and a typical structure of an intelligent stability assessment
system is proposed. The procedure starts from data preparation, selection
of input and output, significant features and learning algorithm are also dis-
cussed. Considering high uncertainty of parameters involved in future grids
14 Chapter 1. Introduction

security assessment, Monte Carlo simulation approach is also used by the

authors.
In [47], the authors developed an intelligent framework for real-time DSA
of power systems with large penetration of wind power. The framework con-
sists four major components, a DSA engine whose role is to perform real-time
DSA of the power system, a wind power and load demand forecasting en-
gine for offline and online predicting wind power generation and electricity
load demand, a database generation engine for generating instances to train
the DSA engine, and a model updating engine for online updating the DSA
engine. In the paper, the authors used an algorithm called extreme learning
machine to overcome inadequacy of other soft computing approaches such
as Artificial Neural Network (ANN), Support Vector Machines (SVM), Deci-
sion Tree (DT), Fuzzy rule and data mining.
Decision tree is one of the most popular ISs used for TSA [48, 49].
Authors in [48] proposed an online dynamic security assessment scheme
for large-scale interconnected power systems using phasor measurements
and decision trees. The scheme builds and periodically updates decision
trees offline to decide critical attributes as security indicators. The scheme
uses a new classification method involving each whole path of a decision
tree instead of only classification results at terminal nodes to provide more
reliable security assessment results for changes in system conditions.
In [49], authors used Phasor Measurement Units (PMUs) and decision
trees to develop a real-time security assessment tool to assess four impor-
tant post-contingency security issues, including Voltage Magnitude Violation
(VMV), Thermal Limit Violation (TV), Voltage Stability (VS) and Transient
Stability (TS). The proposed scheme is tested on a real power system rep-
resented by a series of Operating Conditions (OCs) during a representative
day. Robustness tests for the offline trained DTs are performed on a group
of changed OCs that were not included for training the DTs and the idea of
tuning critical system attributes for preventive controls is also presented to
improve system security.
Support Vector Machine (SVM) is another popular tool used by researchers
[50, 51]. In [50], authors present a novel approach to enable frequent compu-
tational cycles in online dynamic security assessment by using the terms of
the Transient Energy Function (TEF) as input features to a machine learn-
ing algorithm (support vector machine is used). The aim is to train a single
classifier that is capable of classifying stable and unstable operating points
independent of the contingency. The network is trained based on the current
1.2. Security Assessment in Planning and Operation. 15

system topology and the loading conditions. It is shown that the classifier
can be trained using a small set of data when the terms of the TEF are used
as input features.
In [51], online power system TSA problem is mapped as a two-class clas-
sification problem - offline training and online application - and a novel data
mining algorithm using the Core Vector Machine (CVM) is proposed to solve
the problem based on PMU big data. Compared with other Support Vector
Machines, the proposed CVM based assessment algorithm has higher preci-
sion and the least time consumption and space complexity.
Authors in [5, 6, 7] used the ELM algorithm for pre-fault and post-fault
online TSA.
DSA application of the ELM algorithm proposed in [5] has shown to have
faster learning speed compared to other ISs. In the paper, a new transient
stability assessment model using the increasingly prevalent extreme learn-
ing machine theory is developed. It has significantly improved the learning
speed and can enable effective on-line updating. The proposed model is com-
pared with some state-of-the-art methods in terms of computation time and
prediction accuracy. The simulation results show that the proposed model
possesses significant superior computation speed and competitively high ac-
curacy.
A TSA model using an ELM-based ensemble in [6] is proposed by au-
thors. The model is developed for real-time dynamic security assessment
of power systems. The IS structures a series of extreme learning machines
and generalizes the randomness of single ELMs during the training. The
proposed model learns and works very fast and can estimate the credibility
of its DSA results, allowing an accurate and reliable pre-fault DSA mecha-
nism: credible results can be directly adopted while incredible results are de-
cided by alternative tools such as time-domain simulation. This makes the IS
promising for practical application since the potential unreliable results can
be eliminated for use. Case studies considering classification and prediction
are, respectively, conducted on an IEEE 50-machine system and a dynamic
equivalent system of a real-world large power grid.
In study of post-fault TSA, authors in [7] used an ELM-based ensemble
and proposed a new decision-making rule. The case study demonstrated
feasibility of the ELM application for post-fault TSA, which requires faster
learning speed than pre-fault TSA applications. This paper develops a novel
IS to balance the post-fault TSA response speed and accuracy requirements.
A set of classifiers are sequentially organised, each is an ensemble of extreme
16 Chapter 1. Introduction

learning machines. A self-adaptive TSA decision-making mechanism is de-

signed to progressively adjust the response time.

1.3 Challenges in Future Grids Planing and Oper-

ation
Power systems are undergoing a major transformation driven by the in-
creasing uptake of renewable energy sources, DC power transmission, and
the decentralization of electric power supply, like rooftop PV, energy storage,
home energy management systems, and electric vehicles. Renewable gen-
erators and smart grid associated techniques are shaping the big networks
around the globe.
We interpret future grids to be any grid type structures with the above-
mentioned transformational changes which are significantly different to the
existing power systems.
Our power system planning and operation engineers are challenged by
technical issues they haven’t seen. The new features of our power system
brought by distributed generation, electric vehicles, utility scale storage, a
large amount of power electronics application and participation of end users
in system operation should be addressed to ensure a sustainable and reliable
future power system.
How future grids will look like, however, is still uncertain as the evolution
depends not only on technological development but also on the regulatory
environment. The future grid planning aim is to find an optimal transmission
and/or generation expansion plan for an existing grid, the aim in future grids
scenario analysis is to analyze possible evolution pathways to inform power
system planning and policy making. Therefore, future grids’ planning may
involve large amount of scenarios and the existing planning tools may no
longer suitable.
Other than the raised future grids’ planning issues, operation of future
grids using conventional tools is also challenged by the new features of fu-
ture grids such as intermittent generation, demand response and fast re-
sponding power electronic plants which lead to much more diverse oper-
ation conditions compared to the existing networks. Among all operation
issues, monitoring stability as well as security of a power system and action
1.4. Recent Studies on Modeling of Future Grids 17

with deliberated preventive or remedial adjustment is of vital important. On-

line dynamic security assessment can evaluate security of a power system al-
most instantly when current or imminent operation conditions are supplied.
DSA is proved to be more efficient than traditional extensive off-line secu-
rity studies and is widely utilized by utilities in power systems operation.
However, given the new features of future grids, the existing dynamic secu-
rity assessment tools also need to be examined and refurbished to suit future
grids.

1.4 Recent Studies on Modeling of Future Grids

So far, power system planners and operators have, in general, well man-
aged security issues of the existing power systems. However, our power
systems are undergoing a revolution as a result of emerging renewable en-
ergy sources, more flexible networks and new load profiles. How to handle
the new challenges of our future grids in planning and operation security
assessment is a key to have reliable future grids.
A proper mathematical model of a power system which is closest to its
physical network is vital when analyzing the real world system by using the
model. As discussed previously, many new features which are not part of
our existing power systems will be playing important roles in future grids
and capture these new features in a power system model for planning or
operation security assessment is the first important step. Many studies have
been carried out recently on future grids modeling and some are given in the
following subsection.
The future grid test bed for this research is of fundamental and the first
task to be completed. Though the results of this research work is not case
specified, but can apply to any future grids, a test bed is required to carry out
necessary simulation works. In [37], the authors concluded that there is no
international research so far has developed a standardized comprehensive
modeling framework for future grids which closes to the modeling frame-
work for classical systems: a suite of definitions, equations and software for
power flow, stability analysis, dispatch, security and reliability. Also even
some well-known modern power system study software have included new
features to cover renewables, however, deficiencies remain in the capability
to represent completely the features of future grids in the long-term.
18 Chapter 1. Introduction

A few published future grid models analysed by the authors [52, 53, 54,
55], the conclusion is that those models all have pros and cons as explained
in the next a few paragraphs.
The first model proposed in [52] - Australian Energy Market Operator
(AEMO)’s 100 percent renewables study, four scenarios are considered dis-
tinguished by the development of renewable generation technologies and
the demand projection. The focuses are again on economic analysis and bal-
ance. In [53] is called Future Zero-carbon Electrical Grid of Australia (ZCA)
in 2020. In this model renewable generation, fixed distributed storage and
mobile storage (Electric Vehicles) are considered. However the study focused
only on economic aspects and selection of the wind and solar sites with the
highest probability of wind speed and solar radiation. Moreover the study
didn’t cover the grid performance, stability and security assessment.
In [54], the authors produced simulations of scenarios with 100 per cent
RE considering a copper plate model for the National Electricity Market (NEM).
This model study ignored the core of the future grid – network, but only
concentrated on generation–demand balance. In [55], the researchers tried to
determine the least cost mix of wind, solar, geothermal, gas and hydro gen-
erations for California in 2050. The generation mix includes dis-patchable
hydro, pumped hydro, natural gas, geothermal, and centralized solar ther-
mal with storage. The research conclusion is based on balancing and hourly-
time domain simulations. Again stability is not considered. Regarding the
role played by the storage in future grid, the authors of reference [56, 57, 58]
have tried different ways to evaluate the cost, performance of the grid with
different scale of storage systems are embedded.
From the above studies, the researchers mostly focused on to evaluate the
feasibility of RES generation to balance the future projected demand. The
overall performances of the grid were not addressed. Most of the studies
are based on a final stage of the evolution of the grid, thus the performances
of the grid during transition are not studied. However it’s obvious that the
transition from the classical to so called future grid will take at least decades
depending on policy, economical condition and technical development. An-
other major concern is that all of the above studies have used conventional
load models and neglected DR. In the long-term it can be expected that DR
will play a major role alongside storage and so affect the result of power sys-
tem studies significantly [37].
1.5. Focus of the Presented Work 19

1.5 Focus of the Presented Work

In this research work, a set of comprehensive new future grid security
assessment tools are proposed from security assessment in future grids plan-
ning and operation point of view.
As mentioned in the previous sections, time-domain simulation is widely
used for power system planning since only limited number of critical sce-
narios and contingencies need to be investigated. And if those scenarios and
contingencies are secure then the system under study is regarded as secure.
However, the greater complexity of future grids goes way beyond just di-
mensional scale. This will require new tools such as for scanning large num-
bers of network scenarios.
On the other hand, the new features of future grids impose more diverse
and fast changing operating conditions on existing online security assess-
ment tools, such as the widely used Online Dynamic Security Assessment
systems. Investigation is needed to evaluate performance of conventional
existing online security assessment tools and new tools need to be developed
for future grids operation.
The focus of this dissertation are, for future grid planning, to develop a
framework using machine learning to effectively assess the security of future
grids by fast analyzing a large amount of the scenarios; for future grids oper-
ation, to propose new ML based DSA approaches to address technique issues
brought by future grids’ diverse and fast changing operation conditions. Un-
supervised learning, supervised learning and semi-supervised learning techniques
are utilized in a set of proposed planning and operation security assessment
tools.
One of the major research aims is to propose a framework to assess the
security of future grids for planning purpose by analyzing large amount of
the scenarios, considering the new features which are not part of the present
system. The inherent intermittent production of Renewable Energy Source
(RES), inertia less, fast transition of system structure and diverse load flow
pattern, inverter dominated and new employed control schemes will greatly
affect the way to analyze security of the future grid. Another major aim is to
develop new IS based DSA tools, therefore to give the future grid operators
awareness of potential stability issues in a faster and more accurate way. The
aims of this research work are listed below:

• Design planning study scenarios of a future grid for a CSIRO project.

20 Chapter 1. Introduction

• Conduct transient stability study of the designed scenarios using time-

domain simulation method and analysis transient stability of the grid
affected by different new features.

• Propose an IS (unsupervised learning) based framework to perform fast

scanning of a large amount of scenarios for future grid planning. (Mar-
ket model is included in the framework.)

• Conduct fast scanning of small signal stability and voltage stability

margin using the proposed method.

• Propose an IS (supervised learning) based DSA tool to tackle reliability

and accuracy issues caused by topology change and renewable genera-
tion penetration in future grids.

• Propose an IS (semi-supervised learning) based DSA tool to tackle se-

curity assessment speed and accuracy issues caused by renewable gen-
eration penetration in future grids.

1.6 Structure of the Thesis

Other than the introduction chapter, the thesis has another 5 chapters.
During the period of candidature, works are carried out on power system
security assessment by conventional and state-of-the-art techniques. Innova-
tive approaches are proposed in chapters 3 to 5 and papers based on these
innovation are published.
In the second chapter, a CSIRO project focused on developing a future
grid model and scenarios for the security assessment of the future grid is
given. Scenario development will be discussed. Time-domain simulation
method is applied in the project evaluating transient stability performance.
The simulation considered 18 scenarios and hourly operating conditions are
evaluated over typical months.
A large number of scenarios are evaluated using time-domain simulation
in future grids planning and simulation burden is unmanageable even for
offline application. The third chapter presents a novel unsupervised learning
algorithm based fast scanning tool. Small signal stability and voltage stability
margin are evaluated for a future grid model based on the IEEE 14-generator
benchmark network system.
New features of future grids such as intermittent renewable generation,
demand response, utility scale storage, etc. challenges the existing IS based
1.7. Methodology 21

DSA tools. The fourth chapter evaluate the performance of a conventional

IS based DSA tool and propose novel techniques to overcome the reliability
and accuracy issues if traditional supervised learning algorithm is used in a
DSA tool.
To achieve fast security assessment and high accuracy in future grids op-
eration, a novel semi-supervised learning algorithm based DSA model is pro-
posed in the fifth chapter. Semi-supervised learning algorithm uses both la-
beled and unlabeled operating conditions in the IS training process and sig-
nificantly reduced time-domain simulation required to prepare training data
for the DSA.
The last chapter concludes the thesis by revisiting key achievements pre-
sented in previous chapters and identify possible future works in developing
tools for future grids security assessment in both planning and operation.

1.7 Methodology
Conventional time-domain simulation method is used in the CSIRO fu-
ture grid project security assessment work. A large amount of scenarios need
to be assessed and which leads to heavy computational burden. To overcome
the issue, the first research target is to develop a framework for fast scanning
of future grids security, thus to have an overview of a future grid security
performance from the planning point of view. A novel ML based security
scanning tool for future grids planning is proposed. Further, future grids op-
eration conditions are more diverse than existing networks and which jeop-
ardizes reliability and accuracy of conventional IS based DSA tools. The sec-
ond research target is to propose novel ML based DSA tools for future grids
on-line security assessment. The methodologies are adopted to carry out the
research work are provided:
There are different ways to define the types of machine learning algo-
rithms but commonly the algorithms can be divided into categories accord-
ing to their purpose. The main categories are unsupervised Learning, super-
vised learning, semi-supervised Learning and reinforcement Learning.

• To achieve fast scanning of security performance of the future grid sce-

narios, unsupervised machine learning algorithm is considered. The
algorithm is mainly used in pattern detection and descriptive model-
ing. However, there are no output categories or labels here based on
which the algorithm can try to model relationships. Clustering algo-
rithm will be used to categories large number of operating conditions
22 Chapter 1. Introduction

and only ’typical’ operating conditions are evaluated by time-domain

simulation. Feature selection is required to pick best features for oper-
ating condition clustering.

• For future grids on-line security assessment, supervised machine learn-

ing algorithm is considered since the algorithm is more accurate than
the unsupervised algorithm when evaluate imminent operating condi-
tions. Supervised learning algorithms try to model relationships and
dependencies between the target prediction output and the input fea-
tures such that we can predict the output values for new data based
on those relationships which it learned from the previous data sets.
Novel feature selection method will be discussed, other advanced ma-
chine learning methods such as ensemble learning will be considered
to achieve a more reliable and accurate DSA tool for future grids.

• In the previous two algorithms, either there are no labels for all the
observation in the data set or labels are present for all the observa-
tions. Semi-supervised learning falls in between these two. Supervised
learning algorithms require large number of training samples which are
normally from time-domain simulation. For future grids on-line secu-
rity assessment, the requirement of time-domain simulation for train-
ing samples can be relaxed by semi-supervised learning algorithm. A
novel DSA model using semi-supervised learning algorithm will be
discussed. Data editing method will be used to overcome the disad-
vantage of the semi-supervised algorithm.

• In supervised and semi-supervised algorithms, neural network will be

used as the core of the machine learning based tools. A new learning
algorithm - Extreme Learning Machine (ELM) will be considered.

• A simplified 14-generator network model of southern and eastern Aus-

tralia is used as a starting point for the CSIRO scenarios study. How-
ever, works are required to upgrade the network step by step and fi-
nally reach a desired future grid which is based on an AEMO 100 Per-
cent Renewable Study [52], and other models [53]. In this future grid
model, variable energy sources (e.g. wind and solar), new transmission
technologies (e.g. Flexible Alternating Current Transmission System
(FACTS) and HVDC) and responsive loads will be considered. The ex-
isting FACTS are kept as they are but new control loops may be added.
1.7. Methodology 23

Standard RES models will be updated to realize ancillary services func-

tionality. Load centres will be located at locations mentioned in the
study, Distributed conventional synchronous generators driven by re-
newable energies such as gas turbines, which are used to pick up loads
when RES are not available, are located close to load centres.

• Another Institute of Electrical and Electronics Engineers power system

benchmark network model - New England System - is used in novel
DSA tools using supervised and semi-supervised algorithms due to its
simplicity. However, the model is updated with renewable generation
and operating conditions are generated following the way in the CSIRO
future grid project.

• In terms of simulation tool, PowerFactory is a very powerful simulation

system for power system analysis. It incorporates a comprehensive list
of simulation functions including the following major packages: Eigen-
value Analysis, Contingency Analysis, RMS Simulation, Load Flow
Analysis, Protection Analysis and more other functionality. The Pow-
erFactory software will be the major simulation tool to carry out neces-
sary time-domain studies. DSAToolsTM is another most popular power
system analysis tools and provides comprehensive system security as-
sessment capabilities including Powerflow & Short circuit Analysis Tool
PSAT, Voltage Security Assessment Tool VSAT, Transient Security As-
sessment Tool TSAT and Small Signal Analysis Tool SSAT. The power-
ful package of SSAT provides the user deep insight of the system small
signal stability which will be used particularly in the research.
 25

Chapter 2

The CSIRO Future Grid Study

Project

2.1 Introduction
In the CSIRO project, I was involved in the study scenarios design and
was responsible for transient stability study from modeling, simulation, re-
sults analysis to paper writing. In this chapter, future grid study scenario
design and transient stability study conducted in the CSIRO project is pre-
sented. Different power system planning scenarios are discussed and tran-
sient stability performance of the scenarios are compared. Conventional time-
domain simulation method is used in the transient stability study.
The study work carried out in the second part showed very high compu-
tational burden of the conventional time-domain method which leads to the
research focus of this dissertation on machine learning methods application
in future grids security assessment.

2.2 Project Background

In conventional power systems, large thermal and hydro power plants
have provided balancing and stability control. Among different system needs,
a priority after basic balancing of power and energy is to ensure that power
flows and dynamics are within bounds and stable (for angle, voltage and fre-
quency) in normal operation and after events (faults, failures). However, the
conventional planning and control models, which are well-known and stan-
dardised, will be challenged by all new features of future grids: renewable
energy sources (which are less predictable) and distributed generation, cost
constraints on ‘poles and wires’ and new loads such as EVs which add new
peaks.The further requirements here are ramping ability and stability. Thus,
26 Chapter 2. The CSIRO Future Grid Study Project

the role of modelling and analysis related to balancing and stability for FG
scenarios remains of central importance.

2.2.1 Existing Future Grid Studies

Many studies on the FG have been published by researchers. But, the
majority of those research are focused on simple balancing studies by using
a range of copper plate transmission models [53, 59, 60, 61, 52, 62]
The Zero Carbon Australia 2020 Stationary Energy Plan (ZCA2020 Plan)
[53] aims to find a detailed and practical roadmap to decarbonise the Aus-
tralian stationary energy sector within a decade. In the plan, wind power
and Concentrating Solar Thermal (CST) with molten salt storage are the two
primary technologies used, Detailed modelling was undertaken to ensure
that the new renewable energy supply can meet all demand projected un-
der the ZCA2020 Plan, 24 hours a day, 7 days a week, 365 days a year. The
ZCA2020 Stationary Energy Plan describes how to repower Australia’s sta-
tionary energy sector using 100% renewable sources by 2020. The authors
acknowledge that the Plan detailed herein is not the only way that Australia
could achieve zero emissions from the stationary energy sector.
In [59], least cost options are presented for supplying the Australian Na-
tional Electricity Market with 100% renewable electricity using wind, photo-
voltaics, concentrating solar thermal with storage, hydroelectricity and bio-
fuelled gas turbines. authors use a genetic algorithm and an existing simula-
tion tool to identify the lowest cost (investment and operating) scenarios of
renewable technologies and locations for NEM regional hourly demand and
observed weather in 2010 using projected technology costs for 2030. A sim-
plified transmission network is used to estimate the annual cost to balance
supply and demand across NEM regions.
Combinations of renewable electricity sources are modeled (inland wind,
offshore wind, and photovoltaics) with electrochemical storage (batteries and
fuel cells), incorporated into a large grid system in [60]. The first target of this
study is to find an combination can provide smooth output with help from
storage. The second target is to seek minimal cost, calculating true cost of
electricity without subsidies and with inclusion of external costs. In total 28
billion combinations of renewables and storage are evaluated, each tested
over 35,040 hours (four years) of load and weather data.
2.2. Project Background 27

In [61], authors explore the potential for a 100% renewable electricity gen-
eration system with substantially increased levels of wind penetration, fossil-
fuelled electricity production was removed from an historic 3-year data set,
and replaced by modelled electricity production from wind, geothermal and
additional peaking options for the New Zealand power system. Genera-
tion mixes comprising 53–60% hydro, 22–25% wind, 12–14% geothermal, 1%
biomass and 0–12% additional peaking generation were found to be feasible
on an energy and power basis, whilst maintaining net hydro storage. Ap-
plication of the approach applied in this research to countries with different
energy resource mixes is discussed, and options for further research are out-
lined.
In 2011, the Australian Government announced its Clean Energy Future
Plan. As one initiative under that plan, AEMO conducted a study [52] which
explores two future scenarios featuring a National Electricity Market fueled
entirely by renewable resources. This study considers two scenarios with
differing views about how quickly renewable technologies will develop over
time. Accordingly, power systems with differing configurations are also con-
sidered to emerge in each scenario. As pointed out in the report, the findings
are tightly linked to the underlying assumptions and the constraints within
which the study was carried out. Any changes to the inputs, assumptions
and underlying sensitivities would result in considerably different outcomes.
The study in [62] examines the challenges of integrating significant vol-
umes of wind power generation onto the power system of Ireland. The report
provide the first significant modeling of power system behavior at unprece-
dented instantaneous penetrations of wind. According to the authors, the
findings are a key element towards meeting Ireland’s ambitious 2020 renew-
able energy targets.

2.2.2 Limitations of Existing Studies

All the above works neglected network related issues such as line con-
gestion and stability in their studies. Further, most studies use conventional
demand models, and ignore the influences of emerging demand-side tech-
nologies and the synergies that may arise between them when modeling net
future demand.
Moreover, all assume specific market arrangements by which RESs are
integrated into grid operations. This does not allow for the change and evo-
lution of such market institutions in response to technological developments
28 Chapter 2. The CSIRO Future Grid Study Project

that might better support the least-cost delivery of electrical power.

The CSIRO project is the observation that to the best of our knowledge,
no international research so far has developed a standardized comprehensive
modeling framework for future grids close to what have been accustomed to
for classical systems: a suite of definitions, equations and software for power
flow, stability analysis, dispatch, security and reliability.

2.3 Future Grids Security Assessment Framework

In this project, a future grid model is built and scenarios are designed and
evaluated using a simulation platform presented in [63] which consists of
market simulation, load flow calculation study and security assessment alto-
gether. The electricity market model is built in PLEXOS based on a modified
14-generator model, and the dispatch results from the market are used for
power flow study and security assessment.

2.3.1 Test System

In this work, we apply our scenario and sensitivity-based study frame-
work on the Australian FG. A modified 14 generator model of the Australian
NEM, which was initially developed for small signal stability studies [64], is
employed as the test bed. A single line diagram of the 14-generator model
of the NEM is shown in Fig.2.1. Areas 1 to 5 represent Snowy Hydro (SH),
New South Wales (NSW), Victoria (VIC), Queensland (QLD) and South Aus-
tralia (SA), respectively. All excitation system or Automatic Voltage Regu-
lation (AVR) and Power System Stabiliser (PSS) of generators are adopted
from [64]. Standard steam turbine governor model IEEEG1, gas turbine gov-
ernor GAST and hydro turbine governor HYGOV are used. The NEM has
been split into 16 zones according to AEMO’s planning document to capture
differences in generation technology capabilities, costs, weather and so on in
[65], as shown in Fig.2.2.

2.3.2 Simulation Preparation and Procedures

AEMO’s forecast hourly load profile in year 2040 is aggregated across
each region of the NEM [66]. The demand data for each region is distributed
among loads of the 14-generator model in a region based on their default
percentage values in the region. The modified 14-generator model is then
2.3. Future Grids Security Assessment Framework 29

F IGURE 2.1: Simplified 14-generator model.

modeled in PLEXOS and DIgSILENT PowerFactory for the market simula-

tions, balancing and security assessment, respectively.
Three major steps are involved in the future grid security assessment pro-
cedure and listed below.

• The first step after generation of scenarios is to perform a market sim-

ulation to provide the initial conditions/equilibria for system studies.
30 Chapter 2. The CSIRO Future Grid Study Project

F IGURE 2.2: NEM 16-zones by AEMO.

The market model used in this study is based on a recent work pre-
sented in [67] which uses a modified Unit Commitment (UC) problem
that includes the aggregated impact of prosumers (a generic demand
model which represents the aggregated effect of price-responsive users
equipped with rooftop PV-battery systems).

• In the next step of the future grid security assessment, load flow study
2.3. Future Grids Security Assessment Framework 31

is carried out using the dispatch results from the market model. Pow-
erflow and Short-Circut Analysis Tool (PSAT) of Dynamic Security As-
sessment Software (DSATools) are used for this purpose. The outputs
of the load flow will be used as initial conditions for power system se-
curity assessment.

• In the third step, using the steady state conditions as results of the load
flow study, power system security performance can be assessed to as-
sure that the future grid can be operated securely.

The following stability analysis were performed using simulation soft-

wares DSATools and DIgSILENT:

• Small signal stability (SSS): Damping ratio of the least stable rotor angle
mode in the system is calculated in DIgSILENT for each hour of the
simulated year using eigenvalue analysis method (QR method [26]).

• Transient stability: The Extended Equal Area Criteria (EEAC [68]) and
time-domain simulation in Transient Security Assessment Tool (TSAT)
were employed for TS assessment of the future grid, and TS index (TSI)
was calculated for each hour of the simulated year.

• Long-term voltage stability: Long-term VS in the system was evaluated

with the Voltage Security Analysis Tool (VSAT), and loadability margin
was calculated for each hour of the simulated year.

• Frequency stability (FS): The system FS was evaluated using time-domain

simulation in DIgSILENT. Minimum Rate of Change of Frequency (Ro-
CoF) and frequency nadir were used to assess the system frequency
behaviour after a contingency.

The calculation, interpretation and detailed explanation of the above tran-

sient stability indices are provided in Simulation Results. In the next section,
the test-bed assumptions and modeling as well as FG scenarios for the Aus-
tralian NEM are explained.

2.3.3 Assumptions and Scenarios Description

In this project, our FG scenarios are inspired by the published studies of

• 100% renewable scenario for the Australian NEM [69],

32 Chapter 2. The CSIRO Future Grid Study Project

• general outcomes of the CSIRO FG Forum [70, 71],

• AEMO 100% RES Study [52].

The study in [69] proposes the least cost mix of diverse RES technolo-
gies (including WF, hydro, biogas, utility PV and concentrated solar thermal
power plant with thermal storage) for an Australian FG. However, in that
research, the balancing studies are simplified through relieving some techni-
cal constraints including the grid model, ramp rate and minimum up/down
time of the generators, etc. Those constraints can change the dispatch results,
the energy share of different resources, and even the least-cost mix of diverse
RES technologies for the future of the NEM. Further, a simplified balancing
study without considering network and stability constraints can not guaran-
tee the feasibility of an operation scenario.
The proposed scenario and sensitivity based analysis is an approach which
aims to overcome the limitations of those highly simplified approaches. that
can address the structural balancing and stability issues and enlighten the
path that should be followed in the future in terms of generation and net-
work expansion, market design and the operation of FGs.
In our study, we obtained electricity demand, wind and solar traces from
AEMO’s National Transmission Network Development Plan (NTNDP) in
2040 [66]. The modified 14 generator model of the NEM, representing 100%
renewable generation portfolio is shown in Fig. 2.1. In this Figure, the Areas
1 to 5 represent Snowy Hydro (SH), New South Wales (NSW), Victoria (VIC),
Queensland (QLD) and South Australia (SA), respectively. We matched the
14-generator model with the 16 zones of the AEMO’s NTNDP [66] to extract
the corresponding wind, solar and demand traces for the market model.
In order to implement the suggested generation portfolio in [69] (which
is based on demand and weather data in 2010), first, we scaled up the capac-
ity of each generation technology based on the demand energy/peak power
growth in each region of the NEM in 2040. Second, since the 14-generator
model was originally developed based on the Australian grid in 2010, we
had to reinforce the transmission system to ensure the system balancing. In
the following sections, the features for sensitivity study are described.

• Network Strength Sensitivity

The suggested network augmentation in the AEMO 100% RES Study
[52] was used as a guideline for grid reinforcement. In addition, we
monitored the voltage angle differences between busbars connected by
2.3. Future Grids Security Assessment Framework 33

lines/transformers, and ensured that voltage angle differences between

such busbars remain within acceptable bound at all times. This has led
us to two levels of the network strength, i.e. Weak Network (Nw) and
Strong Network (Ns) for the sensitivity aspect of the grid strength.
In the Nw sensitivity, the transmission lines are augmented just enough
to ensure that the balance between demand and supply is maintained,
although, the capacity factor of RESs is low (15%-20%). The reduced
capacity factor is due to the curtailment of wind energy in some hours
caused by the congested transmission lines.
In the Ns sensitivity, transmission lines are further augmented to guar-
antee the system balancing as well as a reasonable capacity factor for
RESs, i.e. between 25%-35%. The augmented corridors in the 14 gener-
ator model of the NEM are also shown in Fig. 2.1. Table 2.2 compares
the installed capacity of different RES technologies for the scaled up
generation portfolio in Nw and Ns sensitivities. As it can be seen, the
installed capacity of RESs in the Nw sensitivity is considerably higher
than the scaled up generation portfolio in [69], which is due to very low
capacity factor of RESs. Observe in the table that the generation portfo-
lio in the Ns sensitivity is roughly the same as the scaled up generation
portfolio in [69].

• Prosumers sensitivities
Due to the emerging situation in Australia, where rooftop PV pene-
tration is increasing significantly, we considered scenarios cases where
prosumers equipped with rooftop PV play a significant role in re-shaping
the future demand. Further, the rooftop PV owners are increasingly
discouraged to export power back into the grid due to very low Feed
in Tariffs (FiTs) and increasing grid electricity costs. Therefore, it is
expected that most rooftop PV owners will install small-scale Battery
System (BS) to utilize Integrated PV-battery systems (IPBS) in order to
maximize their self-consumption. It appears likely that such a change
will occur globally, as also acknowledged in [72]. With this prospect,
we considered four different rooftop PV-battery uptake sensitivities as
follows:

– Medium uptake of rooftop PV, low uptake of BS (PmBl)

– Medium uptake of rooftop PV, high uptake of BS (PmBh)
– High uptake of rooftop PV, low uptake of BS (PhBl)
34 Chapter 2. The CSIRO Future Grid Study Project

– High uptake of rooftop PV, high uptake of BS (PhBh)

Prosumers are considered amongst residential and commercial customers,

whereas the industrial customers are left unaffected. The penetration of
prosumers in each region of the NEM are adopted from AEMO stud-
ies in [73, 74]. In the sensitivities with low uptake of BS, only a por-
tion of prosumers (10% to 40%, depending on the area of the NEM) are
equipped with IPBS, and, for 1kW rooftop PV, 1.8kWh BS is consid-
ered. In the sensitivities with high uptake of BS, however, it is assumed
that all the prosumers utilize IPBS. Furthermore, we performed a se-
ries of simulations to choose the ratio of household battery to rooftop
PV for prosumers in those scenarios, and realized that 1kW rooftop
PV/3kWh battery system allows prosumers to store excess generation
from rooftop PV without spilling.

• Utility storage sensitivities

Different penetrations of utility storage, i.e. zero (conventional demand),
low (Sl), and high (Sh) are considered in this study. Currently, there is
no utility storage installed in the NEM, and there are limited studies
conducted regarding its future prospect in the Australian FG. Never-
theless, some studies suggested that installation of utility storage will
take place in a foreseeable future in the NEM [70, 71]. In order to choose
the location of utility storage, we populated the grid with utility storage
and performed a series of simulations over the simulated year. Then,
the busbars where utility storage could reach its maximum State of
Charge (SoC) in most of the horizons were selected. The chosen bus-
bars turn out to be close to RESs (with many surplus hours), as they
provide cheap electricity in the grid. The surplus energy from RESs can
be stored in the utility storage, and then can be used during peak hours
and/or hours with lack of generation from RESs in order to reduce the
total electricity cost in the system. For choosing the capacity of util-
ity storage in each region, generation portfolio (including the surplus
power from RESs) as well as the load energy/peak demand (power) in
each region of the NEM were taken into account.
In the rest of this chapter, we used the introduced abbreviations in Table
2.1 to refer to different FG sensitivities discussed. Some examples are
given below to explain how the scenarios should be interpreted from
the abbreviations in Table 2.1:
2.4. Future Grid Transient Stability Study 35

– NwSlPzBz: Weak Network and Utility Storage penetration low

and PV penetration zero and Household Battery storage penetra-
tion zero.
– NsSzPhBh: Strong Network and Utility Storage penetration zero
and PV penetration high and Household Battery storage penetra-
tion high.

TABLE 2.1: Abbreviations Used in Study Scenarios Description

Symbol (element) Description Symbol (penetration) Description

N Network w/s weak/strong
S Utility Storage z/l/h zero/low/high
P PV z/m/h zero/med/high
B Household BS z/l/h zero/low/high

TABLE 2.2: Comparison of Generation Portfolio

Technology Scaled up in (GW) Network Weak (GW) Network Strong (GW)

WF 55.7 90.1 55
Utility PV 4.3 9.2 6.4
CST 3.1 5.2 3.6
GT 26.5 33 33
Hydro 7.5 9 9

2.4 Future Grid Transient Stability Study

In this project, the objective of the transient stability analysis is to examine
how various factors integrated in the scenarios affect the future grid transient
stability.
In the designed scenarios, the synchronous machine is used for CSP, hy-
dro and gas power stations. Though small in number, these conventional
synchronous generators with massive rotating rotors play an important role
in the future grid not only because they are working as dispatchable genera-
tors but also because they provide inertia to stabilize the system.
Researchers have done studies on how renewable generation impact on
power system transient stability. In [75], the authors investigated the ef-
fect of high PV penetration on a large system. Both utility scale and dis-
tributed rooftop PV are modeled in their work. In [76], the authors present
a stochastic-based approach to evaluate the probabilistic transient stability
36 Chapter 2. The CSIRO Future Grid Study Project

indices of a power system incorporating WFs. In [77], the authors study if

potential areas in a region of a Mediterranean partner country (MPC) are
suitable (in terms of dynamic performance) for solar and wind generation
integration. In [78], the impact of DFIG wind farm and its geographical dis-
persion on the power system transient stability in different cases are studied.
In [79], the wind farm layout and wake effect on the power system transient
stability are tackled. In general, simulation results of these studies reveal
that, renewable generation penetration level, system topology, disturbance
type as well as the location are all important factors in determining the na-
ture of the impact on the system dynamic performance.
However, the existing studies all focus on very few factors which are con-
sidered as the features of the future grid, such as including PV farm, wind
farm in a study. Another common feature of the exiting studies is that they all
followed conventional stability assessment way by only study a few typical
operating conditions and try to draw conclusion based on the results.
Different to the previous studies, in this project we consider more major
potential features of the future grid, such as demand response, household
battery storage, utility storage, etc. Moreover, the stability study is based on
a scanning process rather than a few isolated operating points. The results
of the stability scanning would give deeper insight of how the renewable
generation reshape stability performance in different scenarios.

2.5 Transient Stability and Stability Indices

Large-disturbance rotor angle stability or transient stability, as it is com-
monly referred to, is concerned with the ability of the power system to main-
tain synchronism when subjected to a severe disturbance, such as a short cir-
cuit on a transmission line. The resulting system response involves large ex-
cursions of generator rotor angles and is influenced by the nonlinear power-
angle relationship [17].
In all studied scenarios in this project, majority generators in the systems
are non-synchronous machines; the synchronous machine is used for CSP,
hydro and gas power stations. Though small in numbers, these conventional
synchronous generators with massive rotating rotors play an important role
in the future grid to not only working as dispatchable resources but also pro-
vide inertia to stabilize the system. Study transient stability of the different
scenarios and analyse how different factors impact on the transient stability
indices are important.
2.6. Extended Equal-Area Criterion Brief 37

Time-domain simulation and direct method are playing important roles

in offline and online power system stability assessments. To achieve fast tran-
sient stability assessment, many work have been done by researchers con-
centrated on the stability criterion based on the system dynamic response,
including energy function analysis [80], phase-plane trajectories [81, 82, 83],
equal area criterion citePaudyal2010, Mariotto2010 and projection energy func-
tion based on system trajectories [84]. Among the different transient stability
indices, Extended Equal-Area Criterion (EEAC) [39] is approved to be an ef-
ficient and reliable method.

2.6 Extended Equal-Area Criterion Brief

The extended equal area criterion is a direct type method. It aims at en-
hancing and broadening the advantages of the Lyapunov criterion, by fur-
nishing analytical expressions for ultra-fast analysis, sensitivity analysis and
means to preventive control. To reach these objectives, the EEAC uses some
conjecture, assumption, and approximation together with the equal area cri-
terion (or equivalently the Lyapunov direct criterion) [39]. Extended Equal-
Area Criterion is one of optional transient stability criterion in TSAT simula-
tion software.

2.7 Simulation Results

The previously designed scenarios are used in this section to carry out
transient stability study. Synchronous generator dynamic model is kept from
the 14-generator system model for hydro, gas and CSP generators. The WECC
type IV power converter dynamic model is used for renewable generators.
Due to time-consuming time domain simulations, a typical week of each sea-
son in the year 2040 is chosen to conduct transient stability scanning. The
equilibrium operating points for transient stability study are the result of the
market model explained in previous sections. TSAT is a professional power
system study tool which is used for transient stability study in this project.
In TSAT settings, transient security index for base case analysis is selected
as stability margin and the stability margin algorithm is selected to be angle
margin which uses EEAC as the transient stability criterion. A three phase
short circuit fault is applied on all 79 buses in all scenarios. The fault starts at
100ms and lasts for 100ms then it is removed with no equipment taken out
of service.
38 Chapter 2. The CSIRO Future Grid Study Project

Each time-domain simulation gives a stability level for that operating

point; the stability level ranges between -100 (most unstable) to 100 (most
stable). The stability levels of all operating points of a scenario are averaged
in two ways for comparison purpose later on. For a particular contingency
the stability levels of the system through out the weeks (672 hours) are aver-
aged, the new stability index gives general stability level of the system due
to this contingency; vulnerable buses can be identified using this index. The
formula to calculate this stability level index is given below.
P672
T SI c (h)
h=1
T SI1 = T SI(c) = (1)
672
Another stability level index is calculated by averaging the stability levels
as results of 79 contingencies applied on buses (79 buses) for each hour in
the week. This new stability level index gives general stability level of the
system at each hour across the week. The formula to calculate the index is
given below.
P79
c=1 T SI h (c)
T SI2 = T SI(h) = (2)
79

100% -100

90% -80

80% -60

70% -40

60% -20

50% 0

40% 20

30% 40

20% 60

10% 80

100
N lPh z

l
l

N Ph h

hP Bz
Ph h

Pz z

hP hBl
h
N Ph l
N lPz h

D l
z
N mD l
N zPm Bz

l
N hD l

Bz

hB
hB
Sh hB

B
B

zP hB
B

B
Sh zB

zP zB

B
Sz mB

sS hB
S hB

zP m
Sz Dm

Sz Dh

w Dz

z
S Dz

N zDz
sS m

hD
sS D

N PhD
D

sS D

N PzD
D

sS D
sS mD

N hD
N Pz
N zPz

m
P

l
z

l
z
P

sS
S

sS
S

sS
w
w

w
w

w
w

N
w
w
N

F IGURE 2.3: Hourly transient stability indices of all studied sce-

narios.
2.7. Simulation Results 39

In Fig. 2.3 transient stability index TSI1 is used. A colormap is used to

denote different stability level as simulation results for all 18 scenarios. For
each scenario, simulation results of 79 faults are averaged hour by hour and
then sorted in order. The colormap uses deep blue to stand for very unstable
(-100) and deep yellow to stand for very stable (+100). Each column in the
figure corresponds to one scenario and the ordered hourly stability indices
are given different color according their values.
For weak network scenarios, NwSzPzDzBz, NwSzPmDmBl and NwSzPhD-
hBl are less stable among scenarios. For a stronger network, NsSzPzDzBz is
the most unstable scenario.
Figure 2.4 uses another stability index TSI2 which is the hourly averaged
transient stability index for each fault (bus). The figure shows the stability
level of all 79 buses for all 18 scenarios. Same as in Figure 2.1, colormap is
used to demonstrate different stability levels in all cases. Since most of the
studied operating points are stable for applied faults, the average stability
levels are all positive however which does not mean all operating points are
stable.

100

NsShPhDhBl
NsSlPhDhBl 90
NsShPzDzBz
NsSlPzDzBz
80
NsSzPhDhBh
NsSzPhDhBl
NsSzPmDmBh 70
NsSzPmDmBl
NsSzPzDzBz
60
NwShPhDhBl
NwSlPhDhBl
NwShPzDzBz 50
NwSlPzDzBz
NwSzPhDhBh
40
NwSzPhDhBl
NwSzPmDmBh
NwSzPmDmBl 30
NwSzPzDzBz
Bus01

Bus10

Bus20

Bus30

Bus40

Bus50

Bus60

Bus70

Bus79

F IGURE 2.4: Projection of transient stability indices on all buses

of all studied scenarios.
From Figure 2.4, it is easy to identify the most vulnerable buses. Bus
40 Chapter 2. The CSIRO Future Grid Study Project

10 and bus 11 are the most vulnerable buses in the weak network scenar-
ios, however this is not observed in strong network scenarios. In all scenar-
ios, bus 41 seems to be the most vulnerable one and especially for scenarios
NwSzPhDhBl, NsSzPzDzBz, NsSlPzDzBz and NsShPzDzBz.
In general, we can observe that network strength plays a big role in the
system transient stability performance. Demand response also helps to sta-
bilize the system. However, how utility level storage and residential battery
capacity affect the transient stability is not quite clear and need further study.

2.8 Security Assessment Simulation Burden

The transient stability study of the future grid is based on AEMO’s fore-
casted demand and renewable generation in the year 2040 with one hour
interval for the year. The test system is relatively small with 14 aggregated
generators, 28 loads, 59 buses and 90 transmission lines. The transient sta-
bility is done by time-domain simulation using EEAC method. A high per-
formance PC is used to perform the simulation. From the simulation, power
flow study took about 15 minutes to complete; about 4 days are required
for transient stability study considering 59 contingencies; about 5 hours re-
quired to carry out the Small signal stability study and about 9 days required
to finish the voltage stability study.
Considering a real world power system which is much bigger in scale, the
computation burden of a full range of security scanning for planning purpose
to find an optimal upgrading pathway is unbearable, even in the simulation,
a direct method EEAC is used. Obviously, in order to achieve a future grid
security assessment for planning requires a new method to overcome the
computational burden. The next chapter presents a new machine learning
tool for fast stability scanning using un-supervised method.
 41

Chapter 3

Un-supervised machine Learning

Method for Fast Stability Scanning

3.1 Introduction
This chapter is based on my POWERCON paper [85] and the first journal
paper [86] focusing on fast scanning of small signal and voltage stability in-
dices for long term planning study. Clustering is core of the task and it is an
unsupervised machine learning method.
Unsupervised learning is a type of machine learning algorithm used to
draw inferences from data sets consisting of input data without labeled re-
sponses. The goal for unsupervised learning is to model the underlying
structure or distribution in the data in order to learn more about the data. In
the power system security assessment area, the data sets are operating con-
ditions and contingencies, the responses are referred to the system security
levels under the given operating conditions and contingencies.

3.2 Background
Power systems are undergoing a major transformation driven by the in-
creasing uptake of renewable energy sources, DC power transmission, and
the decentralization of electric power supply underpinned by the informa-
tion and communication technologies and demand-side technologies, like
rooftop PV, energy storage, home energy management systems, and electric
vehicles. How future grids will look like, however, is still uncertain as the
evolution depends not only on technological development but also on the
regulatory environment.
 Chapter 3. Un-supervised machine Learning Method for Fast Stability
42
Scanning

Therefore, one of the challenges associated with future grid planning is

that the structure of a future grid cannot be simply extrapolated from the
existing one.

3.2.1 Existing Studies

As an example, the emergence of prosumers1 might change the demand
profile, which results in a significantly different stability performance, as
demonstrated in [87]. Instead, for future grids planning, several possible
evolution paths need to be accounted for. Future grid planning is different to
the conventional power system planning and the later only analyze a hand-
ful of the most critical scenarios. To account for a wide range of possible
future evolutions, scenario analysis has been proposed in many industries,
e.g. in finance and economics [88], and in energy [89, 90, 91, 92]. As opposed
to the conventional power system planning, where the aim is to find an opti-
mal transmission and/or generation expansion plan for an existing grid, the
aim in scenario analysis is to analyze possible evolution pathways to inform
power system planning and policy making. Given the uncertainty associated
with long-term projections, the focus of future grid scenario analysis should
focus on analyzing what is technically possible, although it might also con-
sider an explicit costing [93]. Therefore, future grids’ planning may involve
large amount of scenarios and the existing planning tools may no longer suit-
able.
Future grid analysis is a growing research area. Melbourne Energy In-
stitute [53] have proposed a possible plan for a future Australian grid rely-
ing 100% on renewable energy sources. The Centre for Energy and Environ-
mental Markets at the UNSW [54, 94] has shown for the Australian National
Electricity Market that balancing a 100% RES power system is technically
possible. The PJM study [95] has shown that the PJM network can be pow-
ered 90-99.9% of the time entirely on RESs, at a cost comparable to today’s.
The existing studies, however, only focus on balancing and use a simplified
copper plate model of the transmission network. They also neglect stability
analysis, which limits their value.
Stability analysis is an important task in power system planning. In con-
ventional stability analysis, only a small number of worst-case critical con-
ditions is typically analyzed. If stable under those conditions, the system is
assumed stable in all possible credible operating conditions. The selection
1
Consumer with generation (e.g.rooftop-PV) and battery storage (producer-consumer).
3.2. Background 43

of the critical conditions is most often based on the historical performance,

and planners’ experience and judgment [96, 97, 75, 98, 99]. In power systems
with significant penetration of intermittent RES, the generation dispatch and
the associated power flows change many times throughout the day and often
follow rather different seasonal patterns, which renders past operational ex-
perience of limited value. Although the authors of a future grid study [100]
selected a few critical operation points for stability analysis, they also pointed
out that there is no guarantee that these cases are necessarily the most diffi-
cult ones.

3.2.2 Motivation Behind the Study

Chronological time series scanning offers a way for the stability analysis
of a power system with a constantly varying operating conditions, and to
capture the inter-seasonal variations in renewable generation. With time se-
ries scanning, it is possible to capture stability performance over a long hori-
zon. The authors in [101] have demonstrated the value of using time-series
analysis for steady-state voltage stability analysis of a power system with
high penetrations of wind. They have shown that in contrast to traditional
power systems without intermittent generation, in a system with a high RES
penetration, the worst case operating point shifts. The time-consuming time-
series simulation, however, was not discussed in [101]. Instead, the worst
case points were manually picked from several years worth of data, and
the simulations were performed around these points to reduce the compu-
tational burden.
The Future Grid Research Program funded by the Australian Common-
wealth Scientific and Industrial Research Organisation is to propose a com-
prehensive modeling framework for future grid scenario analysis. The aim
of the project is to explore possible future pathways for the evolution of the
Australian grid out to 2050 by looking beyond simple balancing. To this end,
a simulation platform has been proposed in [63] that consists of a market
model, power flow analysis, and stability analysis. Preliminary results have
shown, however, that time-series scanning over a one-year horizon is com-
putationally very expensive as demonstrated in Chapter 2.. To speed-up the
computation, we propose a machine learning based framework for fast sta-
bility scanning. The efficacy of the framework is demonstrated on a simpli-
fied 14-generator model of the Australian National Electricity Market.
 Chapter 3. Un-supervised machine Learning Method for Fast Stability
44
Scanning

3.3 Contribution of This Study

The main contribution of this study is a framework for fast stability scan-
ning using ML. The ML techniques used in the study (clustering, feature se-
lection, PSO-k-means) have been used in the past in many power engineering
applications [102, 103, 104, 105, 106, 107, 1, 6, 2, 3, 108, 4, 109, 110]. Imple-
menting them directly to the problem of fast stability scanning, however, is
infeasible because the number of operating conditions and the number of
features required to describe an operating condition are much larger than in
the existing applications. To address that, we propose a self-adaptive PSO
k-means clustering algorithm that considers both the adjusted feature ranks
and weights for clustering and optimal selection of the number of clusters.
The rest of the chapter is organized as follows: Section 3.4 outlines the
simulation platform for future grid scenario analysis. Section 3.5 gives an
overview of the application of ML in power systems and describes the per-
tinent ML algorithms. Section 3.6 proposes a novel fast stability scanning
framework. In Section 3.7, the efficacy of the proposed framework is demon-
strated on a simplified 14-generator network model of the Australian Na-
tional Electricity Market. Section 3.8 concludes the chapter.

3.4 Simulation Platform

We use the simulation platform for future grid scenario analysis originally
proposed in [63] as the basis, summarized in Algorithm 1. The platform
consists of four modules: (i) scenario generation, (ii) market simulation, (iii)
load flow analysis, and (iv) stability analysis, described in more detail later.
The other three modules remain the same.

Algorithm 1 Future grid scenario analysis.

Input: Network data, generation data, wind, solar and demand traces for
each scenario s ∈ S in the studied year.
Output: Stability indices for each time slot t ∈ T , for each scenario s ∈ S.
1: for s ← 1, |S| do
2: for t ← 1, |T | do
3: Market simulation (generation dispatch);
4: Load-flow analysis;
5: end for
6: for t ← 1, |T | do
7: Stability analysis (voltage, angle, frequency);
8: end for
9: end for
3.4. Simulation Platform 45

F IGURE 3.1: 14-generator test system.

 Chapter 3. Un-supervised machine Learning Method for Fast Stability
46
Scanning

3.4.1 Test System

We use a modified 14-generator IEEE test system that was initially pro-
posed in [64] as a test bed for small-signal analysis. The system is loosely
based on the Australian National Electricity Market, the interconnection on
the Australian eastern seaboard. The network is stringy, with large trans-
mission distances and loads concentrated in a few load centers. It consists
of 59 buses, 28 loads and 14 generators, each representing a power station
consisting of between 2 to 12 units, resulting in a total of 74 synchronous
machines. The single-line diagram of the test-bed is illustrated in Fig. 3.1,
in which Areas 1 to 5 represent Snowy Hydro , New South Wales, Victoria,
Queensland and South Australia, respectively. Areas 1 and 2 are electrically
closely coupled, hence the system has four distinct areas.

3.4.2 Scenario Description (Line 1 in Algorithm 1)

Given that the focus of the study is fast stability scanning, we only analyze
one future grid scenario. We augmented the test system by replacing conven-
tional synchronous generators at selected buses with Wind Farms (WF) and
PV farms, and a concentrated solar thermal plant, as shown in Fig. 3.1, re-
sulting in 30% RES energy penetration. To increase the transfer capacity of
the network, we added HVDC links between buses 412 and 211, 216 and 313,
305 and 508, reinforced the existing AC transmission corridors and added
static var compensator to improve voltage control. We used wind, solar and
demand predictions for the year 2030 from the Australian Energy Market
Operator’s National Transmission Network Development Plan [111].

3.4.3 Time-series Analysis (Lines 2-5 in Algorithm 1)

Time-series analysis consists of market simulation and load-flow analysis
using the generation dispatch results. To capture the inter-seasonal varia-
tions in renewable generation and the demand, we need to analyze a full
year, which results in |T | = 8760 assuming hourly resolution.

3.4.3.1 Market Model (Line 3 in Algorithm 1)

The aim of the market model is to emulate the outcome of an efficient

electricity market without assuming any particular market structure. The
model is based on a unit commitment problem aiming to minimize total
electricity generation cost, and is subject to the following constraints: power
3.4. Simulation Platform 47

balance, spinning reserve, power generation limit, start-up and shut-down

constraints, ramp rate limits, generator minimum up time restrictions, and
generator minimum down time restrictions. To achieve an acceptable com-
putational performance, the resulting mixed-integer optimization problem is
solved using a rolling horizon approach with hourly resolution. The decision
horizon is two days, where the solution for the first day is retained, and the
solution of the next day overlaps with the next two-day horizon. We assume
that generators bid at their respective short-run marginal cost, which we as-
sume to be zero for RES. A more complete description of the model is given
in [112].

3.4.3.2 Load-flow Analysis (Line 4 in Algorithm 1)

Load-flow analysis uses the dispatch results of market simulation and

the load traces from [111]. RES are assumed to operate in a voltage-control
mode. With hourly resolution, we obtain 8760 operating points, or instances,
representing the year 2030. Each operating point is represented by a set of
steady-state power system variables, or features. The operating points re-
sulting from the time-series analysis are used for stability analysis.

3.4.4 Stability Analysis (Lines 6-8 in Algorithm 1)

In this study, we focus on small-signal and static voltage stability, al-
though the simulation platform can also cover large-disturbance angle (tran-
sient) stability and frequency stability [113].

3.4.4.1 Modal Analysis

Small disturbance (or small-signal) rotor angle stability is concerned with

the ability of a power system to maintain synchronism under small distur-
bances [17]. Small signal stability problems are usually due to lack of damp-
ing. Inter-area oscillation modes may cause large power swing across inter-
connectors and can lead to system collapse or splitting. In this study, the
system exhibits a poorly damped inter-area mode between NSW and QLD,
which is the focus in stability scanning. We use modal analysis of a power
system model linearized around the current operating point.
 Chapter 3. Un-supervised machine Learning Method for Fast Stability
48
Scanning

3.4.4.2 Steady-state Voltage Stability

Voltage stability refers to the ability of a power system to maintain steady

voltages at all buses in the system from a given initial operating condition[17].
Voltage stability problems are typically associated with lack of reactive power
support, which can result from heavily load transmission lines. In systems
with high RES penetration, as in this study, this is of particular importance
given the constantly varying power infeed. Several stability indices have
been proposed for voltage stability assessment, giving a measure of the dis-
tance of the current operating point from the voltage collapse point [114]. In
this study, we used the loadability margin for stability scanning.

3.5 Clustering and Feature Selection

ML has been used in different power systems applications [102, 103, 104,
105, 106, 107, 1, 6, 2, 3, 108, 4, 109, 110]. In online dynamic security assess-
ment2 , ML is used for classification, to map a system operating condition
into a suitable stability index, for example for voltage stability [102, 103] and
on-line transient stability assessment [104, 105].
The classification of a system security status consists of three steps: (i) a
large database is generated using time-domain simulation to create a training
set; (ii) a set of features that best describe an operating condition is selected
as the inputs of the classifier; and (iii) the classifier is trained using an appro-
priate tool, e.g. an artificial neural network [102, 103, 106], a support vector
machine [107], or a decision tree [1].
To cover a large amount of possible operation conditions and to achieve
an acceptable level of accuracy, however, the training set is normally very
big—thousands of operation points for a relatively large conventional sys-
tems with no RES [6]. One possible way to address the problem is to reduce
the size of the training data set by limiting or fixing the load or generation
variation range for imminent hours only [106, 1].
For a study of a future power system with high RES penetration, the pos-
sible operating space is even larger. Therefore, a direct application of ex-
isting DSA methods becomes infeasible. Instead, as proposed in this study,
clustering is required to reduce the number of operating points for stability
scanning.
2
Online security assessment involves both dynamic and static security assessment. The
term dynamic security assessment is usually used to denote both.
3.5. Clustering and Feature Selection 49

Clustering, in particular the k-means algorithm, has been used in power

systems for scenario reduction before, for example to group substation load
profiles based on the chronological demand curves’ magnitudes [2, 3]; to
group similar P and Q patterns before classification [108]; to group similar
bus voltage and frequency response signals in order to locate outliers [4]; and
to group similar wind farm power and load level patterns to reduce simula-
tion time for the transfer capability assessment [109, 110]. When the number
of features is small, implementing a clustering algorithm is straightforward.
For example, [2, 3, 4] require only one feature, and [108, 109, 110] two fea-
tures. In our application, on the other hand, the operating scenarios are de-
fined by hundreds of correlated features, which requires some modifications
to the conventional clustering algorithms. In particular, we need to consider
the importance or the weight of each feature on the stability index, which we
discuss in the next section. Before describing the proposed framework for
fast stability scanning, we first describe the two pertinent ML algorithms, i.e.
k-means for clustering and ReliefF for feature selection.

3.5.1 Clustering
Clustering is the task of grouping a set of objects into clusters based on
their similarity [115]. A cluster is described by its internal homogeneity and
the external separation, i.e., patterns in the same cluster should be similar to
each other while patterns in different clusters should not [115]. When cluster-
ing a large amount of data, their similarity is usually expressed as a distance.
After clustering, all elements within a particular cluster can be represented
by the center of this cluster or a cluster centroid. In power systems, cluster-
ing is a popular ML algorithm used for dimensionality reduction. It has been
used in load forecasting [3], to accelerate the convergence of the Monte Carlo
simulations in transfer capability analysis [109], and to study the influence of
power flows on the damping of critical oscillatory modes [116].

3.5.1.1 k-means Algorithm

Among many data clustering methods, k-means algorithm is one of the

most often used methods for clustering. This method is very simple and es-
pecially suitable for large data sets and can be easily implemented in solving
many practical problems.
For a given data set X = {xi | xi ∈ Rn , i = 1, 2, ..., n}, the algorithm
partitions the data into k clusters, C1 , C2 , ..., Ck , where c1 , c2 , ..., ck are cluster
 Chapter 3. Un-supervised machine Learning Method for Fast Stability
50
Scanning

centroids or cluster means, defined as:

1 X
cj = x, (1)
Nj x∈C
j

where Nj is the number of data points in cluster j. Conventionally, k is an

input parameter to the algorithm. The similarity of the data in a cluster is
defined as their Euclidean distance to the cluster centroid. In Cartesian co-
ordinates, the Euclidean distance between two points xi and xj is defined
as: v
u n
uX
d(xi , xj ) = t wh (xih − xjh )2 , (2)
h=1

where feature weights wh are set to 1 in the conventional k-means algorithm,

and h denotes the dimensionality of the feature.
The k-means algorithm can be cast as an optimization problem with the
following objective:
Xk X
arg min kx − ci k2 (3)
C
i=1 x∈Ci

This is a NP-hard problem, for which several efficient heuristic solution

techniques have been proposed [115]. It is efficient in clustering large data
sets, however being a non-convex problem, it often terminates in local op-
tima.

3.5.1.2 Particle Swarm Optimization (PSO)

PSO is a population-based stochastic search process used to solve global

optimization problems where conventional mathematical programming ap-
proaches fail [117]. In the PSO, a swarm consists of a number of potential
solutions to the optimization problem, where each particle of the swarm cor-
responds to a potential solution. In the context of clustering, a single particle
represents a group of cluster centroids. The aim of the PSO is to find the
position of a particle that results in the best evaluation of a given objective
function, in our case the Sum of the Mean Squared Error (SMSE) defined as:
 
1 X Nc
 1
X
Je = d(xi , cj ), (4)
Nc j=1 |Cj |
x ∈C i j

where Nc is the size of the cluster centroid vector, cj is a cluster centroid

defined in (1), |Cj | is the number of data vectors belonging to cluster Cj , and
3.5. Clustering and Feature Selection 51

d(·) is the Euclidean distance defined in (2). To search for the best solution in a
multi-dimensional space, the particles “fly” through the space with different
speeds and directions. In the searching process, the fitness (4) of each particle
is evaluated and stored. The historical best position of each particle pbest and
the global best position gbest among all the particles are used to adjust the
flying speed and the direction of the particles.
The velocity of each particle is updated according to:

vi (n + 1) = w · vi (n) + c1 · rand1 · (pbest − pi (n))

(5)
+c2 · rand2 · (gbest − pi (n))

where c1 and c2 are constants, rand1 ∈ [0, 1] and rand2 ∈ [0, 1] are randomly
generated numbers, and w is the inertia factor defined as:

wmax − wmin
w = wmax − niter · . (6)
Niter

The particles’ position are iteratively updated as follows:

pi (n + 1) = pi (n) + vi (n + 1). (7)

In [118, 119], the authors have demonstrated that the combination of PSO
and k-means clustering can improve the clustering performance or, to some
extent, overcome the weaknesses of the k-means algorithm. We build on that
by proposing an improved self-adaptive PSO-k-means clustering algorithm,
discussed in more detail in Section IV.B.

3.5.2 Feature Selection

An operating condition of a power system is defined by a set of system
variables, or features, e.g. generator active and reactive powers, bus voltage
magnitudes and angles, load levels, etc. Feature selection is a process of se-
lecting a subset of relevant features that is necessary and sufficient to describe
the target concept by reducing the dimensionality of the input data and en-
hancing generalization by reducing over-fitting [120]. Feature selection has
attracted significant attention in DSA, e.g. in [121, 107, 122].

3.5.2.1 Relief Algorithm

A popular feature selection algorithm with little application in power sys-

tems is ReliefF [123, 120]. The main idea of the original Relief algorithm
 Chapter 3. Un-supervised machine Learning Method for Fast Stability
52
Scanning

[123] is to estimate features’ ability, represented by features’ weights, to dis-

tinguish between instances, power system operating conditions in our case,
that are near to each other.

Algorithm 2 RReliefF feature selection algorithm [120]

Input: For each training instance r ∈ R a vector of attribute values a ∈ A
and predicted values λ ∈ L.
Output: For each training instance r ∈ R a vector w ∈ R|A| of estimations of
the qualities of attributes a ∈ A.
1: Set all w to 0;
2: for i ← 1, m do
3: Randomly select instance ri ;
4: Select k instances qj nearest to ri ;
5: for j ← 1, k do
6: ndc ← ndc + diff (λ(·), ri , qj ) · d(ri , qj )
7: for l ← 1, |A| do
8: nda da
l ← nl + diff (l, ri , qj ) · d(ri , qj )
9: ndca
l ← ndca +
10: diff (τ (·), ri , qj ) · diff (l, ri , qj ) · d(ri , qj )
11: end for
12: end for
13: end for
14: for l ← 1, a do
15: wl ← ndca /ndc − (nda − ndca )/(m − ndc )
16: end for

The original Relief algorithm [123] is limited to two-class problems. Its

extensions, ReliefF and RReliefF can also deal with multi-class and regression
problems, respectively [120]. The pseudo code for the RReliefF algorithm
used in this study is shown in Algorithm 2, where ndc , nda , and ndca denote
the weights for the prediction values of different prediction (line 6), different
attribute (lines 8) and for different prediction and different attribute (line 9
and 10), respectively.
The term d(ri , qj ) takes into account the distance between the two in-
stances ri and qj . It is defined as:

d1 (ri , qj )
d(ri , qj ) = Pk (8)
l=1 d1 (ri , qj )

Closer instances should have greater influence, so the influence of in-

stance rj is exponentially decreased with the distance from the given instance
ri :
2
d1 (ri , qj ) = e−(rank(ri ,qj )/σ) (9)
3.6. A Novel Fast Stability Scanning Framework 53

where rank(ri , qj ) is the rank of the instance qj in a sequence of instances

ordered by the distance from ri and σ is a user defined parameter controlling
the influence of the distance.

3.6 A Novel Fast Stability Scanning Framework

In the original simulation platform [63], stability analysis is performed on
all operating points, which is time consuming. We propose a framework for
fast stability scanning to achieve a significant computational speed-up. The
framework consists of three parts: (i) feature selection, (ii) clustering, and (iii)
stability analysis. The pseudo code of the framework is shown in Algorithm
3.

Definition 1 Let R = {ri | ri ∈ R|A| , i = 1, 2, . . . , |R|} denote a steady-state

power system operating condition, uniquely defined by a set of attributes A = {ai |
r(ai ) ∈ [−1, 1]|R| , i = 1, 2, . . . , |A|}, where r(ai ) is a normalized numerical value of
attribute ai across all operating conditions. For each operating condition ri ∈ R, we
compute a stability index λi ∈ R. The task of fast stability scanning is to cluster
R into a set of representative clusters C represented by cluster centroids c ∈ R|A| , so
that |C| < |R|, and to compute a stability index λ̂ using cluster centroids c ∈ C, so
that |λ − λ̂| ≤  for all r ∈ R, where  is a predefined tolerance.

Algorithm 3 Fast stability scanning framework.

Input: Set of operating conditions R, feature selection performance ρ and
tolerance f , set of features A.
Output: Stability index λ for each for each r ∈ R, minimum cluster distance
c , minimum data distance d .
1: while ρ ≥  do
2: Randomly select a training instance ri ;
3: Run feature selection using RReliefF (Algorithm 2);
4: Update feature weights for all a ∈ A (10);
5: end while
6: Run self-adaptive PSO-k-means clustering (Algorithm 4);
7: for c ← 1, |C| do
8: Calculate λ(c);
9: end for
10: for r ← 1, |R| do
11: Assign λ(c) to r(c);
12: end for

A time-series analysis of one full year with an hourly resolution results

in |R| = 8760. A minimum feasible set A includes voltage magnitudes and
 Chapter 3. Un-supervised machine Learning Method for Fast Stability
54
Scanning

angles at all buses in the system, active and reactive demands, and active and
reactive powers of all generators in the system. Without the loss of general-
ity, however, A can also include derived variables, such as transmission line
flows.
The framework proposed in this study bears similarities and differences
with online DSA. They both involve knowledge base generation and feature
selection. The first difference is in the offline simulation. As a supervised
learning method, DSA requires a big knowledge base to achieve high accu-
racy mapping, which requires a lot of offline simulation. Fast scanning, on
the other hand, is an unsupervised learning method, so the offline simulation
is only needed for the feature selection and to generate the operating condi-
tions for the stability analysis, which has a much lower computational bur-
den. The second difference is in the application. DSA is an operational tool,
which requires fast mapping of current or imminent operating conditions
and a very high accuracy since the mapping result is the basis for preventive
or emergency control. Fast scanning, on the other hand, is a planning tool
that aims to scan a lot of scenarios across long horizons to provide planners
with the stability level of the system under study.

3.6.1 Feature Selection for Weighted Clustering (Lines 1-5 in

Algorithm 3)
Compared to conventional DSA, we propose two innovations in feature
selection: (i) both feature ranks and weights are used in clustering, and (ii)
the size of the required training set for feature selection is determined adap-
tively to reduce the simulation time.
In this study, candidate features considered for clustering include active
and reactive powers of the loads, thirteen synchronous generators including
one CSP, six wind farms, two utility PV farms, HVDC links, and inter-area
active and reactive power flows. Due to the dimensionality of the space of
operating conditions, feature weights have to be considered in clustering.
In [102, 103, 104, 105], feature ranks are used to select a subset of candidate
features used in the classifier that determines the feature weights. In this
study, both feature ranks and weights resulted from ReliefF algorithm are
used for clustering. Feature weights also require preprocessing, due to two
reasons: (i) the accumulated effect of many unimportant features may mask
the effect of a smaller number of dominant features, and (ii) to improve the
3.6. A Novel Fast Stability Scanning Framework 55

representativeness of the cluster centroids’, the degree of segmentation for

features with large variance should be increased.
We propose the following weight adjustment for all a ∈ A:

var(r(ai ))
w̃i = C · wi · (10)
log (2 · rank(ai ))
where C is a tunable parameter, and w̃i and wi are adjusted and original
feature weights, respectively.

3.6.2 Self-adaptive PSO-k-means Clustering (Line 6 in Algo-

rithm 3)
The conventional k-means clustering algorithm has two inherent draw-
backs: (1) its clustering performance depends on randomly assigned initial
cluster centroids, which can lead to unreliability; (2) the algorithm is based
on gradient descent and can thus easily terminate in local optima.
In the PSO-k-means algorithm, the solution of the PSO can be used as the
initial k-means cluster centroids, which can avoid the algorithm trapping in
local optima. However, like any other global optimisation algorithm, the PSO
is prone to premature convergence. This may be improved by increasing the
size of the swarm but at the cost of an increased computational burden. An-
other issue is to determine the cluster numbers and how to deal with empty
clusters. To address these issues, we propose a self-adaptive PSO-k-means
clustering algorithm, described in Algorithm 4.
The algorithm starts with the initialization of the PSO particles. Random
cluster centroids (operating points in our case) are assigned as the particles’
initial position C0 , and local best pbest,0 , global best gbest,0 are calculated using
a random initial velocity V0 .
The PSO (Lines 2 to 13) is ran first to locate the best initial position, which
is then used by the k-means clustering in the second stage (Lines 14 to 20).
In the PSO run, the position and the direction of each particle are updated
in every iteration.
The issue with the conventional PSO algorithm is that a particle may fly
out of the load-flow solution space, resulting in a divergent load flow and
hence an infeasible cluster. To overcome this, the nearest feasible position
within the solution space is used instead of the invalid position (Line 6). To
the premature convergence of the conventional PSO algorithm, we adopt a
technique proposed in [124] that monitors the fitness variance of all the par-
ticles in the swarm in each iteration and uses it as an indicator of premature
 Chapter 3. Un-supervised machine Learning Method for Fast Stability
56
Scanning

Algorithm 4 Self-adaptive PSO-k-means clustering.

Input: PSO iteration limit MaxIter
Output: Cluster centroids C.
1: Initialize C0 , V0 , pbest,0 , gbest,0 ;
2: while iteration ≤ MaxIter do
3: for i ← 1, SwarmSize do
4: Update Vi , Ci (5);
5: Update pbest,i , gbest,i if required;
6: Search space limit check;
7: end for
8: Calculate swarm fitness variance (4);
9: Calculate mutation probability pm [124];
10: if pm > rand ∈ [0, 1] then
11: Mutate gbest (9);
12: end if
13: end while
14: The best particle position is used as initial cluster centroids for k-means;
15: repeat
16: Perform k-means clustering;
17: Remove empty clusters;
18: Create new cluster for data points d(r, c(r)) > d ;
19: Combine clusters if d(ci , cj ) < c ;
20: until convergence

convergence. A mutation probability pm is calculated according to [124] and

used as a trigger for a mutation of gbest (Lines 10 to 12). The mutation of gbest
is defined as:
 η
gbest,k = gbest,k · 1 + (11)
2
where η is a normally distributed random variable.

3.6.3 Stability Scanning (Lines 7-11 in Algorithm 3)

Compared with the initial number of operating points, the number of rep-
resentative clusters resulting from clustering is much smaller. The stability
analysis is performed on cluster centroids using conventional stability anal-
ysis. The stability index λ(c) is assigned to every operating point r(c) repre-
sented by the cluster centroid c.
Given |C| < |R|, the computational time is significantly reduced.
3.7. Simulation Results 57

TABLE 3.1: Features and weights for SSS

Initial Initial Adjusted Adjusted

Feature name
weights rank weights rank
Sync11_P 0.106 1 7.709 1
WF04_P 0.098 2 2.267 2
Sync11_Q 0.062 3 1.748 3
PV02_P 0.054 4 1.020 4
PV01_Q 0.041 8 0.538 5
PV01_P 0.040 11 0.465 6
WF04_Q 0.039 12 0.397 7
PV02_Q 0.035 17 0.396 8
Sync09_P 0.048 5 0.310 9
Sync08_Q 0.047 6 0.297 10

3.7 Simulation Results

In order to evaluate the efficacy of the proposed fast stability scanning
framework, we performed small signal stability and steady-state voltage sta-
bility analysis of a simplified model of the NEM in the year 2030 described in
Section III. Fast stability scanning is performed using the representative clus-
ter centroids and the results are compared with the time-consuming time-
series stability analysis, that uses all 8760 operating points.
For small-signal stability, the damping ratio of the inter-area oscillation
mode between Areas 2 and Area 4 is used as the stability index, whereas
for voltage stability, we used the loading margin assuming a uniform load
increase at all load buses in the system, where all generators increase their
production in proportion to the base case. We first present the results of
feature selection and clustering, followed by the results of stability scanning.

3.7.1 Feature Selection

Tables 3.1 and 3.2 show the top 10 features’ initial weights and ranks, and
the adjusted weights and ranks for SSS and VS, respectively.
The results confirm the necessity of feature selection before clustering.
Notice that the feature weights after the feature selection for SSS and VS dif-
fer significantly, which reflects the different features’ impact on SSS and VS.
It is interesting to observe that the generator Sync11 (CSP) and Wind Farm
04, both located in northern QLD, have a significant impact on the oscillation
mode between Areas 2 and 4.
 Chapter 3. Un-supervised machine Learning Method for Fast Stability
58
Scanning

TABLE 3.2: Features and weights for VS

Initial Initial Adjusted Adjusted

Feature name
weights rank weights rank
WF06_P 0.130 1 7.416 1
WF05_P 0.116 3 2.372 2
WF06_Q 0.109 4 2.066 3
Inter-P3 0.128 2 1.973 4
HVDC3S_Q 0.096 5 1.110 5
WF02_P 0.068 6 0.704 6
WF05_Q 0.036 7 0.506 7
WF03_P 0.023 11 0.299 8
WF02_Q 0.027 8 0.250 9
Inter-P2 0.025 10 0.145 10

In order to find the dominant features, the size of the training set is pro-
gressively increased by randomly picking the operating points from the time-
series analysis until the resulted feature ranks and weights converge. Com-
pared to conventional DSA where the size of the training set for feature se-
lection is fixed, our approach avoids unnecessary computation thus reducing
the computational burden, and also prevents overfitting.
Fig. 3.4 shows the convergence process. Observe that a sufficient accu-
racy is achieved after 300 iterations. Note that the stability index need to
be calculated using conventional methods for all operating points used for
feature selection.

3.7.2 Clustering
Self-adaptive PSO-k-means weighted clustering is used to find typical
generation-load patterns. Clustering reduces the number of data points from
8760 operating points resulting from the time-series analysis to 555 and 421
clusters, for SSS and VS, respectively, which represents a dimensionality re-
duction of 95.2% and 93.7%, respectively.
In order to show statistically the advantage of the proposed PSO k-means
algorithm over the conventional k-means, we ran the k-means and the PSO-
k-means using 100 random initial seeds.
Fig. 3.2 compares the best, the average and the worst SMSE of the clus-
tering results of the conventional k-means and the proposed self-adaptive
PSO-k-means. Observe that the k-means algorithm starts from a randomly
assigned cluster centroid that is normally far away from the global optimum.
Therefore, the SMSE of the k-means is much larger than the PSO-k-means
3.7. Simulation Results 59

Sum of mean squared errors K-means(worst)

K-means(average)
0.18 K-means(best)
PSO-K-means(worst)
PSO-K-means(average)
PSO-K-means(best)
0.17

0.16

0.15
1 5 10 15 20
Clustering iteration
F IGURE 3.2: Comparison of the clustering results: conventional
k-means vs. the proposed self-adaptive PSO-k-
means.

SMSE in the first a few iterations. The PSO-k-means, on the other hand, starts
with a much smaller SMSE, and has a better performance overall. The results
also show that the performance of the proposed method is much more stable
with a consistently better performance than the conventional k-means.

3.7.3 Small-signal Stability

For the sake of illustration, a section of the damping ratio of the inter-
area oscillation mode between Areas 2 and 5 between hours 5201 and 5700
is shown in Fig. 3.3 (a), which reveals a close agreement between the fast
scanning results (worst case data) and the time series analysis.
To verify that statistically, the damping ratios were calculated for 500 ran-
domly selected operating conditions and compared with the values obtained
from fast stability scanning.
Fig. 3.3 (b) compares the error distribution of the damping ratio as result
of fast scanning using the conventional k-means (best case data, blue bins)
and the proposed PSO-k-Means algorithm (worst case data, red bins).
Observe that the error the proposed PSO-k-means algorithm is kept be-
low 14%, with the highest density in the 0-4% range, while for the conven-
tional k-means, the error can be as high as 18%. The average percentage error
is 3.2% to 4.5% for PSO-k-means and k-means, respectively.
 Chapter 3. Un-supervised machine Learning Method for Fast Stability
60
Scanning

(a)
Damping ratio 0.14

0.12

0.1

Fast scanning result

0.08 Time series scanning result

5201 5300 5400 5500 5600 5700

Hours in the year 2030
(b)
100
operation point number

k-means algorithm(Best)
80 PSO-k-means algorithm(Worst)

60

40

20

0
0 5 10 15 20
Percentage of damping ratio error

F IGURE 3.3: SSA critical damping ratio fast scanning results:

(a) time series, (b) error distribution using PSO-k-
means and k-means.

3.7.4 Voltage Stability

To illustrate the performance of fast stability scanning for voltage stability
analysis, Fig. 3.5 (a) shows the loading margin between hours 7201 and 7700.
Again, in order to verify the fast scanning accuracy, we calculated the loading
margin for 500 randomly selected operating conditions and compared the
results with the values obtained with fast stability scanning. Fig. 3.5 (b)
compares the error distribution of the load margin using the proposed PSO-
k-means (red bins) and the conventional k-means (blue bins). Observe that
the error is mostly kept below 4%, with the highest density in the 0-3% range
3.7. Simulation Results 61

(a)
0.2 Sync11-P PV02-P WF04-Q Syc08-Q
WF04-P PV01-Q PV02-Q
Feature weights

Sync11-Q PV01-P Sync09-P

0.15

0.1

0.05

0
50 150 250 300 350 400
Number of instances used for feature selection

(b)
0.5
WF06-P HVDC3S-Q WF02-Q
0.4 WF05-P WF02-P Inter-P2
Feature weights

WF06-Q WF05-Q
Inter-P3 WF03-P
0.3

0.2

0.1

0
50 150 200 250 350
Number of instances used for feature selection
F IGURE 3.4: Convergence of feature selection: (a) SSS, (b) VS.

for the PSO-k-means.

Similar to the small-signal stability, the proposed PSO-k-means algorithm
performs much better. In this case, the average percentage error decreases
from 5.5% to 1.2% and the maximum error decreases from 18% to 9% com-
pared to the conventional k-means.
 Chapter 3. Un-supervised machine Learning Method for Fast Stability
62
Scanning

(a)

12000
Load margin (MW)

10000

8000
Fast scanning result
Time series scanning result
6000
7201 7300 7400 7500 7600 7700
Hours in the year 2030
(b)
200
k-means algorithm(Best)
operation point number

PSO-k-means algorithm(Worst)
150

100

50

0
0 5 10 15 20
Percentage of load margin error

F IGURE 3.5: VSA loading margin fast scanning results: (a) time
series, (b) error distribution using PSO-k-means
and k-means.

3.7.5 Worst Case Operating Point Shift

Conventionally in power system planning, worst case conditions are con-
sidered when the system is the most stressed, and stability studies are con-
ducted under these conditions. In order to clearly see the relationship be-
tween the critical damping ratio and the system generation/demand level,
with the constructed inter-area oscillation mode damping ratio trace, the
minimum damping ratio happens at hour 5466 in the year 2030. In Fig. 3.6,
the damping ratio trace between hour 5201 and 5700 is given, total demand
3.7. Simulation Results 63

in NEM of the same time slot is compared with the damping ratio. It can
be observed that the minimum damping ratio does not coincide with the lo-
cal maximum load level, nor the maximum load level in the year 2030, the
observation of the worst case point shifting is in accordance with [101].

35

0.14

Demand active power(GW)

0.13 30
Damping ratio

0.12

0.11 25

0.1

0.09 20

0.08 Damping ratio

Demand active power
0.07 15
5201 5300 5400 5500 5600 5700
Hours in the year 2030

F IGURE 3.6: SSA: Critical mode damping ratio vs. demand.

Similarly, we plotted the loading margin and the total system demand for
a period of 500 hours in Fig. 3.7. Observe that there is little correlation be-
tween high/low demand level and the low/high loading margin, which jus-
tifies the time series approach compared to a conventional approach where
only a small number of the most critical conditions is analyzed.

14 35
Demand active power(GW)
Loading margin (GW)

12 30

10
25

8
20

6 Loading margin
Demand active power
15
7201 7300 7400 7500 7600 7700
Hours in the year 2030
F IGURE 3.7: VSA: Loading margin vs. demand.
 Chapter 3. Un-supervised machine Learning Method for Fast Stability
64
Scanning

3.7.6 Simulation Burden of Stability Scanning

The simulations were performed on a 64-bit Xeon 2.60GHz workstation
with 256GB RAM. Compared to full-time series stability analysis, the compu-
tational burden has been reduced from 220 min to 21 min and from 960 min
to 90 min, for SSS and VS, respectively, which represents about a ten-fold re-
duction with a satisfactory accuracy. It is observed that the feature selection
(30 seconds) and the clustering (5 minutes) computation does not affect the
reduction of computation much.

3.8 Conclusion
Unlike the conventional power system planning that aims to find the op-
timal transmission and/or generation expansion plan, the future grid analy-
sis considers scenarios that are not mere extrapolations of the existing grid.
Next, to capture the intra-seasonal variation in the RES output, we need to
use time series analysis as opposed to picking a small number of the most
critical operating condition, as it is done conventionally. The challenge of fu-
ture grid stability analysis is the sheer number of operating conditions that
need to be analyzed. In this chapter, we have proposed a novel framework
for fast stability scanning of future grids scenarios. The framework is based
on a feature selection algorithm that makes it possible to perform cluster-
ing using both feature ranks and weights. To reduce the number of clusters,
we proposed an improved self-adaptive PSO-k-means clustering technique
that determines the optimal cluster number. The case study demonstrated
the suitability of the proposed framework. Considering the level of detail
required for future grid analysis, an acceptable accuracy is achieved with a
more than a ten-fold speed-up. However, in power system operation su-
pervised machine learning method can achieve higher accuracy for online
stability assessment which is presented in the next chapter.
 65

Chapter 4

Supervised machine Learning

Method for TSA

4.1 Introduction
This chapter is based on my AUPEC paper [125] which focused on solv-
ing reliability of dynamic transient stability assessment due to power system
topology change. Classification using an trained intelligent system ensemble
is core of the task and it is an supervised machine learning method.
Supervised learning can be regarded as of function approximation, where
basically an intelligent system is trained based on input data sets and asso-
ciated responses, and in the end of the process establish function that best
describes the input data sets. In the power system security assessment area,
the data sets are operating conditions and contingencies, the responses are
referred to the system security levels under the given operating conditions
and contingencies.

4.2 Background
Dynamic security assessment provides power system operators with se-
curity information of power systems for current or imminent operating con-
ditions considering various system topologies and contingencies. The se-
curity information is basis for preventive or emergency control to prevent
systems0 insecurity. Transient stability assessment is one of the most impor-
tant tasks in DSA.
Unlike energy function-based methods or time-domain simulation, Intel-
ligent systems trained via machine learning using a large number of training
instances can map operating conditions to power systems’ security status,
which provides a much faster, versatile and easier to implement alternative
66 Chapter 4. Supervised machine Learning Method for TSA

to achieve DSA. In the past 10 years, TSA using IS has attracted a lot of in-
terest. Recent studies [48, 49, 5, 104, 126] demonstrated effectiveness of IS
application in DSA.

4.2.1 Existing Studies

Decision tree is one of the most popular ISs used for TSA [48, 49]. Au-
thors in [48] proposed a path-based DT classification method instead of the
conventional node-based method. In [49], authors used both the classifica-
tion tree and regression tree for transient stability and voltage stability pre-
diction. Support vector machine is another popular tool used by researchers
[50, 51]. In [50], authors used the SVM as IS for TSA and some energy-based
terms as features. In [51], authors proposed a core vector machine for tran-
sient stability assessment and results showed better predictive performance
in case study compared to other vector machine approaches.
Authors in [5, 6, 7] used extreme learning machine algorithm for pre-fault
and post-fault online TSA. DSA application of the ELM algorithm proposed
in [5] has shown to have faster learning speed compared to other ISs. Then a
TSA model using an ELM-based ensemble in [6] is proposed by authors. In
addition to fast and high accuracy of the ELM-based ensemble, the authors
also proposed a decision-making rule, which eventually gives 100% predic-
tive accuracy in their case studies. In study of post-fault TSA, authors in
[7] used an ELM-based ensemble and proposed a new decision-making rule.
The case study demonstrated feasibility of the ELM application for post-fault
TSA, which requires faster learning speed than pre-fault TSA applications.

4.2.2 Motivation Behind the Study

Affecting factors that determine reliability and accuracy of an online IS-
based TSA include Stability Database (SDB) for IS training, IS training al-
gorithms, feature selection algorithms used to select effective feature set as
IS inputs and stability decision-making algorithms. To cover a wide range
of operating conditions, existing studies generate SDB considering different
network topologies and generation-load patterns. In order to maintain high
classification accuracy considering different network topologies, most of ex-
isting studies choose topology-independent variables as candidate features.
However, as we will present in Section 4.6, TSA classification performance
4.3. Contribution of the Study 67

might be unreliable due to power system topology change in some cases. In-
deed, network topology change greatly increases diversity of power system
operating conditions and requires IS has better differentiability.

4.3 Contribution of the Study

In this study, we aim to design a reliable and high accuracy TSA model
considering various system topologies. The contributions of this study are
three-fold: (i) first, from feature selection algorithm aspect, a hybrid filter-
wrapper feature selection method is proposed; (ii) then, from IS training
algorithm aspect, a boosting learning algorithm is used during IS training
process; (iii) last, from stability decision-making algorithm aspect, an ELM-
based ensemble with a new decision making rule based on weighted outputs
of ELMs is proposed to achieve 100% predictive accuracy.
The rest of the chapter is organized as follows: Section 4.4 introduces
the feature selection and IS used in this study; in Section 4.5 we give test
system description and generation of the SDB; in Section 4.6, testing results
of the proposed online transient stability assessment system and analysis of
the results are given; the last section concludes this chapter.

4.4 Extreme Learning Machine Ensemble Learning

and Feature Selection

4.4.1 Extreme Learning Machine Theory

Different feedforward neural networks in conjunction with the Back-Propagation
(BP) learning algorithm are very popular among applications in power sys-
tems including the TSA. However, very slow learning speed of conventional
feed-forward neural network due to BP weights adjustment iteration and
poor generalization issue are bottlenecks in applications. Authors in [127]
proposed an algorithm called Extreme Learning Machine which discarded
gradient descent approach in the BP algorithm.
Cost function E of a Single hidden Layer Feedforward Neural Network
(SLFN) is defined as:
 2
N
X Ñ
X
E=  βi φ(wi · xj + bi ) − tj  , (1)
j=1 i=1
68 Chapter 4. Supervised machine Learning Method for TSA

where N is total instance (a record in an IS training set) number; Ñ is total

hidden layer node number; βi is weight vector of link connecting the ith hid-
den node and output layer nodes; φ() is activation function of the hidden
layer nodes; wi is the weight vector of link connecting input layer and the ith
node in the hidden layer; xj is the j th input vector which has N dimension;
bi is bias of the ith node in the hidden layer and tj is the ith instance target or
class identification.
Authors in [127] proved that minimize the cost function (1) is equivalent
to find specific ŵi , b̂i and β̂(i = 1, · · ·, Ñ ) such that

kH(ŵ1 , · · ·, ŵÑ , b̂1 , · · ·, b̂Ñ )β̂ − T k

(2)
= min kH(w1 , · · ·, wÑ , b1 , · · ·, bÑ )β − T k ,
wi ,bi ,β

where H is a hidden layer output matrix of a SLFN; β is a output weight

matrix, and T is a class identification matrix of a set of instances. Definition
for H, β and T can be found in [127].
According to the ELM theory, if a hidden layer activation function is in-
definitely differentiable then one can assign random values as the hidden
layer input weights and biases directly instead of adjusting the parameters
via gradient descent iteration. Thus a SLFN supervised learning process is
equivalent to calculating minimum-norm least-squares solution β̂ in (3).

β̂ = H † T, (3)

where H † is the Moore-Penrose generalized inverse of matrix H.

The ELM algorithm greatly reduced training time of a SLFN neural net-
work, which is critical for online stability assessment task. Authors have
demonstrated successful application of the ELM algorithm in [5, 104, 7].

4.4.2 Feature Selection for IS

Power system operating condition is defined by a set of system variables,
or features, e.g. generator powers, bus voltages, etc. Feature selection is a
process of selecting a subset of most relevant features that is necessary and
sufficient to describe the target concept.
Feature selection methods can be categorized as embed, filter and wrap-
per approach. The filter method evaluation criteria of individual feature
goodness normally include correlation, distance and entropy, etc. On the
other hand, the wrapper method selects an optimal subset features instead
4.4. Extreme Learning Machine Ensemble Learning and Feature Selection69

of an individual feature depending on accuracy of trained ISs output. Gen-

erally the filter methods lead to a faster learning pipeline but individual fea-
tures’ random combination does not work well as the wrapper methods due
to dependency of features.

4.4.2.1 Filter and Wrapper Feature Selection

As a filter type feature selection method, Relief algorithm estimates fea-

tures’ ability to distinguish different instances (power system operating con-
ditions in our case), represented by features’ weights [120]. Extensions of
original Relief algorithm are Relief-F and RRelief-F can also deal with multi-
class and regression problems, respectively. Results of the algorithm give a
relative weight to each candidate feature and ranks of all features in terms of
their weights.
As a wrapper type method, Sequential Floating Forward Selection (SFFS)
[128] is developed based on Plus-L and Minus-R Selection (LRS) method but
with flexible algorithm parameters. The SFFS method applies a number of
backward steps as long as resulting subset is better than previously evaluated
one after each forward step. However, heuristic methods require assessment
of possible subsets via predictive accuracy and are very time consuming if a
large number of candidate features exist.

4.4.2.2 Hybrid Filter-Wrapper Feature Selection

For power systems DSA application with many candidate features, the
wrapper type methods are not feasible in terms of computation burden. The
time complexity of heuristic feature selection methods are O(N 2 ); where N
is candidate feature number. In this study, we propose a combined filter-
wrapper algorithm using the RRelief-F method to find out top weighted fea-
tures out of all candidate features first, then use the SFFS to locate an optimal
subset for each ELM in an IS ensemble. Due to randomness of individual
ELM’s parameters, the feature subsets for ELMs are different. Algorithm of
the proposed method is in Algorithm 5.

Definition 2 Let R = {ri | ri ∈ R|A| , i = 1, 2, . . . , |R|} denote a steady-state

power system operating condition, uniquely defined by a set of features A = {ai |
r(ai ) ∈ [−1, 1]|R| , i = 1, 2, . . . , |A|}, where r(ai ) is a normalized numerical value of
feature ai across all operating conditions. For each operating condition ri ∈ R, we
compute a stability index λi ∈ R. The task of feature selection is to find an optimal
feature subset X for each ELM in IS ensemble.
70 Chapter 4. Supervised machine Learning Method for TSA

Algorithm 5 Combined feature selection algorithm

Input: Training instance r ∈ R and corresponding stability index λ.
Output: For each ELM, a set of features Xk ∈ A.
1: Set all w to 0;
2: for i ← 1, m do
3: Randomly select instance ri ;
4: Select k instances qj nearest to ri ;
5: for j ← 1, k do
6: ndc ← ndc + diff (λ(·), ri , qj ) · d(ri , qj )
7: for l ← 1, |A| do
8: nda da
l ← nl + diff (l, ri , qj ) · d(ri , qj )
9: nl ← ndca +
dca

10: diff (τ (·), ri , qj ) · diff (l, ri , qj ) · d(ri , qj )

11: end for
12: end for
13: end for
14: for l ← 1, a do
15: wl ← ndca /ndc − (nda − ndca )/(m − ndc )
16: end for
17: Atop ⊂ A, Xk = ∅, k = 0;
18: while σ ≤ acc do
19: Yk ← Atop − Xk
20: fms ← argmaxy∈Yk [J(Xk + y)]
21: Xk ← Xk + fms
22: fls ← argmaxx∈Xk [J(Xk − x)]
23: if J(Xk − fls ) > J(Xk ) then Xk+1 ← Xk − fls
24: go to 22
25: else
26: go to 19
27: end if
28:
end while
4.4. Extreme Learning Machine Ensemble Learning and Feature Selection71

Lines 1-12 in Algorithm 5 perform RRelief-F algorithm to find feature

weight w for each candidate feature in A. Line 13 chooses top ranked features
Atop for SFFS algorithm based on feature weights. In SFFS loop (Lines 14-22),
prediction accuracy σ is checked with an user defined threshold acc. Lines
15-16 find most significant feature fms within candidate features Yk . Lines
17-18 performs backwards evaluation to remove least significant feature fls .
J(X ) performs prediction using feature subset X and calculate accuracy.

4.4.3 Ensemble Learning and Rule for Classification

In machine learning, an IS-based classifier has categorical output while
a predictor has continuous output. We use a set of ELM-based predictors
to construct an ensemble and the ensemble works as a classifier to evaluate
power system OCs stability status. Each predictor gives outputs when sub-
ject to system conditions. The predictors in the ensemble must be able to
reflect system properties from different aspects by ensuring diversity of the
predictors. Therefore, the predictors in the ensemble normally have differ-
ent outputs under same system condition and final output of the ensemble is
based on all the predictors’ outputs.
Ensemble learning is an effective method to significantly improve gener-
alization of a neural network [129]. In this study, an IS ensemble has a group
of ELMs, and each ELM has different parameters and input set X which are
resulted from previous feature selection. For operating conditions applied to
the ensemble, the ELMs have different outputs. Ensemble learning combines
outputs of the ELMs to achieve better predictive performance comparing to
a single ELM. Each ELM in an IS ensemble is called a weak predictor. In
this study, we use average value of all weak predictor’s outputs as gauge for
operating condition stability classification.
We define transient stability status of any operating conditions as a binary
class and label stable OCs with +1 and unstable OCs with −1. Therefore,
when use the ensemble output for operating condition hard classification,
threshold between stable and unstable is 0. However, to avoid misclassifica-
tion we regard near zero ensemble outputs as incredible; and for operating
conditions lead to incredible ensemble outputs time-domain simulation is re-
quired to determine its real stability status. In this way, overall TSA reliability
and accuracy can be maintained at a high level for online implementation.
72 Chapter 4. Supervised machine Learning Method for TSA

In [130] a boosting algorithm is used to train weak predictors. In boost-

ing training algorithm, likelihood of an operating condition is used for train-
ing depends on its sampling probability. To train a set of weak predictors
one after another, the boosting algorithm increases the sampling probability
of those operating conditions which are misclassified by previous predictor
when train the next predictor hence increase accuracy for those operating
conditions. Other than using the boosting algorithm, we propose to give a
weight for each weak predictor’s output in calculating the ensemble output
for decision making. By doing so, incredible OCs number is reduced and
less time-domain simulation is required which is proved in Section 4.6. The
proposed algorithm is given in Algorithm 6.

Algorithm 6 ELM ensemble classification process.

Input: Training set RT R = (ri , λi ), Testing set RT E = (rk , λk ), λi ∈ λT R , λk ∈ λT E
Output: Classification results of testing set, λresult
1: for m ← 1, M do
2: p(1) (m) = 1/M
3: end for
4: for j ← 1, N do
5: hj ← ELMt (pj , RT R )
6: λj ← hj (RT R )
7: Iid ← Find(λj 6= λT R )
8: p(j+1) (Iid ) = Adj. ∗ pj (Iid )
9: λj ← hj (RT E )
10: end for
P
11: λ = j=1,N (λj · wj )/N
12: for s ← 1, M do
13: if λ(s) ≤ LowLimit then
14: C(s) ← Unstable
15: else if λ(s) ≥ HighLimit then
16: C(s) ← Stable
17: else
18: C(s) ← Simulation(RT E (s))
19: end if
20: end for
21: λresult ← C

Operating condition set R and corresponding stability indices λ is di-

vided into training set RT R and testing set RT E . In the algorithm, Lines 1-2
initialize sampling probability for each operating condition in RT R . Lines
3-8 achieve boosting training for N weak predictors. The weak predictors
are trained one after another using ELM method ELMt and training set RT R
based on sampling probability p. Lines 5-7 evaluate the trained predictor
and locate misclassified operating conditions then adjust corresponding sam-
pling probability. Line 9 calculate the ensemble output by averaging all
4.5. Simulation Platform and Stability Database 73

predictors’ outputs considering user defined weights w. Lines 10-17 per-

form proposed classification rule. User defined parameters LowLimit and
HighLimit separate credible and incredible ensemble outputs and conduct
time-domain simulation for incredible operating conditions.

4.5 Simulation Platform and Stability Database

4.5.1 The 39-bus Test System

The IEEE 39 bus New England System is a well-known benchmark power
system. Researchers have been using the system for variety of studies includ-
ing online TSA [104, 126]. The IEEE 39-bus system has 10 generators, 39 buses
and 46 lines which consolidate a typical meshed power system. In this study,
all network parameters and dynamic controller parameters are from [131].

4.5.2 Stability Database

For an intelligent system based online TSA, an accurate mapping from a
given operating condition to corresponding system stability status relies on
an unbiased, realistic and comprehensive stability database which is used to
train the IS. Almost all existing studies emphasized importance of generation
of the database [5, 48, 126]. In this chapter, we try to study how power sys-
tem topology change would impact on TSA classification performance and
consequently design a reliable online TSA system.
Other than a base case topology with all components in the test system
in service, we consider another 10 topologies by taking different components
out of service. Table 4.1 summarizes the 11 topologies we used for SDB gen-
eration.
TABLE 4.1: Test System Topologies

Topology Out serviced Topology Out serviced

Base case — 6 Generator 10
1 Lines 1 - 39 7 Generator 9
2 Lines 3 - 4 8 Generator 6
3 Lines 16 - 17 9 Generator 5
4 Lines 14 - 15 10 Generator 3
5 Lines 21 - 22

For operating condition design, we considered 8 basic loading varying

between 50% to 120% with a 10% interval using default values (base case)
74 Chapter 4. Supervised machine Learning Method for TSA

from [131]. Generators proportionally supply loads based on the base case
ratio. Taking into account of uncertainties in power system, around each
basic loading, 20 more loading randomly generated within ±20% of the basic
loading. The random loading variance is satisfied by distributing random
increments to the generators. Thus in total we generated 160 generation-load
patterns. We then conducted power flow study by applying the generation-
load patterns to the 11 topologies and recorded 1760 operating conditions.
Next step is to perform time-domain simulation for each OC and record
stability status(stable or unstable) when the system is subject to contingen-
cies. The contingencies include transient three phase short circuit events on
all 39 buses; fault duration is randomly generated for a specific contingency
and between 100ms to 200ms. An operating condition and corresponding
stability status form one sample in the SDB. In total, the stability database
includes 1760*39 = 68640 instances. Instances resulted from the 11 topologies
are stored in different SDBs.
In reality, for most contingencies and OCs, a power system is normally
stable. Therefore, the number of stable OCs in a stability database is nor-
mally much bigger than the number of unstable OCs. To handle the class
imbalance issues with the original SDB, re-sampling method presented in
[132] is adopted.

4.6 Results and Analysis

In this section, we give results of feature selection, TSA classification accu-
racy and analyze how different methods lead to different classification accu-
racy. In the following case studies, one online IS ensemble-based TSA scheme
is created with 30 ELMs and each ELM has 30 inputs describing an OC and
one output giving corresponding transient stability status.

4.6.1 Results of Feature Selection

Candidate features for TSA normally are classified into two categories.
One is pre-contingency steady-state variables such as bus voltage magni-
tudes, bus voltage angles, generator active powers and reactive powers, etc.
Another is post-contingency variables such as rotor angles and speed tra-
jectories, bus voltage magnitude trajectories, etc. Researchers normally use
the second category variables for post-contingency stability assessment for
emergency control purpose [7]. In this chapter, we focus on pre-contingency
4.6. Results and Analysis 75

transient stability assessment and use the first category variables as candi-
date features. More specifically, we included bus variables, generator rotor
angles, active and reactive powers, load active and reactive powers. Con-
ventionally these variables are considered as topology independent variables
which are expected to better reflect topology change and give accurate clas-
sification result when evaluate OCs of a range of topologies [5, 104, 126, 49].

0.05

0.04
Feature weights

0.03

0.02

0.01

-0.01

-0.02
25 50 75 100 125 150 183
Rank of candidate features
F IGURE 4.1: Ranked weights of candidate features as result of
the RRelief-F feature selection.

Fig. 4.1 gives results of the RRelief-F as a filter type feature selection
method. Feature weight is a gauge of correlating level between individual
feature and stability status of the test system. The figure shows that majority
of the candidate features have a little or negative correlation with the system
stability status compared to the top 50 features.
We then use the top ranked 50 features as candidate features for the SFFS
algorithm which aims to find 30 optimal feature subsets for ELMs in the en-
semble as their inputs. Using 50 features instead of 183 features, the SFFS
forward-backward searching process time complexity reduced from O(1832 )
to O(502 ) which is more than 10 times reduction. We evaluate one ELM clas-
sifier using different feature sets resulted in the SFFS feature subset selection
76 Chapter 4. Supervised machine Learning Method for TSA

95
Classification accuracy (%)

94

93

92

91

90
5 10 15 20 25 28
SFFS iteration
F IGURE 4.2: Classification accuracy improvement during the
SFFS process.

process. Fig. 4.2 shows classification accuracy improves when feature set is
refined in the SFFS process.

4.6.2 Evaluation with Different System Topologies and Con-

tingencies
To show how system topology change impact on TSA classification per-
formance, we first use samples in the 11 SDBs separately to train and test the
ensemble, without the boosting algorithm in the process. For each ELM, 30
out of the 50 top weighted features are randomly chosen as inputs. For each
contingency on 1 of 39 buses, 10-fold cross validation [133] is carried out,
and average testing classification accuracy is calculated. The average testing
accuracy is calculated by,
P11
t=1 T S c (t)
T S(c) = , (4)
11
4.6. Results and Analysis 77

where T S c (t) represents classification accuracy of the TSA which is trained

and tested by samples in the tth topology0 SDB resulted by the cth contin-
gency.

100
Average classification accuracy ( % )

99

98

97

96

95
Predictor trained by single topology instances
Predictor trained by combined topologies instances
94
10 20 30 39
Fault location
F IGURE 4.3: Classification accuracy affected by topology
change.

Next, the TSA is trained and tested by using mixed samples of all topolo-
gies for each contingency. The bottom curve in Fig. 4.3 indicates that clas-
sification accuracy compared to the single topology case decreases by 1% to
3% for most of the contingencies. The result shows that TSA performance
deteriorated due to topology change. In other words, if network topology
changes, the TSA classification performance is not reliable despite topology
independent features are used.

4.6.3 Evaluation with Hybrid Feature Selection and Boosting

Algorithm
To overcome TSA performance deterioration due to topology change, We
first try to improve classification accuracy by using optimal feature subsets
as results from the hybrid RRelief-F + SFFS feature selection. Different to
78 Chapter 4. Supervised machine Learning Method for TSA

randomly picked features, feature subset for each ELM have shown a bet-
ter differentiability in the previous section. Compared to previous case, the
optimal feature subsets help to improve classification accuracy by 1% in Fig.
4.4.
To further improve TSA performance and take advantage of ensemble
learning, we then use the boosting technique, which enhances learning on
misclassified samples in training process. Fig. 4.4 shows classification accu-
racy of the ELMs ensemble increases to over 99% for all contingencies, which
is even better than the single topology case without boosting learning.

100

99
Classification accuracy ( % )

98

97

96

95 Predictor using features selected by RRelief-F

Predictor using features selected by hybrid selection
Using boosting algorithm and hybrid selection

94
10 20 30 39
Fault location
F IGURE 4.4: Classification accuracy improved by SFFS and
boosting learning.

4.6.4 Decision-making for 100% Predictive Accuracy

By using combined feature selection and boosting learning, the resulted
classification accuracy is over 99%, but there is a way to improve the classi-
fication accuracy further to 100% at the cost of time-domain simulation for
all incredible operating conditions. We investigated all misclassified OCs
resulting from the 10-fold cross-validation for the 39 bus contingencies and
 4.6. Results and Analysis 79

classification results of these OCs are depicted in Fig. 4.5. All these results
are well within a boundary between [−0.15, +0.15].

0.2

0.15
Classification results

0.1

0.05

-0.05

-0.1

-0.15

-0.2
1 10 20 30 40 47
Misclassified operating points for all bus contingencies
F IGURE 4.5: Prediction results of all misclassified OCs for faults
on the buses.

Therefore, we change threshold of classification rule from a single bound-

ary 0 to [−2, −0.2] for unstable and [+0.2, +2] for stable OCs. All the OCs
with prediction results located in this incredible zone [−0.2, +0.2] will have
true stability status determined by time-domain simulation.
Obviously, the number of OCs within the incredible zone is more than the
misclassified OC number. It is still time consuming to do time-domain sim-
ulation for all these OCs in some cases. By using weighted weak predictors
outputs in the proposed Algorithm 6, we reduced number of incredible OCs
and hence limited requirement of the time-domain simulation. Fig. 4.6 com-
pares 100 OCs with IS prediction results between [−0.3, +0.3] without (circles
◦) and with (diamonds ) weighted outputs considered. It is clear that sepa-
ration between stable and unstable OCs are increased and less OCs are in the
incredible zone. Indeed, the average number of incredible OCs reduced from
98 to 22 for each contingency. Although we only give results of contingencies
80 Chapter 4. Supervised machine Learning Method for TSA

F IGURE 4.6: Ensemble classification using weighted outputs

improves class separation.

on the buses in the chapter due to space limit, results for contingencies on
the lines are similar.

4.7 Conclusion
In this study, for intelligent system based transient stability assessment,
simulation results show that using system topology independent features
as IS inputs in some cases is unreliable when system is subjected to topol-
ogy change. In order to overcome classification unreliability due to topology
change, we first proposed a hybrid filter-wrapper feature selection method,
which helped to improve TSA classification performance when the test sys-
tem is evaluated with multiple topologies. Furthermore, simulation results
show that the boosting learning method in conjunction with the rule for clas-
sification we proposed based on weighted weak classifiers’ output can signif-
icantly improve TSA reliability and helped to reduce time-domain simulation
for achieving 100% classification accuracy.
4.7. Conclusion 81

In a process to prepare the SDB, we run time domain stability study to

find stability status (label) for each OC. This step is called OC labeling. Sim-
ilar to other published papers on intelligent system based DSA, supervised
learning (only use labeled OC) is used to train the intelligent system which
requires a large number of time domain simulation to prepare a SDB. A reli-
able online DSA system needs to be updated to accommodate fast changing
power system operating conditions. Therefore, preparing a large SDB for
timely DSA system updating becomes prohibitive. Next chapter is focusing
on using semi-supervised learning (use both labeled and unlabeled OC) in
online DSA framework which requires much less time consuming simula-
tion for preparing a SDB.
 83

Chapter 5

Semi-supervised machine Learning

Method for TSA

5.1 Introduction
This chapter is based on the second journal paper [134] which focused on
reducing labeled samples required for intelligent system training and main-
tain high accuracy of transient stability classification at the same time. Clas-
sification using an trained intelligent system ensemble is core of the task and
semi-supervised machine learning method is used.
In machine learning, in between unsupervised learning and supervised
learning is the semi-supervised learning algorithm. In many practical situa-
tions, the cost to get responses for given data sets or labeling is quite high,
since it is often time consuming to do that. So, in the absence of responses in
the majority of the data sets but present in few, semi-supervised algorithms
are the best candidates for the model building. In the power system security
assessment area, the data sets are operating conditions and contingencies, the
responses are referred to the system security levels under the given operating
conditions and contingencies.

5.2 Background
Historically, power system security assessment was based on offline time-
domain simulations. With an increasing penetration of large-scale renew-
ables, however, such as wind and solar generation, as well as distributed
energy sources, such as rooftop PV, battery storage and flexible loads, power
systems are becoming less predictable. Using conventional offline security
assessment to achieve the same level of accuracy and reliability would sig-
nificantly increase the computational burden due to an increased number of
84 Chapter 5. Semi-supervised machine Learning Method for TSA

possible operating conditions. Therefore, dynamic security assessment has

emerged to properly capture the security of fast changing operating condi-
tions. To reduce the computational burden associated with time domain sim-
ulations or energy function methods used for security assessment, machine
learning is typically used to train a classifier used to classify power system
operating conditions, which can be done in real time.

5.2.1 Existing Studies

Transient stability assessment is one of the most important tasks of DSA.
Several recent studies [49, 50, 108, 51, 6] have demonstrated the effectiveness
of different machine learning techniques for DSA. For example, in [49] de-
cision trees are used for transient and voltage stability assessment; in [50],
a support vector machine-based classifier is used for TSA; radial basis func-
tion neural network is used in [108] to assess voltage stability; in [51], a core
vector machine is proposed for TSA; and a computationally efficient DSA
framework based on extreme learning machines is proposed in [6].
A key feature of a classifier is its generalization ability, which refers to
the ability of the classifier to give reliable and accurate predictions using pre-
viously unseen data. The generalization ability depends on the classifier’s
structure, the learning algorithm used, the training set size and its quality
[9]. Most existing DSA studies focus on designing supervised learning al-
gorithms to achieve more accurate and faster DSA. In supervised learning,
a training set consists of a group of variables describing system conditions
and the corresponding labels (security indexes in DSA). The labels are nor-
mally obtained using computationally expensive time-domain simulations.
In [49], for example, 476 data points with a 3-minute resolution capturing
day-ahead operating conditions are evaluated considering 181(N − k) con-
tingencies, where k ∈ {2, 3, 4}. In order to prepare the training set for voltage
and transient stability assessment using decision trees, 476×181×2 = 172, 312
time-domain simulations are therefore required. In [6], 6, 345 operating con-
ditions are evaluated and 6, 345 × 100 = 6, 345, 000 time-domain simulations
are required for TSA considering 100 contingencies.

5.2.2 Motivation Behind the Study

To accommodate fast changing power system conditions, DSA classifier
must be updated regularly to ensure its robustness, which requires generat-
ing new training samples and retraining. If retraining is done online, say,
5.3. Contribution of the Study 85

using short-term forecasting of operating conditions, the update becomes

meaningless if retraining is too slow.
Clearly, evaluating a large amount of operating conditions in order to
cover a wide range of diverse operating conditions in a training set quickly
becomes computationally prohibitive. Reducing the number of samples in
the training set is also not an option as this would reduce the classifier’s gen-
eralization ability [9].
A possible solution is to use parallel or distributed computing techniques,
as suggested in [6, 135]. However, that requires either spatial or time parti-
tioning of the problem, neither of which is trivial. It also significantly in-
creases the hardware requirements.

5.3 Contribution of the Study

Another alternative is to use semi-supervised learning [10], which uses both
labeled and unlabeled samples. In this study, we propose a new DSA frame-
work based on a combination of semi-supervised learning and data editing
to reduce the data noise in the learning process. To improve the general-
ization ability of the classifier, we use a large number of unlabeled operating
conditions, which can be computed efficiently. As a result, the proposed DSA
framework requires significantly less labeled operating conditions to achieve
a similar generalization ability.
We use an ELM based classifier, however the proposed framework can
easily accommodate other machine learning techniques. To the best of our
knowledge, this is the first application of semi-supervised learning for DSA,
which is the main contribution of this study.
The rest of the chapter is organized as follows: Section 5.4 reviews the
machine learning techniques used in the chapter; Section 5.5 details the pro-
posed DSA framework; the test system and the training data used in the case
study are given in Section 5.6; Section 5.7 presents the results of the case
study; and Section 5.8 concludes.
86 Chapter 5. Semi-supervised machine Learning Method for TSA

5.4 Review of Pertinent Machine Learning Tech-

niques

5.4.1 Training Set Size and Generalization performance

According to the statistical learning theory, the generalization performance
of a classifier is affected by the training set size and its quality, and the learn-
ing algorithm used [9]. For example, the error between the estimated net-
work and the target function f for a common class of artificial neural net-
works is shown to be bounded by [136]:

Cf2
   
nd
O +O ln N , (1)
n N

where n is the number of neural network nodes, and d is the input dimension
of the neural network, N is the number of training observations, and Cf2 is
the first absolute moment of the Fourier magnitude distribution of f . O(·)
describes the limiting behavior of a function in term of its arguments. Simi-
larly, the bound on the generalization error for a radial basis function neural
network is [137]:
  r !
1 nd ln (nN ) − ln δ
O +O , (2)
n n

where 0 < δ < 1.

It is apparent from (1) and (2) that the bigger the training set size, the
better the generalization performance of a classifier, as confirmed by the sim-
ulation results in Section 5.7. A possible way to reduce the size of the training
set is to use semi-supervised learning, using unlabeled samples. However,
this introduces data noise into the training set, which can deteriorate its qual-
ity. The next two sections discuss semi-supervised learning and data editing,
which can be used to deal with erroneously labeled data points in the train-
ing set.

5.4.2 Supervised, Unsupervised and Semi-supervised Learn-

ing
x
Let X = {xi |xi ∈ R|A | , i = 1, 2, . . . , |X |} define a set of steady-state
power system operating conditions, uniquely defined by a set of attributes
5.4. Review of Pertinent Machine Learning Techniques 87

Ax = {axi |axi ∈ R, i = 1, 2, . . . , |Ax |}. Normally, xi is assumed to be inde-

pendently and identically distributed in X . For each operating condition
xi ∈ X , one can compute (normally by time-domain simulation) a system re-
y y y
sponse yi ∈ Y, where Y = {yi |yi ∈ R|A | , i = 1, 2, . . . , |X |}, and Ay = {ai |ai ∈
R, i = 1, 2, . . . , |Ay |} is the set of attributes describing system response yi ∈ Y.


Let L = (x1 , y1 ) , . . . , (xi , yi ) , . . . , x|X | , y|X | define a labeled training set,
where a 2-tuple (ordered pair) (xi , yi ) ∈ L is called a labeled sample. Differ-
ent to time-domain simulation method, a machine learning based classifier
h is used to evaluate any unlabeled sample x by assigning it a class c ∈ C,
|C| ∈ Z+ . In the context of DSA, the aim of supervised learning is to use the
classifier h which is trained by a large labeled training set L to classify (la-
bel) unseen power system operating conditions (not in L), where h(x) ∈ R
and C = {0, 1}. A classification rule CR() is used to assign a class c to op-
erating condition x, c = CR(h(x)). Assumed in our study, h(x) > 0.9 and
h(x) < −0.9 assume, respectively, a stable (c = 1) and an unstable (c = 0)
operating condition, while the values in between are undecided. In unsuper-
vised learning, on the other hand, we use an unlabeled training set U = X to
estimate the density of distribution X . Supervised learning has been widely
used in DSA [49, 50, 108, 6, 51].
In many applications generating labeled samples is computationally ex-
pensive. In TSA, for example, it requires running time-domain simulations
for a large number of operating conditions, obtained using computationally
efficient power flow analysis. To improve the generalization performance of
a classifier we can use semi-supervised learning using both labeled samples
L and unlabeled samples U. Typically, the number of unlabeled samples is
much bigger than the number of the labeled samples, |U|  |L|.
Semi-supervised learning is based on cluster and manifold assumptions
[10]. The main idea of the cluster assumption is that there is a high prob-
ability that all samples in the same cluster have the same system response.
According to the manifold assumption, samples within a small neighbor-
hood in the space of power system operating conditions X have a similar
system response. Under these assumptions, unlabeled samples U are used to
increase the density of the training set space, which leads to a more accurate
assessment.
88 Chapter 5. Semi-supervised machine Learning Method for TSA

5.4.3 Tri-training
One of the most popular semi-supervised learning algorithms is co-training
[138], which trains two classifiers separately on two independent sets of at-
tributes of the training set. The issue with co-training is that it requires time
consuming cross validation [133]. To overcome that, [139] proposed a new
semi-supervised learning algorithm called tri-training, using three classifiers
instead of two. After the three classifiers are trained using a small set of la-
beled samples, a large number of unlabeled samples is evaluated by the three
classifiers. An unlabeled sample can be added to the training set used to train
one classifier only if the other two classifiers agree on its classification. Unlike
in co-training, the labeling confidence doesn’t need to be explicitly measured.
The issue is that one classifier may get a sample with a wrong label if the
other two classifiers both give wrong classification results. Samples with a
wrong label are called noise, with the noise rate of a training database defined
as the ratio of the number of noisy samples to the total number of samples.
According to [140], even in the worse case, if a sequence of m labeled samples
is drawn, classification noise rate increase due to wrongly labeled samples
can be compensated if the amount of samples m satisfies:
 
2 2N
m≥ ln , (3)
ε (1 − 2η)2
2 δ

where ε is the hypothesis’s (in our case the classifier) worst-case classification
error rate, η < 0.5 is an upper bound on the noise rate, N is the number
of hypotheses, and δ is the hypothesis’ confidence limit. A hypothesis Hi
that minimizes the disagreement with the sequence will have the Probably
Approximately Correct learning (PAC learning) property [140]:

Pr [d(Hi , H ∗ ) ≥ ε] ≤ δ, (4)

where probability Pr [d(·)] is taken over all evaluation runs of the symmetric
difference between the two hypothesis sets Hi and H ∗ (the ground-truth). In
other words, (4) tells us that the probability of hypothesis Hi being within 
from H ∗ is at least 1 − δ.
Based on the PAC learning theory [140] and the co-training algorithm
proposed in [138], we summarize the tri-training criteria derivation process
proposed in [139]. Let c = 2µln(2N/δ). Further, introducing µ to make the
5.4. Review of Pertinent Machine Learning Techniques 89

equality in (3) hold we get:

c
m= . (5)
ε2 (1 − 2η)2

Rearranging (5) and introducing v as a gage of the classification error de-

pending on m and η we have:

c
v= = m(1 − 2η)2 . (6)
ε2

It follows from (6) that the larger the sample set m and the smaller the sample
noise rate η, the larger the v and the lower the classification error ε.
In the t-th iteration of the tri-training, sample noise rate η (t) is defined as:

ηL |L| + e(t) L(t)

η (t) = , (7)
|L ∪ L(t) |

where L denotes the initial labeled sample set with size |L|, ηL is the labeled
sample set noise, L(t) denotes the pseudo labeled sample set, and e(t) denotes
the upper bound of the classification error rate in the t-th iteration. Substi-
tuting (7) into (6) for the t-th and the (t − 1)-th iterations gives:
!2
ηL |L| + e(t) L(t)
u(t) = L ∪ L(t) 1−2 (8)
|L ∪ L(t) |

and !2
ηL |L| + e(t−1) L(t−1)
u(t−1) = L ∪ L(t−1) 1−2 . (9)
|L ∪ L(t−1) |


In order to have a decreasing classification error rate e(t) < e(t−1) , u(t) >
u(t−1) needs to be satisfied.
In (8) and (9) we made a few assumptions. First, ηL is very small since in
a labeled sample set there is very little noise. Second, L(t) is always bigger
than L(t−1) . This can be guaranteed by choosing an increasing number of
unlabeled samples in iterations. The last assumption is that e(t) ≥ 0 and
e(t−1) < 0.5. Therefore, if e(t) L(t) < e(t−1) L(t−1) then u(t) > u(t−1) .
To determine if an unlabeled sample could be labeled for a classifier we
require:
e(t) L(t−1)
0 < (t−1) < < 1. (10)
e |L(t) |
When L(t) is much bigger than L(t−1) then e(t) L(t) may be larger than
e(t−1) L(t−1) . To avoid the situation from happening, a sub-sampling step
90 Chapter 5. Semi-supervised machine Learning Method for TSA

is used to reduce L(t) by using only a subset of L(t) and therefore ensure
e(t) L(t) < e(t−1) L(t−1) . The size of L(t) after the sub-sampling is given as:
& '
e(t−1) L(t−1)
L(t) = −1 . (11)
e(t)

5.4.4 Data Noise and Data Editing

According to the PAC learning theory, noisy samples can be used in ma-
chine learning, which is the basis for most semi-supervised algorithms [140].
A drawback of semi-supervised learning, however, is inevitable noise that
comes with unlabeled samples [141], which has a negative impact on the tar-
get classifier. The tri-training algorithm uses two classifiers with an accept-
able confidence to label the unlabeled samples in order to enrich the training
set for the other classifier. The number of erroneously labeled samples in
iterations increases since classifiers are trained by increasingly less accurate
samples that only include a limited number of initially labeled samples. To
address this, data editing can be used to reduce the noise introduced in the
training [139]. The aim of data editing is to improve the quality of the train-
ing set by identifying and eliminating mislabeled samples prior to training
the classifier [142]. The idea can also be used in semi-supervised algorithms
to reduce the noise introduced in the pseudo labeling process.
In our approach, we use a data editing technique called depuration [142],
in which the algorithm changes labels of erroneously labeled samples or even
removes “suspicious” samples from the training set based on a well-known
K-Nearest Neighbor (KNN) algorithm. In the KNN algorithm, a data imper-
fection is determined by using labels of k nearest neighbors of a sample xi . If
more than k 0 , (k + 1)/2 ≤ k 0 < k, neighbors of xi have the same label c, the
label of sample xi is changed to c. The algorithm is summarized in Algorithm
1. To find k nearest neighbors for x, the Euclidean distance between x and
the other samples in the training set are calculated. In Cartesian coordinates,
the Euclidean distance between two points xi and xj is defined as:
v
u n
uX
d(xi , xj ) = t wd (xid − xjd )2 , (12)
d=1

where wd are feature weights, and n denotes the dimensionality of the feature
vector.
5.5. A New Dynamic Security Assessment Framework 91

Algorithm 7 Depuration Data Editing.

Input: Initial labeled training set L; Parameters k and k 0 .
Output: Edited training set S.
1: S ← X
2: for all s ∈ S do
3: for all c ∈ C do
4: N ← k nearest samples in {X \ s}
5: nc ← Number of samples in N in class c
6: if nc > k 0 then
7: Change class c(s) to c
8: else if S ← {S \ s} then
9: end if
10: end for
11: end for

In this chapter, we combine tri-training and data depuration to achieve a

reliable and more accurate DSA with less labeled training data compared to
conventional supervised algorithms.

5.5 A New Dynamic Security Assessment Frame-

work
A schematic diagram of the proposed DSA framework is shown in Fig.
5.1. It consists of an offline training module and an online application mod-
ule. Online classifier retraining is also achievable but will not be discussed in
this chapter.
The core difference of the proposed framework compared to the existing
ones [49, 50, 108, 6, 51] is the training process. In conventional approaches,
only labeled samples are used to train the classifier. In the proposed frame-
work, also unlabeled samples are used to enrich the training set and data
editing is used to reduce the noise introduced in the training process.

5.5.1 Training Set Generation

The training set used to train a classifier needs to be unbiased, realistic
and should cover a wide range of possible operating conditions. In our study,
we follow the approach proposed in [48], which is to forecast load and gener-
ation data to generate the training set for security assessment. Generator dis-
patch levels are obtained using a market model based on an optimal power
flow.
92 Chapter 5. Semi-supervised machine Learning Method for TSA

Prepare training set

Forecasting Market Simulation

Security Archive Security Simulation

Unlabeled OPs Labeled OPs

IS training

Data Semi-supervised
Updating
Editing Learning Algorithm

IS parameters
Online application
Trained Online
Real-time
Security
Measurements
Assessment IS

Online or Offline Decision Making

F IGURE 5.1: Block diagram of the proposed DSA framework.

5.5.2 Extreme Learning Machine Classifier

Various machine learning techniques have been used to train the classi-
fier for DSA, including artificial neural networks, support vector machines,
5.5. A New Dynamic Security Assessment Framework 93

and decision trees [49, 50, 108, 6, 51]. They generally exhibit good perfor-
mance, however they tend to suffer from an excessive training time and a
complex parameters tuning. A comparison of different classifiers is given
in [5], showing that ELM have a superior computational performance and
a competitively high accuracy. This can be attributed to the random assign-
ment of weights and biases in the hidden layer, which obviates parameter
tuning using gradient descent and hence increases the learning speed. ELM
have been successfully applied in several DSA applications [5, 6, 7]. In the
case study we use an ELM-based neural network as a classifier for TSA.

5.5.3 Feature Selection

An operating condition of a power system is defined by a set of system
variables, or features, e.g. generator active and reactive powers, bus voltage
magnitudes and angles, load levels, etc. Feature selection is a process of se-
lecting a subset of relevant features that is necessary and sufficient to describe
the target concept (security status in our case) by reducing the dimensionality
of the input data and enhancing generalization by reducing over-fitting [120].
A popular feature selection algorithm is ReliefF [120], which is also used in
this chapter. The main idea of the ReliefF algorithm is to estimate features’
ability, represented by features’ weights, to distinguish between instances,
power system operating conditions in our case, that are near to each other.
In this chapter, we use ReliefF to select the top ranked features as inputs to
the ELM-based classifier. In addition to that, feature weights are used in the
weighted KNN algorithm to locate the nearest neighbors for data editing.

5.5.4 Combined Tri-training and Data Editing Algorithm

We first explain how we combine the tri-training and data editing algo-
rithms in the proposed DSA framework. Pseudo code of the proposed algo-
rithm is given in Algorithm 8.

5.5.4.1 Initialization (lines 1-4 in Algorithm 8)

Three classifiers (h1 , h2 and h3 ) are trained using the initial labeled train-
ing set using a conventional supervised learning algorithm (ELM in our case).
Diversity of these classifiers is vital to any co-training based on semi-supervised
learning. In the tri-training algorithm, the diversity is obtained by using dif-
ferent training sets for the three classifiers using bootstrap sampling [143].
94 Chapter 5. Semi-supervised machine Learning Method for TSA

Function BootstrapSample() randomly picks different training samples for

the three classifiers and the classifiers are trained using function Learn(). The
classifiers will then be refined in the tri-training process.

5.5.4.2 Accuracy Evaluation (line 9 in Algorithm 8)

In each iteration, the classification error of two classifiers hj , hk ; j, k ∈

{1, 2, 3} is evaluated by function MeasureErr(). The tri-training process con-
tinues for the other classifier hi ; i ∈ {1, 2, 3}; i 6= j, k if the classification error
reduces. Given that the samples are independent and identically distributed,
the classification error is approximated using the labeled samples as follows:

|{x|cj (x) = ck (x) 6= c(x)}|

MeasureErr(hj , hk ) = , (13)
|{x|cj (x) = ck (x)}|

where x ∈ L and L is the original labeled training set; c(x) is the true label
of x obtained by simulation; | · | is cardinality of a set; the numerator is the
number of samples on which hj and hk give the same but wrong classification
result; the denominator is the number of all samples on which hj and hk give
the same classification result.

5.5.4.3 Pseudo Labeling and Data Editing (lines 11-22 in Algorithm 8)

If performance of classifiers hj and hk is improved in the previous itera-

tion, they are used to label all unlabeled samples u ∈ U, where U is the initial
unlabeled training set. Evaluation results of hj and hk on an unlabeled sam-
ple u are regarded as credible if their outputs are either both bigger than 0.9
(stable) or both smaller than −0.9 (unstable). If hj and hk both give the same
credible classification for u then (u, CR(hj (u))) can be put in the training set
Li for the other classifier hi . However, this might introduce noisy samples
into the training set if both hj and hk give credible but incorrect classifica-
tion.
Function WtKNN(u, L) performs a weighted KNN algorithm to find three
nearest neighbors in L for u based on a weighted Euclidean distance. NeighborLabel
is decided by voting on the three nearest neighbors’ labels. The unlabeled
sample u and its pseudo label CR(hj (u)) or CR(hk (u)) is put in Li if the
pseudo label is same as the NeighborLabel. Otherwise, time-domain simu-
lation function SecuritySim() is called to find the true label TrueLabel(u) and
the training sets L and U are updated accordingly.
5.6. Simulation Platform and Stability Database 95

5.5.4.4 Tri-training Criteria (lines 23-33 in Algorithm 8)

The tri-training algorithm achieves a constant classification error reduc-

tion using unlabeled samples in iterations by defining the following crite-
ria: (i) to satisfy (10), the size of the pseudo labeled training set |Li | should
(t) (t)
increase constantly in iterations; (ii) the product ei |Li | should always be
(t−1) (t−1) (t) (t−1) (t−1) (t)
less than ei |Li |. Therefore, if Li is bigger than ei |Li |/ei then
(t)
a function SubSample() is used to reduce the size of Li as defined in (11).
The update flag Updatei is set and iteration continues until the criteria are
not met.

5.5.4.5 Classifier Update (lines 36-41 in Algorithm 8)

For each classifier, if the update flag Updatei is set, then the classifier is
retrained by the learning algorithm Learn() with the initial labeled training
set L and the pseudo labeled training set Li . Unlike the standard tri-training
algorithm, in this framework, newly labeled samples are added during the
data editing process to L and pseudo labeled samples Li are produced in the
data depuration algorithm.

5.6 Simulation Platform and Stability Database

5.6.1 The 39-bus Test System

The IEEE 39-bus New England test system (Figure 5.1) is used to demon-
strate the performance of the proposed framework. We consider two sce-
narios: a conventional one and a renewable one, in which four conventional
generators(on buses B32, B33, B36 and B37) are replaced with wind farms.
The purpose of using the renewable scenario is to demonstrate how an in-
creased diversity of operating conditions due the increased penetration of
intermittent renewable generation affects DSA performance. The network
parameters and dynamic controller parameters are taken from [131].

5.6.2 Training Set

The 39-bus system is divided into six regions based on network coherence
[144]. The load in each region is increased by 50 % compared to the base case
[131], and each region is assigned an hourly load trace from [111] except the
region includes generator G1 and bus B39 which represents an external net-
work. In the renewable scenario, hourly wind traces from [111] are used to
96 Chapter 5. Semi-supervised machine Learning Method for TSA

Algorithm 8 Tri-training & Data Editing Algorithm.

Input: Labeled sample set L; Unlabeled sample set U.
Output: Classifiers hi , i ∈ {1, 2, 3}.
1: for i ← 1 : 3 do
(0)
2: Li ← BootstrapSample(L)
3: hi ← Learn(Li )
4: end for
(t−1) (t−1)
5: t ← 1; ei ← 0.5; Li ← ∅.
6: while Updatei = T RU E do
7: for i ← 1 : 3 do
8: Updatei ← FALSE
(t)
9: ei ← MeasureError( hj , hk ); j, k 6= i; j, k ∈ {1, 2, 3}
(t) (t−1)
10: if ei < ei then
11: for all u ∈ U do
12: NeighborLabel ← WtKNN (u, L)
13: if hj (u), hk (u) ≥ 0.9 ∨ hj (u), hk (u) ≤ −0.9 then
14: if CR(hj (u)) = NeighborLabel then
(t) (t−1)
15: Li ← Li ∪ {(u, CR(hj (u)))}
16: else
17: TrueLabel(u) ← SecuritySim(u)
18: L ← L ∪ {(u, TrueLabel(u))}
19: U ← U \ {u}
20: end if
21: end if
22: end for
(t−1)
23: if |Li | = 0 then
 
(t)
(t−1) ei
24: |Li |← (t−1) (t) +1
ei −ei
25: end if
(t−1) (t)
26: if |Li | < |Li | then
(t) (t) (t−1) (t−1)
27: if ei |Li | < ei |Li | then
28: Updatei ← TRUE
29: else   
(t−1) (t−1)
(t) (t) e |Li |
30: Li ← SubSample Li , i (t) − 1
ei
31: Updatei ← TRUE
32: end if
33: end if
34: end if
35: end for
36: for i ← 1 : 3 do
37: if Updatei = TRUE
 then 
(t)
38: hi ← Learn BootstrapSample(L) ∪ Li
(t−1) (t) (t−1) (t)
39: ei ← ei ; |Li | ← |Li |
40: end if
41: end for
42: t=t+1
43: end while
5.6. Simulation Platform and Stability Database 97

F IGURE 5.2: IEEE New-England 39 Bus System.

model the generation profile of the four wind farms. To increase the num-
ber of operating points, the 168 operating conditions representing one week
are interpolated from hourly to a three-minute resolution, which gives 3360
points. To mimic load uncertainty, the three-minute load levels are modified
by ±20% around the base case. To further increase the diversity of operat-
ing conditions, we assume ten network topologies given in Table 4.1, which
yields 33600 operating points for both the conventional and the renewable
scenario. A market simulation is then used to create dispatch levels for the
conventional generation (wind farms are assumed to have dispatch priority).
Given that the optimal power flow used in market dispatch doesn’t always
converge, the number of operating conditions for time domain simulations is
reduced to 29698 and 30917 for the conventional and the renewable scenario,
respectively. In time-domain simulations, we assume random faults (three-
phase to ground short circuits), varying both the fault location (transmission
line or bus) and the fault duration (between 100 ms to 200 ms). For trans-
mission line faults, the line is disconnected after the fault has been cleared.
98 Chapter 5. Semi-supervised machine Learning Method for TSA

For each time-domain simulation, the transient stability status (stable or un-
stable) is recorded and later used in the classification. In the case study, a
subset of the labeled operating conditions is assigned to a labeled training
set L. Also, a subset of the operating conditions (not in L) without labels is
assigned to an unlabeled training set U.

TABLE 5.1: Test System Topologies

Topology Out of service Topology Out of service

Base case — 5 Lines 21 - 22
1 Lines 1 - 39 6 Generator 10
2 Lines 3 - 4 7 Generator 9
3 Lines 16 - 17 8 Generator 6
4 Lines 14 - 15 9 Generator 5

In reality, for most contingencies and operating conditions, a power sys-

tem is normally stable. Therefore, the number of stable operating conditions
in the training set is normally much bigger than the number of unstable oper-
ating conditions. To handle the class imbalance, re-sampling [132] is adopted.
The basic idea of re-sampling is to repeatedly use samples from a smaller
class (unstable conditions in our case) during the training of the classifier.

5.7 Case Study Results and Analysis

The proposed DSA framework is suitable for transient stability, voltage
stability and other online security applications. In the case study, a TSA
based on the proposed DSA framework is demonstrated. To demonstrate
the advantages of the proposed DSA framework, we build another TSA tool
which uses conventional supervised learning algorithm and name it con-
ventional TSA. The conventional TSA has three classifiers consisting a small
ELM-based ensemble [6].

5.7.1 Results of the Feature Selection

Two categories of features are often used in DSA: pre-contingency (steady-
state) variables, and post-contingency variables. In this study, we use pre-
contingency variables for TSA. More specifically, we use bus voltages, gener-
ator rotor angles, active and reactive powers, load active and reactive powers
as candidate features. These variables are considered topology independent,
which is expected to better reflect a topology change and result in more ac-
curate classification [104, 126].
5.7. Case Study Results and Analysis 99

0.08 Renewable system

Conventional system
0.07

0.06
Feature weights

0.05

0.04

0.03

0.02

0.01

0
10 20 30 40 50 60 70 80 90 100
Feature ranks
F IGURE 5.3: Top 100 features as a result of feature selection.

The results of feature selection using RRelief-F are shown in Fig. 5.3 for
both the conventional and the renewable scenario. Observe that, compared
to the top-ranked features, the majority of features have little correlation with
the stability status as evidenced by the lower respective feature weights. To
increase the diversity of the three classifiers in the tri-training algorithm, we
randomly select 30 out of 50 top ranked features for each classifier as inputs
to the ELM-based classifiers in the proposed DSA.

5.7.2 Impact of the Training Set Size

The conventional supervised learning based TSA is used to demonstrate
how the labeled training set size |L| affects the classification performance.
The TSA classifier is trained using different labeled training set sizes, ranging
from under 1000 to more than 15000. The diversity of an ensemble of the
three ELM-based classifiers is obtained, first, by selecting different features as
inputs and, second, by randomly selecting weights and biases in the hidden
layer. Observe in Fig. 5.4 how the the TSA performance improves when the
labeled training set size becomes larger.
 100 Chapter 5. Semi-supervised machine Learning Method for TSA

0.99

0.98
Classification accuracy

0.97

0.96

0.95

0.94

0.93

0.92
8
37

46

64

92

38

83

29

75

9
92

82

36

14
12

15

21

30

46

61

77

92
10

12

15
Number of labeled samples used in training
F IGURE 5.4: Classification accuracy vs the training set size.

5.7.3 Impact of the Increased Penetration of Renewables

To get a sense of the impact of the increased penetration of renewable gen-
eration on the variability of the operating conditions, Fig. 5.5 shows a com-
parison between the conventional and the renewable scenario showing bus
voltage phase angle variances for all base case topology feasible operating
conditions obtained from the power flow study. Observe that the variability
is consistently larger for the renewable scenario. Two labeled training sets
generated in Section 5.6.2, one for the conventional and one for the renew-
able scenario are used separately to train and evaluate a supervised learn-
ing based TSA. As shown in Fig. 5.6, the same training set size results in
a worse classification performance for the renewable scenario. Therefore, a
larger training set is required to achieve the same classification performance,
which comes at a cost of an increased computational burden.
5.7. Case Study Results and Analysis 101

90
Conventional generation scenario
Variance of bus voltage phase angle
80
Renewable generation scenario
70

60

50

40

30

20

10

0
17 19 20 23 24 25 26 27 28 29 33 34 36 37 38
Bus IDs
F IGURE 5.5: Bus voltage phase angle variance for the conven-
tional and the renewable scenario.

5.7.4 Semi-supervised Learning Improves Online TSA Per-

formance by Using Unlabeled Samples
Next we demonstrate that semi-supervised learning using tri-training al-
gorithm can help to improve online TSA performance using unlabeled sam-
ples. Same amount of labeled samples drawn from the renewable training
set are used to train the conventional TSA and the semi-supervised learn-
ing based TSA. Other than the labeled samples, a large number of unlabeled
samples are also used in the tri-training process. Performance comparison of
the two TSA is given in Fig. 5.8. The results show clearly advantageous of
the semi-supervised learning algorithm using unlabeled samples in all cases.

5.7.5 Data Editing

In the previous section, data editing in the new TSA training is not ac-
tivated and the addition of unlabeled samples in semi-supervised learning
introduces noise. As shown in Fig. 5.7, the number of noisy samples in each
 102 Chapter 5. Semi-supervised machine Learning Method for TSA

0.992
Conventional generation scenario
Renewable generation scenario
0.99
Classification accuracy

0.988

0.986

0.984

0.982

0.98
2

4
9

82

36

91

45

00
30

46

61

77

92

10

12

13

15

17
Number of labeled samples used in training
F IGURE 5.6: Comparison of the classification results for the
conventional and the renewable scenarios.

iteration of the tri-training increases as progressively more unlabeled sam-

ples are added to the training set. The number of unlabeled samples in each
iteration is, respectively, 308, 616, 1232, 2464, 4928, and 9856. We can observe
how data editing greatly reduces the number of noisy samples in Fig. 5.7.

5.7.6 Comparison of TSA Classifiers

Now we compare the performance of the conventional TSA classifier with
the TSA classifier based on the proposed DSA framework. In the data editing
process, a time-domain simulation is called to find true label of an operating
condition if two classifiers both give credible output but does not match the
neighbor’s label (Algorithm 2 lines 13 to 17). This operating condition be-
comes a labeled sample and is therefore put in L (Algorithm 2 line 18). So
final number of the labeled samples used in the learning varies from the ini-
tial number of labeled samples. Starting with different numbers of labeled
samples, the tri-training learning and data editing algorithm performance
is evaluated first and final number of labeled samples used are recorded.
Then same amount of labeled samples are used to evaluate the conventional
5.8. Conclusion 103

500
Without data editing
Number of noisy samples introduced
With data editing

400

300

200

100

0
1 2 3 4 5 6
Training iteration
F IGURE 5.7: Comparison of the number of noisy samples per
iteration in tri-training with and without data edit-
ing.

TSA and new TSA tool without data editing as in Section 5.6. The results
are shown in Fig. 5.8. Observe how semi-supervised learning with data
editing consistently outperforms the conventional TSA classifier. Especially
for when the number of labeled samples is small, semi-supervised learning
greatly increases classification performance of the new TSA model by using
unlabeled samples.

5.8 Conclusion
In the case studies of this chapter, we first demonstrate how labeled train-
ing set size impacts on performance of the neural network-based TSA and
penetration of renewable generation increase diversity of power system op-
erating condition. The result is that more labeled training samples are re-
quired to achieve higher classification performance for conventional DSA
tools. We then compared the classification performance of the conventional
TSA tool and our new TSA tool based on semi-supervised learning. Results
 104 Chapter 5. Semi-supervised machine Learning Method for TSA

1
Supervised training results
0.995 Tri-training results (semi-supervised)
Tri-training + data editing results
Classification accuracy

0.99

0.985

0.98

0.975

0.97

0.965

0.96
09

27

00

00

00

00
15

17

21

35

65

95
Number of labeled samples used in training
F IGURE 5.8: Compare prediction results of supervised, semi-
supervised and semi-supervised with active learn-
ing algorithms.

show the new TSA requires much less labeled training samples to achieve
same performance compared to the conventional TSA. Finally, we demon-
strate the data editing algorithm can help to reduce the noise introduced in
the semi-supervised learning process and hence lead to a better classification
result. The new framework greatly reduces online simulation for the training
set preparation and improves the classification performance of DSA. The new
framework provides an alternative when timely DSA updating is required.
We used an ELM-based neural network ensemble and tri-training algorithm
in the case study, however, the idea of the semi-supervised learning-based
DSA is also suitable for other machine learning approaches and different type
of semi-supervised algorithms can be used.
 105

Chapter 6

Conclusion

Planning and online dynamic security assessment of future grids using

conventional tools are challenged by the new features of future grids such as
intermittent generation, demand response and fast responding power elec-
tronic plants which lead to much more diverse operation conditions com-
pared to the existing networks. In this research work, a set of comprehensive
new future grids security assessment tools are proposed from security as-
sessment in future grids planning and operation point of view.

6.1 Conclusion of the Presented Works

One of the major research aims is to propose a framework to assess the
security of future grids for planning purposes by analyzing a large amount of
scenarios, considering new features which are not part of the present system.
In Chapter 1, existing techniques and methods for electricity network se-
curity assessment are revisited. However, these assessment tools used for
power system planning and operation are challenged by new features of fu-
ture grids. Motivation and focus of the thesis are given following by method-
ology utilized in this study work.
Chapter 2 introduces a CSIRO project and transient stability study using
conventional time-domain simulation method. The results of the stability
study of a future grid model show dis-advantage of time-domain simula-
tions.
Chapter 3 presents a novel machine learning based framework for fast
stability scanning of future grids scenarios. The framework is based on a fea-
ture selection algorithm that makes it possible to perform clustering using
both feature ranks and weights. The case study demonstrated the suitabil-
ity of the proposed framework. Considering the level of detail required for
future grid analysis, an acceptable accuracy is achieved with a more than a
ten-fold speed-up.
106 Chapter 6. Conclusion

We show that machine learning can be used to speed-up stability scan-

ning in future grids characterized by a high penetration of variable RES. The
proposed methodology can be used by future grid operators to increase the
situational awareness so that they can identify potential stability issues on
time and with a greater accuracy.
In Chapter 4, In order to overcome classification unreliability of an ex-
isting machine learning based DSA frame for future grids online security as-
sessment, we first proposed a hybrid filter-wrapper feature selection method,
which helped to improve TSA classification performance when the test sys-
tem is evaluated with multiple topologies. Furthermore, simulation results
show that the boosting learning method in conjunction with the rule for clas-
sification we proposed based on weighted weak classifiers’ output can signif-
icantly improve TSA reliability and helped to reduce time-domain simulation
for achieving 100% classification accuracy for a future grid.
Chapter 5 presents a semi-supervised machine learning based DSA method
for future grids online security assessment. In the case studies, we first
demonstrate how labeled training set size impacts on performance of the
neural network-based TSA. We also show how an increased penetration of
renewable generation increases the diversity of power system operating con-
ditions. We then compared the classification performance of the conventional
TSA tool and our new TSA tool based on semi-supervised learning. Results
show the new TSA requires much less labeled training samples to achieve
performance comparable to the conventional TSA. Finally, we demonstrate
how data editing algorithm can help to reduce the noise introduced in the
semi-supervised learning process, which leads to a better classification re-
sult. The new framework greatly reduces the need for online simulation for
the training set preparation and improves the classification performance of
DSA. The new framework provides an alternative when timely DSA updat-
ing is required in future grids online security assessment.

6.2 Suggestion for Future Work

Machine learning techniques are used in this thesis to deal with chal-
lenges brought by future grids’ new features which deteriorate the perfor-
mance of the existing techniques used in power system planning and opera-
tion. Due to time limitation of this research study, further works can be done
to improve the study to build a set of security assessment tools for future grid
planning and operation. These are itemized below.
6.2. Suggestion for Future Work 107

• In order to focus on developing machine learning-based security as-

sessment frameworks for future grid planning and operation, less effort
was put on feature selection techniques in this study. Feature selection
approaches used in the study are conventional. For a future grid, to
find out how new features of the grid impact on the feature selection
used in a machine learning based security assessment frame is impor-
tant. Conventional feature selection techniques need to be revisited and
developed for future grids security study.

• In the proposed un-supervised machine learning-based method for fu-

ture grid fast stability scanning, clustering technique is used. The main
types of unsupervised learning algorithms include clustering algorithms
and association rule learning algorithms. The study didn’t compare dif-
ferent algorithms performance in the proposed assessment framework.
More work is suggested to find the most suitable algorithm to be used
for fast future grids planning security assessment.

• Supervised machine learning has different algorithms, such as Near-

est Neighbor, Naive Bayes, Decision Trees, Linear Regression, Support
Vector Machines and Neural Networks. In this thesis, only neural net-
works are used in the proposed security assessment frameworks. It
is suggested to test the proposed frameworks by using various algo-
rithms.

• Classification rule is used to differentiate operating points in terms of

security level. The performance of a classification rule determines the
performance of the security assessment tool. More effort is suggested
to develop optimal classification rules based on observing distribution
of security level of operating points.

• Due to time limit, small network models are used in this study. How-
ever, a bigger network model may be used to verify the performance of
the proposed security assessment tools.
 109

Bibliography

[1] Miao He, Junshan Zhang, and Vijay Vittal. “Robust Online Dynamic
Security Assessment Using Adaptive Ensemble Decision-Tree Learn-
ing”. In: IEEE Transactions on Power Systems 28.4 (Nov. 2013), pp. 4089–
4098.
[2] Ran Li et al. “Development of Low Voltage Network Templates; Part
I: Substation Clustering and Classification”. In: IEEE Transactions on
Power Systems 30.6 (Nov. 2015), pp. 3036–3044.
[3] Minas C Alexiadis, Grigoris K Papagiannis, and Ioannis P Panapakidis.
“Enhancing the clustering process in the category model load profil-
ing”. In: IET Generation, Transmission & Distribution 9.7 (Apr. 2015),
pp. 655–665.
[4] Sudipta Dutta and Thomas J. Overbye. “Feature Extraction and Visu-
alization of Power System Transient Stability Results”. In: IEEE Trans-
actions on Power Systems 29.2 (Mar. 2014), pp. 966–973. ISSN: 0885-8950.
DOI : 10.1109/TPWRS.2013.2283276.

[5] Y. Xu et al. “Real-time transient stability assessment model using ex-

treme learning machine”. In: IET Generation, Transmission & Distribu-
tion 5.3 (2011), p. 314. ISSN: 17518687. DOI: 10 . 1049 / iet - gtd .
2010.0355.
[6] Yan Xu et al. “A Reliable Intelligent System for Real-Time Dynamic
Security Assessment of Power Systems”. In: IEEE Transactions on Power
Systems 27.3 (Aug. 2012), pp. 1253–1263.
[7] Rui Zhang et al. “Post-disturbance transient stability assessment of
power systems by a self-adaptive intelligent system”. In: IET Genera-
tion, Transmission & Distribution 9.3 (Feb. 2015), pp. 296–305.
[8] Wouter F. Schmidt, Martin A. Kraaijveld, and Robert P.W. Duin. “Feed-
forward neural networks with random weights”. In: Proceedings of the
11th IAPR International Conference on Pattern Recognition, 1992. Vol.II.
Conference B: Pattern Recognition Methodology and Systems. Vol. 2. 1992,
pp. 1–4.
110 BIBLIOGRAPHY

[9] V.N. Vapnik. “An overview of statistical learning theory”. In: IEEE
Transactions on Neural Networks 10.5 (1999), pp. 988–999. ISSN: 10459227.
DOI : 10.1109/72.788640.

[10] Olivier Chapelle, Bernhard Schölkopf, and Alexander Zien. Semi-Supervised

Learning. The MIT Press Cambridge, Massachusetts London, England,
2006.
[11] European Network of Transmission System Operators for Electricity.
Annual Report 2016. Tech. rep. 2016.
[12] National Energy Administration. Annual Report 2014. Tech. rep. 2015.
[13] National Bureau of Statistics. Annual Report 2014. Tech. rep. 2015.
[14] China National Renewable Energy Center. Annual Report 2014. Tech.
rep. 2015.
[15] Neil Hodge. “Energy Risks - Power Trip”. In: Emerging Risks - Special
Topic (2015).
[16] Wikipedia. List of Major Power Outages. 2015. URL: https : / / en .
wikipedia . org / w / index . php ? title = List _ of _ major _
power_outages&action=history.
[17] Prabha Kundur et al. “Definition and classification of power system
stability”. In: IEEE Transactions on Power Systems 19.3 (Aug. 2004), pp. 1387–
1401.
[18] I.A.Hiskens D.J.Hill. Voltage Stability Analysis and Control. Tech. rep.
University of Newcastle, 1992.
[19] P.-A. Lof et al. “Fast calculation of a voltage stability index”. In: IEEE
Transactions on Power Systems 7.1 (1992), pp. 54–64.
[20] N Flatabo, R Ognedal, and T Carlsen. “Voltage stability condition in
a power transmission system calculated by sensitivity methods”. In:
IEEE Transactions on Power Systems 5.4 (1990), pp. 1286–1293.
[21] B Gao et al. “Voltage stability evaluation using modal analysis”. In:
ieeexplore.ieee.org ().
[22] BH Chowdhury, CW Taylor IEEE Transactions on Power, and unde-
fined 2000. “Voltage stability analysis: VQ power flow simulation ver-
sus dynamic simulation”. In: ieeexplore.ieee.org ().
[23] R Diao et al. “Decision tree-based online voltage security assessment
using PMU measurements”. In: ieeexplore.ieee.org ().
BIBLIOGRAPHY 111

[24] M Ramaswamy, KR Nayar International Journal of Electrical Power

&, and undefined 2004. “On-line estimation of bus voltages based on
fuzzy logic”. In: Elsevier ().
[25] R Tiwari et al. “Line collapse proximity index for prediction of voltage
collapse in power systems”. In: Elsevier ().
[26] Phraba Kundur. Power system stability and control. McGraw Hill, 1994.
[27] H. Zhou et al. “Stacked Extreme Learning Machines”. In: IEEE Trans-
actions on Cybernetics 45.9 (Sept. 2015), pp. 2013–2025. ISSN: 2168-2267.
DOI : 10.1109/TCYB.2014.2363492.

[28] Yousin Tang and A. P. S. Meliopoulos. “Power system small signal sta-
bility analysis with FACTS elements”. In: IEEE Transactions on Power
Delivery 12.3 (July 1997), pp. 1352–1361. ISSN: 0885-8977. DOI: 10 .
1109/61.637014.
[29] S. Q. Bu et al. “Probabilistic Analysis of Small-Signal Stability of Large-
Scale Power Systems as Affected by Penetration of Wind Generation”.
In: IEEE Transactions on Power Systems 27.2 (May 2012), pp. 762–770.
ISSN: 0885-8950. DOI : 10.1109/TPWRS.2011.2170183.

[30] J. Ma, Z. Y. Dong, and P. Zhang. “Comparison of BR and QR Eigen-

value Algorithms for Power System Small Signal Stability Analysis”.
In: IEEE Transactions on Power Systems 21.4 (Nov. 2006), pp. 1848–1855.
ISSN: 0885-8950. DOI : 10.1109/TPWRS.2006.883685.

[31] J. M. Campagnolo, N. Martins, and D. M. Falcao. “An efficient and ro-

bust eigenvalue method for small-signal stability assessment in paral-
lel computers”. In: IEEE Transactions on Power Systems 10.1 (Feb. 1995),
pp. 506–511. ISSN: 0885-8950. DOI: 10.1109/59.373977.
[32] J Guckenheimer and P Holmes. Nonlinear oscillations, dynamical sys-
tems, and bifurcations of vector fields. 2013.
[33] R Seydel. Practical bifurcation and stability analysis. 2009.
[34] P. C. Magnusson. “The Transient-Energy Method of Calculating Sta-
bility”. In: Transactions of the American Institute of Electrical Engineers
66.1 (Jan. 1947), pp. 747–755. ISSN: 0096-3860. DOI: 10 . 1109 / T -
AIEE.1947.5059502.
[35] P.D. Aylett. “The energy-integral criterion of transient stability limits
of power systems”. In: Proceedings of the IEE Part C: Monographs 105.8
(1958), p. 527. ISSN: 03698904. DOI: 10.1049/pi-c.1958.0070.
112 BIBLIOGRAPHY

[36] AH El-Abiad et al. “Transient stability regions of multimachine power

systems”. In: ieeexplore.ieee.org ().
[37] David J. Hill. CSIRO Future Grid Flagship Cluster Project 1: Power and
Energy Systems Modelling and Security. Tech. rep. The University of
Sydney, 2013.
[38] Hsiao-Dong Chang, Chia-Chi Chu, and G. Cauley. “Direct stability
analysis of electric power systems using energy functions: theory, ap-
plications, and perspective”. In: Proceedings of the IEEE 83.11 (1995),
pp. 1497–1529. ISSN: 00189219. DOI: 10.1109/5.481632.
[39] Y. Xue et al. “Extended equal area criterion revisited (EHV power sys-
tems)”. In: IEEE Transactions on Power Systems 7.3 (1992), pp. 1012–
1022. ISSN: 08858950. DOI: 10.1109/59.207314.
[40] Z Dong and P Zhang. Emerging techniques in power system analysis.
2010.
[41] HongJun Fu XueYong Hou Ping Ju. “Power System Static Voltage
Probabilistic Stability Analysis based on Two Point Estimation”. In:
Hohai University, China (Mar. 2010).
[42] Po-Chen Chen et al. “Analysis of Voltage Profile Problems Due to the
Penetration of Distributed Generation in Low-Voltage Secondary Dis-
tribution Networks”. In: IEEE Transactions on Power Delivery 27.4 (Oct.
2012), pp. 2020–2028. ISSN: 0885-8977. DOI: 10.1109/TPWRD.2012.
2209684.
[43] E. Haesen et al. “A Probabilistic Formulation of Load Margins in Power
Systems With Stochastic Generation”. In: IEEE Transactions on Power
Systems 24.2 (May 2009), pp. 951–958. ISSN: 0885-8950. DOI: 10.1109/
TPWRS.2009.2016525.
[44] Zhao Yang Dong et al. “Using IS to Assess an Electric Power Sys-
tem’s Real-Time Stability”. In: IEEE Intelligent Systems 28.4 (July 2013),
pp. 60–66. ISSN: 1541-1672. DOI: 10.1109/MIS.2011.41.
[45] Kip Morison, Mevludin Glavic, et al. Review of on-line dynamic security
assessment tools and techniques. Tech. rep. CIGRE, 2007.
[46] N. Balu et al. “On-line power system security analysis”. In: Proceedings
of the IEEE 80.2 (1992), pp. 262–282. ISSN: 00189219. DOI: 10.1109/
5.123296.
BIBLIOGRAPHY 113

[47] Yan Xu et al. “An Intelligent Dynamic Security Assessment Frame-

work for Power Systems With Wind Power”. In: IEEE Transactions on
Industrial Informatics 8.4 (Nov. 2012), pp. 995–1003. ISSN: 1551-3203.
DOI : 10.1109/TII.2012.2206396.

[48] Kai Sun et al. “An Online Dynamic Security Assessment Scheme Us-
ing Phasor Measurements and Decision Trees”. In: IEEE Transactions
on Power Systems 22.4 (Nov. 2007), pp. 1935–1943. ISSN: 0885-8950. DOI:
10.1109/TPWRS.2007.908476.
[49] Ruisheng Diao, Vijay Vittal, and Naim Logic. “Design of a Real-Time
Security Assessment Tool for Situational Awareness Enhancement in
Modern Power Systems”. In: IEEE Transactions on Power Systems 25.2
(May 2010), pp. 957–965. ISSN: 0885-8950. DOI: 10 . 1109 / TPWRS .
2009.2035507.
[50] Janath Geeganage et al. “Application of Energy-Based Power System
Features for Dynamic Security Assessment”. In: IEEE Transactions on
Power Systems 30.4 (July 2015), pp. 1957–1965.
[51] Bo Wang et al. “Power System Transient Stability Assessment Based
on Big Data and the Core Vector Machine”. In: IEEE Transactions on
Smart Grid (2016), pp. 1–1. ISSN: 1949-3053. DOI: 10 . 1109 / TSG .
2016.2549063.
[52] Australian Energy Market Operator Limited. 100 per cent renewable
study - modelling outcomes | Department of the Environment and Energy.
Tech. rep. AEMO, 2013, pp. 1–111.
[53] Matthew Wright and Patrick Hearps. Zero carbon Australia stationary
energy plan. Tech. rep. University of Melbourne, 2010.
[54] Ben Elliston, Mark Diesendorf, and Iain MacGill. “Simulations of sce-
narios with 100% renewable electricity in the Australian National Elec-
tricity Market”. In: Energy Policy 45 (2012), pp. 606–613.
[55] Elaine K. Hart and Mark Z. Jacobson. “A Monte Carlo approach to
generator portfolio planning and carbon emissions assessments of sys-
tems with large penetrations of variable renewables”. In: Renewable
Energy 36.8 (Aug. 2011), pp. 2278–2286. ISSN: 0960-1481. DOI: 10 .
1016/J.RENENE.2011.01.015.
[56] A Schwarzenegger. “Research Evaluation of Wind GENERATION, So-
lar Generation, AND Storage Impact on the California Grid”. In: (2010).
114 BIBLIOGRAPHY

[57] MCW Kintner-Meyer, PJ Balducci, and WG Colella. “National assess-

ment of energy storage for grid balancing and arbitrage: Phase 1, WECC”.
In: (2012).
[58] P Denholm, M Hand Energy Policy, and undefined 2011. “Grid flex-
ibility and storage required to achieve very high penetration of vari-
able renewable electricity”. In: Elsevier ().
[59] Ben Elliston, Iain MacGill, and Mark Diesendorf. “Least cost 100%
renewable electricity scenarios in the Australian National Electricity
Market”. In: Energy Policy 59 (Aug. 2013), pp. 270–282. ISSN: 0301-
4215. DOI: 10.1016/J.ENPOL.2013.03.038.
[60] Cory Budischak et al. “Cost-minimized combinations of wind power,
solar power and electrochemical storage, powering the grid up to
99.9% of the time”. In: Journal of Power Sources 225 (Mar. 2013), pp. 60–
74. ISSN: 0378-7753.
[61] I.G. Mason, S.C. Page, and A.G. Williamson. “A 100% renewable elec-
tricity generation system for New Zealand utilising hydro, wind, geother-
mal and biomass resources”. In: Energy Policy 38.8 (Aug. 2010), pp. 3973–
3984. ISSN: 0301-4215. DOI: 10.1016/J.ENPOL.2010.03.022.
[62] EirGrid and System Operator for Northern Ireland (SONI). Clean En-
ergy Solutions Center | All Island TSO Facilitation of Renewable Studies.
Tech. rep. EIRGRID, 2010.
[63] Hesamoddin Marzooghi, David J Hill, and Gregor Verbic. “Perfor-
mance and stability assessment of future grid scenarios for the Aus-
tralian NEM”. In: Australasian Universities Power Engineering Confer-
ence (AUPEC). Sept. 2014, pp. 1–6.
[64] M Gibbard and D Vowles. “Simplified 14-generator model of the SE
Australian power system”. In: The University of Adelaide, South (2010).
[65] AEMO. 2012 NTNDP Assumptions and Inputs. Tech. rep. 2012.
[66] AEMO. 2013 NTNDP Assumptions and Inputs. Tech. rep. AEMO Aus-
tralia, 2013.
[67] Shariq Riaz et al. “Generic Demand Model Considering the Impact of
Prosumers for Future Grid Scenario Analysis”. In: IEEE Transactions
on Smart Grid (2017), pp. 1–1. ISSN: 1949-3053. DOI: 10.1109/TSG.
2017.2752712.
BIBLIOGRAPHY 115

[68] Y. Xue, T. Van Custem, and M. Ribbens-Pavella. “Extended equal area

criterion justifications, generalizations, applications”. In: IEEE Trans-
actions on Power Systems 4.1 (1989), pp. 44–52. ISSN: 08858950. DOI:
10.1109/59.32456.
[69] J. Riesz, B. Elliston, P. Vithayasrichareon, I. MacGill. 100 Renewables
in Australia: A Research Summary - Google Search. Tech. rep. Centre for
Energy and Environmental Markets UNSW, 2016.
[70] CSIRO. Change and choice: The Future Grid Forum’s analysis of Australia’s
potential electricity pathways to 2050. Tech. rep. CSIRO Australia, 2015.
[71] CSIRO. Future Grid Forum: change and choice for Australia electricity sys-
tem. Tech. rep. CSIRO Australia, 2013.
[72] Greenpeace. PowER 2030 - A European Grid for 3/4 Renewable Energy by
2030. Tech. rep. Greenpeace Germany, 2013.
[73] AEMO. Emerging Technologies Information Paper, National Electricity Fore-
casting Report. Tech. rep. AEMO Australia, 2015.
[74] AEMO. Detailed Summary of 2015 Electricity Forecasts. Tech. rep. AEMO
Australia, 2015.
[75] Sara Eftekharnejad et al. “Impact of increased penetration of photo-
voltaic generation on power systems”. In: IEEE Transactions on Power
Systems 28.2 (May 2013), pp. 893–901. ISSN: 0885-8950. DOI: 10.1109/
TPWRS.2012.2216294.
[76] S.O. Faried, R. Billinton, and S. Aboreshaid. “Probabilistic evaluation
of transient stability of a power system incorporating wind farms”.
In: IET Renewable Power Generation 4.4 (2010), p. 299. ISSN: 17521416.
DOI : 10.1049/iet-rpg.2009.0031.

[77] Ahmed Khallaayoun et al. “PV and CSP integration study experience
in a Mediterranean partner country”. In: IET Renewable Power Genera-
tion 10.1 (Jan. 2016), pp. 42–49. ISSN: 1752-1416. DOI: 10.1049/iet-
rpg.2015.0142.
[78] Md Ayaz Chowdhury et al. “Transient stability of power system inte-
grated with doubly fed induction generator wind farms”. In: IET Re-
newable Power Generation 9.2 (Mar. 2015), pp. 184–194. ISSN: 1752-1416.
DOI : 10.1049/iet-rpg.2014.0035.
116 BIBLIOGRAPHY

[79] Libao Shi et al. “Effects of wind generation intermittency and volatil-
ity on power system transient stability”. In: IET Renewable Power Gen-
eration 8.5 (July 2014), pp. 509–521. ISSN: 1752-1416. DOI: 10.1049/
iet-rpg.2013.0028.
[80] K.R. Padiyar and S. Krishna. “Online Detection of Loss of Synchro-
nism Using Energy Function Criterion”. In: IEEE Transactions on Power
Delivery 21.1 (Jan. 2006), pp. 46–55. ISSN: 0885-8977. DOI: 10.1109/
TPWRD.2005.848652.
[81] Kai Sun, Stephen T. Lee, and Pei Zhang. “An Adaptive Power System
Equivalent for Real-Time Estimation of Stability Margin Using Phase-
Plane Trajectories”. In: IEEE Transactions on Power Systems 26.2 (May
2011), pp. 915–923. ISSN: 0885-8950. DOI: 10.1109/TPWRS.2010.
2055900.
[82] Kai Sun and Stephen T. Lee. “Power system security pattern recogni-
tion based on phase space visualization”. In: 2008 Third International
Conference on Electric Utility Deregulation and Restructuring and Power
Technologies. IEEE, Apr. 2008, pp. 964–969. ISBN: 978-7-900714-13-8.
DOI : 10.1109/DRPT.2008.4523546.

[83] Liancheng Wang and A.A. Girgis. “A new method for power system
transient instability detection”. In: IEEE Transactions on Power Delivery
12.3 (July 1997), pp. 1082–1089. ISSN: 08858977. DOI: 10.1109/61.
636874.
[84] D.Z. Fang et al. “Transient stability assessment using projection for-
mulations”. In: IET Generation, Transmission & Distribution 3.6 (June
2009), pp. 596–603. ISSN: 1751-8687. DOI: 10.1049/iet-gtd.2008.
0583.
[85] R. Liu, G. Verbič, and J. Ma. “A machine learning approach for fast fu-
ture grid small-signal stability scanning”. In: 2016 IEEE International
Conference on Power System Technology (POWERCON). Sept. 2016, pp. 1–
6. DOI: 10.1109/POWERCON.2016.7753894.
[86] R. Liu et al. “Fast Stability Scanning for Future Grid Scenario Analy-
sis”. In: IEEE Transactions on Power Systems 33.1 (Jan. 2018), pp. 514–
524. ISSN: 0885-8950. DOI: 10.1109/TPWRS.2017.2694048.
[87] Hesamoddin Marzooghi et al. “Generic Demand Modelling Consid-
ering the Impact of Prosumers for Future Grid Scenario Studies”. In:
(2016).
BIBLIOGRAPHY 117

[88] Liam Fahey and Robert M. Randall. Learning From the Future. Wiley,
1998.
[89] John Foster et al. Delivering a competitive Australian power system Part 2:
The challenges, the scenarios. Tech. rep. School of Economics, University
of Queensland, Australia, 2013.
[90] G. Sanchis. e-Highway2050: Europe’s future secure and sustainable elec-
tricity infrastructure. Project results. Tech. rep. 2015.
[91] “Renewable Electricity Futures for the United States”. In: IEEE Trans-
actions on Sustainable Energy 5.2 (Apr. 2014), pp. 372–378.
[92] Jaquelin Cochran, Trieu Mai, and Morgan Bazilian. “Meta-analysis of
high penetration renewable energy scenarios”. In: Renewable and Sus-
tainable Energy Reviews 29 (2014), pp. 246–253.
[93] Ben Elliston, Jenny Riesz, and Iain MacGill. “What cost for more re-
newables? The incremental cost of renewable generation – An Aus-
tralian National Electricity Market case study”. In: Renewable Energy
95 (2016), pp. 127 –139.
[94] Ben Elliston, Iain MacGill, and Mark Diesendorf. “Least cost 100%
renewable electricity scenarios in the Australian National Electricity
Market”. In: Energy Policy 59 (2013), pp. 270–282.
[95] Cory Budischak et al. “Cost-minimized combinations of wind power,
solar power and electrochemical storage, powering the grid up to
99.9% of the time”. In: Journal of Power Sources 225 (2013), pp. 60–74.
[96] J Bebic. “Power System Planning: Emerging Practices Suitable for Eval-
uating the Impact of High-Penetration Photovoltaics”. In: (2008).
[97] Jaime Quintero et al. “The impact of increased penetration of con-
verter control-based generators on power system modes of oscilla-
tion”. In: IEEE Transactions on Power Systems 29.5 (2014), pp. 2248–
2256.
[98] T. Knuppel et al. “Small-signal stability of wind power system with
full-load converter interfaced wind turbines”. In: IET Renewable Power
Generation 6.2 (2012), p. 79.
[99] M. Klein, G.J. Rogers, and P. Kundur. “A fundamental study of inter-
area oscillations in power systems”. In: IEEE Transactions on Power Sys-
tems 6.3 (1991), pp. 914–921.
118 BIBLIOGRAPHY

[100] NW Miller et al. Western Wind and Solar Integration Study Phase 3A:
Low Levels of Synchronous Generation. Tech. rep. NREL, 2015.
[101] E. Vittal, M. O’Malley, and A. Keane. “A Steady-State Voltage Stability
Analysis of Power Systems With High Penetrations of Wind”. In: IEEE
Transactions on Power Systems 25.1 (Feb. 2010), pp. 433–442.
[102] Debbie Q Zhou, U D Annakkage, and Athula D Rajapakse. “Online
Monitoring of Voltage Stability Margin Using an Artificial Neural Net-
work”. In: IEEE Transactions on Power Systems 25.3 (Aug. 2010), pp. 1566–
1574.
[103] H.A. Shayanfar, H. Razmi, and M. Teshnehlab. “Neural network based
on a genetic algorithm for power system loading margin estimation”.
In: IET Generation, Transmission & Distribution 6.11 (Nov. 2012), pp. 1153–
1163.
[104] Yan Xu et al. “An Intelligent Dynamic Security Assessment Frame-
work for Power Systems With Wind Power”. In: IEEE Transactions on
Industrial Informatics 8.4 (Nov. 2012), pp. 995–1003.
[105] Nima Amjady and Seyed Farough Majedi. “Transient Stability Predic-
tion by a Hybrid Intelligent System”. In: IEEE Transactions on Power
Systems 22.3 (Aug. 2007), pp. 1275–1283.
[106] F. Aboytes and R. Ramirez. “Transient stability assessment in longitu-
dinal power systems using artificial neural networks”. In: IEEE Trans-
actions on Power Systems 11.4 (1996), pp. 2003–2010.
[107] M. Mohammadi and G.B. Gharehpetian. “Application of core vector
machines for on-line voltage security assessment using a decision-
tree-based feature selection algorithm”. In: IET Generation, Transmis-
sion & Distribution 3.8 (Aug. 2009), pp. 701–712.
[108] T. Jain, L. Srivastava, and S.N. Singh. “Fast voltage contingency screen-
ing using radial basis function neural network”. In: IEEE Transactions
on Power Systems 18.4 (Nov. 2003), pp. 1359–1366.
[109] M Ramezani, C Singh, and M.-R. Haghifam. “Role of Clustering in
the Probabilistic Evaluation of TTC in Power Systems Including Wind
Power Generation”. In: IEEE Transactions on Power Systems 24.2 (May
2009), pp. 849–858.
BIBLIOGRAPHY 119

[110] Maryam Ramezani, Hamid Falaghi, and Chanan Singh. “A Determin-

istic Approach for Probabilistic TTC Evaluation of Power Systems In-
cluding Wind Farm Based on Data Clustering”. In: IEEE Transactions
on Sustainable Energy 4.3 (July 2013), pp. 643–651. ISSN: 1949-3029. DOI:
10.1109/TSTE.2012.2231970.
[111] AEMO. 2012 National Transmission Network Development Plan. Tech.
rep. Australian Energy Market Operator, 2012.
[112] Shariq Riaz et al. “Impact study of prosumers on loadability and volt-
age stability of future grids”. In: 2016 IEEE International Conference on
Power Systems Technology (POWERCON). Sept. 2016.
[113] Ahmad Shabir Ahmadyar et al. “Assessment of Minimum Inertia Re-
quirement for System Frequency Stability”. In: 2016 IEEE International
Conference on Power Systems Technology (POWERCON). Sept. 2016.
[114] Thierry van Cutsem and Costas Vournas. Voltage Stability of Electric
Power Systems. Springer, 1998.
[115] R. Xu and D. Wunsch II. “Survey of Clustering Algorithms”. In: IEEE
Transactions on Neural Networks 16.3 (May 2005), pp. 645–678.
[116] J.L. Rueda and D.G. Colomé. “Probabilistic performance indexes for
small signal stability enhancement in weak wind-hydro-thermal power
systems”. In: IET Generation, Transmission & Distribution 3.8 (Aug. 2009),
pp. 733–747.
[117] M Clerc and J Kennedy. “The particle swarm - explosion, stability, and
convergence in a multidimensional complex space”. In: IEEE Transac-
tions on Evolutionary Computation 6.1 (2002), pp. 58–73.
[118] D.W. van der Merwe and A.P. Engelbrecht. “Data clustering using
particle swarm optimization”. In: The 2003 Congress on Evolutionary
Computation, 2003. CEC ’03. IEEE, pp. 215–220. ISBN: 0-7803-7804-0.
DOI : 10.1109/CEC.2003.1299577.

[119] A. Ahmadyfard and H. Modares. “Combining PSO and k-means to

enhance data clustering”. In: Telecommunications, 2008. IST 2008. Inter-
national Symposium on. Aug. 2008, pp. 688–691.
[120] M Robnik-Sikonja and I Kononenko. “Theoretical and Empirical Anal-
ysis of ReliefF and RReliefF”. In: Machine Learning 53.1 (Oct. 2003).
120 BIBLIOGRAPHY

[121] C.A. Jensen, M.A. El-Sharkawi, and R.J. Marks. “Power system secu-
rity assessment using neural networks: feature selection using Fisher
discrimination”. In: IEEE Transactions on Power Systems 16.4 (2001),
pp. 757–763.
[122] K.R. Niazi, C.M. Arora, and S.L. Surana. “Power system security eval-
uation using ANN: feature selection using divergence”. In: Proceed-
ings of the International Joint Conference on Neural Networks, 2003. Vol. 3.
IEEE, 2003, pp. 2094–2099.
[123] K. Kira. “A Practical Approach to Feature Selection”. In: Proceedings of
International Conference on Machine Learning (1992).
[124] Z-S Lu and Z-R Hou. “Particle Swarm Optimization with Adaptive
Mutation”. In: Acta electronica sinica 32.3 (2004), pp. 416–420.
[125] R. Liu, G. Verbič, and Y. Xu. “A new reliability-driven intelligent sys-
tem for power system dynamic security assessment”. In: 2017 Aus-
tralasian Universities Power Engineering Conference (AUPEC). Nov. 2017,
pp. 1–6. DOI: 10.1109/AUPEC.2017.8282442.
[126] Miao He, Vijay Vittal, and Junshan Zhang. “Online dynamic secu-
rity assessment with missing pmu measurements: A data mining ap-
proach”. In: IEEE Transactions on Power Systems 28.2 (May 2013), pp. 1969–
1977.
[127] Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew. “Extreme
learning machine: Theory and applications”. In: Neurocomputing 70.1
(2006), pp. 489–501. ISSN: 09252312. DOI: 10 . 1016 / j . neucom .
2005.12.126.
[128] P. Pudil, J. Novovičová, and J. Kittler. “Floating search methods in
feature selection”. In: Pattern Recognition Letters 15.11 (1994), pp. 1119–
1125. ISSN: 01678655. DOI: 10.1016/0167-8655(94)90127-9.
[129] Thomas G. Dietterich. “Machine-Learning Research”. In: AI Magazine
18.4 (1997), p. 97. ISSN: 0738-4602. DOI: 10.1609/AIMAG.V18I4.
1324.
[130] Robert E. Schapire. “The strength of weak learnability”. In: Machine
Learning 5.2 (June 1990), pp. 197–227. ISSN: 0885-6125. DOI: 10.1007/
BF00116037.
[131] A Pai. Energy function analysis for power system stability. Springer Sci-
ence & Business Media, 2012.
BIBLIOGRAPHY 121

[132] Andrew Estabrooks, Taeho Jo, and Nathalie Japkowicz. “A Multiple

Resampling Method for Learning from Imbalanced Data Sets”. In:
Computational Intelligence 20.1 (Feb. 2004), pp. 18–36. ISSN: 0824-7935.
DOI : 10.1111/j.0824-7935.2004.t01-1-00228.x.

[133] M. Stone. “Cross-validatory choice and assessment of statistical pre-

dictions (with discussion) (Corr: 76V38 p102)”. In: Journal of the Royal
Statistical Society, Series B: Methodological 36 (1974), pp. 111–147.
[134] R. Liu, G. Verbič, and J. Ma. “A New Dynamic Security Assessment
Framework Based on Semi-supervised Learning and Data Editing”.
In: Electric Power Systems Research - Submitted (2019).
[135] J. Shu, W. Xue, and W. Zheng. “A Parallel Transient Stability Simula-
tion for Power Systems”. In: IEEE Transactions on Power Systems 20.4
(Nov. 2005), pp. 1709–1717. ISSN: 0885-8950. DOI: 10.1109/TPWRS.
2005.857266.
[136] AR Barron. “Approximation and estimation bounds for artificial neu-
ral networks”. In: Machine Learning (1994).
[137] Partha Niyogi and Federico Girosi. “On the relationship between gen-
eralization error, hypothesis complexity, and sample complexity for
radial basis functions”. In: Neural Computation 8.4 (1996), pp. 819–842.
[138] A Blum and T Mitchell. “Combining labeled and unlabeled data with
co-training”. In: Proceedings of the eleventh annual conference on Compu-
tational learning theory. ACM, 1998. (1998).
[139] Zhi-Hua Zhou and Ming Li. “Tri-training: exploiting unlabeled data
using three classifiers”. In: IEEE Transactions on Knowledge and Data
Engineering 17.11 (Nov. 2005), pp. 1529–1541. ISSN: 1041-4347. DOI:
10.1109/TKDE.2005.186.
[140] Dana Angluin and Philip Laird. “Learning from noisy examples”. In:
Machine Learning 2.4 (Apr. 1988), pp. 343–370. ISSN: 0885-6125. DOI:
10.1007/BF00116829.
[141] Tongliang Liu and Dacheng Tao. “Classification with Noisy Labels
by Importance Reweighting”. In: IEEE Transactions on Pattern Analysis
and Machine Intelligence 38.3 (Mar. 2016), pp. 447–461. ISSN: 0162-8828.
DOI : 10.1109/TPAMI.2015.2456899.

[142] JS Sánchez et al. “Analysis of new techniques to obtain quality train-

ing sets”. In: Pattern Recognition (2003).
122 BIBLIOGRAPHY

[143] A.M. Zoubir and B. Boashash. “The bootstrap and its application in
signal processing”. In: IEEE Signal Processing Magazine 15.1 (1998),
pp. 56–76. ISSN: 10535888. DOI: 10.1109/79.647043.
[144] SB Yusof, GJ Rogers, and RTH Alden. “Slow coherency based network
partitioning including load buses”. In: IEEE Transactions on Power (1993).