This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

A Review of Machine Learning Approaches to

Power System Security and Stability
Oyeniyi Akeem Alimi1, Student Member, IEEE, Khmaies Ouahada1, Senior Member, IEEE,
and Adnan M. Abu-Mahfouz1, 2, Senior Member, IEEE
1
Department of Electrical & Electronic Engineering Science, University of Johannesburg, Johannesburg 2006, South Africa
2
Council for Scientific and Industrial Research, Pretoria 0184, South Africa
Corresponding author: Oyeniyi Akeem Alimi (e-mail: [email protected]).
“This research was supported by the Council for Scientific and Industrial Research, Pretoria, South Africa, through the Smart Networks collaboration
initiative and IoT-Factory Programme (Funded by the Department of Science and Innovation (DSI), South Africa).”

ABSTRACT Increasing use of renewable energy sources, liberalized energy markets and most
importantly, the integrations of various monitoring, measuring and communication infrastructures into
modern power system network offer the opportunity to build a resilient and efficient grid. However, it also
brings about various threats of instabilities and security concerns in form of cyberattack, voltage instability,
power quality (PQ) disturbance among others to the complex network. The need for efficient methodologies
for quicker identification and detection of these problems have always been a priority to energy
stakeholders over the years. In recent times, machine learning techniques (MLTs) have proven to be
effective in numerous applications including power system studies. In the literature, various MLTs such as
artificial neural networks (ANN), Decision Tree (DT), support vector machines (SVM) have been
proposed, resulting in effective decision making and control actions in the secured and stable operations of
the power system. Given this growing trend, this paper presents a comprehensive review on the most recent
studies whereby MLTs were developed for power system security and stability especially in cyberattack
detections, PQ disturbance studies and dynamic security assessment studies. The aim is to highlight the
methodologies, achievements and more importantly the limitations in the classifier(s) design, dataset and
test systems employed in the reviewed publications. A brief review of reinforcement learning (RL) and
deep reinforcement learning (DRL) approaches to transient stability assessment is also presented. Finally,
we highlighted some challenges and directions for future studies.

INDEX TERMS Classifiers, cyberattacks, deep reinforcement learning, intruder detection system, machine
learning techniques, power quality disturbance, power system, reinforcement learning, test systems,
transient stability assessment, voltage stability

I. INTRODUCTION exposed the networks to various vulnerabilities and threats

Over the past few decades, power system operations are [3]. As a key critical infrastructure, the secure and stable
constantly being modernized so as to accommodate the operation of power system is usually treated with topmost
integration of renewable energy and storage systems (RES), priority by the governments and utility stakeholders as the
liberalized market, numerous measuring and various social, political, and economical activities are
communication technologies devices to name a few [1]. closely tied to the nation’s power system. Adversaries can
While the modernization contributed immensely to safer, access network nodes and alter measurements including
reliable and cleaner energy distribution to users, the control commands, thereby destabilizing the operation,
transition also brings along new challenges to the network’s creating blackouts, financial losses and in some situations,
security and stability [2]. The overreliance of modern national security can be put into jeopardy [4].
power system’s applications such as state estimation, Furthermore, the geometric growth in energy demands, the
Supervisory Control and Data Acquisition (SCADA) introduction of certain disturbing loads, major changes in
systems, Phasor Measurement Unit (PMU) on open network topology, the increasing strains on transmission
communication technologies including the internet have lines, etc. are dangerously pushing the power system

VOLUME XX, 2017 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

towards and beyond stability limits, thus creating power grid protection especially with the challenges
instabilities and power quality disturbances (PQD) attributed to the numerous integration of RES. In the work,
problems [5], [6]. In recent times, these instabilities and several MLTs deployed in the generation, forecasting and
disturbances concerns have drawn the attention of the integration of RES into the power system were studied.
industrial and academic communities as they are leading Also, Saini et al. [18] presented a survey on PQ
causes of outages, economic losses, equipment classifications using machine learning and signal
malfunctioning and failures. According to Electrical Power processing tools. The work focused only on power quality
Research Institute (EPRI), power supply outages crisis analysis.
created an economic loss estimates of $104 billion to $164 To bridge the research gaps, we reviewed a wide range of
billion annually [7]. Several studies in the literature have machine learning architectures and explored the power
discussed the threats, effects and impacts of these system security and stability challenges that have benefitted
disturbances, insecurities and instabilities [8]-[11]. from MLTs. This paper identifies four main power system
Monitoring the power system’s status especially during security and stability domains: SCADA network
load changes and post contingencies have always been a vulnerability and threats, PQD studies, voltage stability
major concern to energy stakeholders and operators. Over assessment (VSA) and transient stability assessment (TSA)
the years, various statistical models and signal processing where MLTs have been extensively deployed. This review
techniques [12]-[14] have been proposed in security and focuses on highlighting the methodologies, achievements
stability studies. Despite demonstrating satisfactory and limitations in the classifier(s) design, dataset generation
performances, the conventional methods have proven to be and test systems employed in the reviewed publications. A
computationally inept, expensive and time-consuming as brief review of reinforcement learning (RL) and deep
they struggle in addressing the emerging analytic needs of reinforcement learning (DRL) application especially in
the complex modern power system. TSA studies is also presented. Conventional methods for
In recent times, machine learning techniques (MLTs) have power system security and stability solutions were not
been vastly used in modelling and monitoring complex extensively reviewed in the paper as there are various
applications. Numerous MLTs such as Artificial Neural works that have covered such approaches in the literature
Network (ANN), Decision Tree (DT), Principal Component [19]-[21].
Analysis (PCA), etc. have been proposed in various The review paper is intended for power system security and
capacities involving power system security and stability stability researchers with the intention of building analytics
assessments. Unlike traditional methods, MLTs have and/or artificial intelligence security solutions for power
proven to be computationally powerful, systematic and system infrastructure, using the current and emerging
explicitly reliable when they are deployed in classification machine learning approaches. Different from the previous
studies. In the literature, various simulation results in power studies, this paper gathers together different approaches,
system’s research works have shown that MLTs have the strategies, procedures, limitations and research gaps on
capacity to learn and understand the changing MLT application to power system security and stability
characteristics of varying loads, network data, etc. that are studies. Specifically, the major contributions of this paper
peculiar to the dynamic nature of modern power system. are stated briefly as follows:
In line with the increasing concerns regarding power  A comprehensive review of the most recent state-of-
system security and stability, quite a number of review the-art ML approaches and the applicability in power
works in the literature have presented detailed description system security and stability domain;
of the problems and notable solutions including machine  The major power system security and stability domains
learning approaches. Sun et al. [9] surveyed relevant cyber (TSA, VSA, PQD and SCADA network vulnerability
security studies with regards to dangers and solutions to and threats) are extensively discussed;
improve the security of power grids. The work presented in  An elaborate review of several MLTs applied to power
[15] focused on surveying tools and techniques to uncover system security and stability problems as regards to the
SCADA systems vulnerabilities. It also addressed different classifier(s) design, dataset generation, preprocessing
methods including machine learning algorithms for techniques, optimization techniques and test systems
SCADA communication security. However, the study is deployed; and
only limited to SCADA communication aspect of the power  The challenges, limitations and research gaps of the
system. In a similar study, Glauner et al. [16] reviewed current machine learning techniques’ applications in
different MLTs, features and datasets in detecting non- power system security and stability studies and the
technical losses which includes electricity theft, defective directions for the successful deployment of MLTs in
meters and billing errors. Also, the study was limited to a future power system security and stability applications.
specific domain as the authors did not include the study of Accordingly, the remainder of the paper is organized as
MLTs on other key power system menaces. With regards to follows; an overview of power system security and stability
RES menaces, the authors in [17] presented a review on is presented in Section II. The section briefly discussed the

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

TSA, VSA, PQ disturbances and SCADA network 1) POWER QUALITY DISTURBANCE

vulnerability analytics. Section III summarizes some of the Power quality disturbances can be defined as the sudden
power system stability and security solutions in the deviation manifested in voltage/current magnitude, phase
literature. Section IV presents a detailed analysis on MLTs angle and frequency from the standard rating. The
for the power system stability and security menaces. disturbances are mostly created due to the introduction of
Section V presents the research gaps and future directions non-linear loads, switching devices, rectifiers and inverters,
and Section VI concludes the paper. etc. into the power system [7], [22], [23]. As explained by
Wang et al. [22], the continuous integrations of disturbing
II. OVERVIEW OF POWER SYSTEM SECURITY AND loads and RES are creating deterioration risks and
STABILITY malfunctioning of machineries at energy generation,
Ensuring that the power system is secured and stable transmission and consumption levels. Varieties of PQD
especially after it has been exposed to varieties of strains and effects include voltage sag, harmonic distortion, notch,
different contingencies is a major challenge that energy flicker, spikes, etc. and they are capable of creating severe
stakeholders are facing in today’s world. Recent incidents problems which include equipment malfunction, short life
worldwide have shown that the geographically dispersed span and failures [24], [25]. According to EPRI, of all power
power system are facing various security and stability crisis supply outages crisis, PQ issues accounts for approximately
that warrants comprehensive protective and preventive 15% of the economic losses estimates of $104 billion to $164
measures. Historically, ensuring the security and stability of billion annually [7]. Thus, the mitigation of PQD have
power system have always been a challenge. Operators of the continued to receive massive attention in the power system’s
old power system used to struggle in efficiently monitoring research community.
the network. In actual fact, operators usually expect reports
on trips and faults from consumers [4]. However, despite the
2) SCADA NETWORK VULNERABILITIES AND THREATS
numerous innovations that epitomizes the modern power
Critical infrastructures (CI) such as the power system, oil and
system in recent times, various evidences and reports have
gas pipelines, water distribution, etc. are monitored and
revealed that the modern power system globally are facing
controlled by SCADA systems which links the CI together as
higher numbers of security and stability challenges. Also, the
a network through advanced Information Technologies (IT)
integration of numerous internet of things (IoT)
[26]. As shown in Fig. 1, the SCADA system architecture for
infrastructures and other sophistically advanced gadgets
the electricity grid basically consist of four major operational
affiliated with the modern power system are creating security
parts namely [27], [28]:
and stability crisis to the network as some of the protocols
 The “Field” devices such as sensors for sensing the
and standards are highly vulnerable to cyberattacks and
status of SCADA equipment under concern (power
intrusion. In addition, the continuous quest for
level, pressure, etc.) and control them according to the
industrialization as well as the building of smart cities
received commands.
continue to create upsurge in unhealthy energy demand. The
 The “Remote Station” devices which include the
industrialization quest also creates the introduction of certain
Remote Terminal Units (RTUs) and Programmable
disturbing loads into the power system. The upsurge in
Logic Controllers. They serve the purpose of sending
energy demands is pushing generation and transmission
and receiving digital data to and from the control centers
facilities towards and beyond operational stability limits and
and the field devices.
the consequences are equipment failures, instabilities and
 The “Control Centre” devices consisting of the Master
PQDs.
Terminal Units (MTU) that issues command to the
Over the years, varieties of security and stability challenges
remote station devices.
have continued to plague the power system. Four major
 Human Machine Interface (HMI): devices which present
categories of power system menaces are identified as highly
processed data to operators usually via graphic user
relevant to the stable and secured operation of the modern-
interface. With the interface, operators can monitor and
day power system namely:
interact with SCADA processes.
• Power Quality Disturbance (PQD)
• SCADA Network Vulnerabilities and Threats
• Transient Stability Assessment (TSA)
• Voltage Stability Assessment (VSA)

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

FIGURE 1. SCADA Architecture [26]

Historically, when SCADA systems were first deployed, the 4. External/Deliberate – this can be described as an attack
major threat was sabotage through the physical destruction of by an external organized group that targets
the utility’s hardware as the old SCADA systems had private vulnerabilities in another nation/state power system such
and dedicated networks that are secured by traditional air as the 2015 cyberattack on Ukrainian power grid
gapped separations [29], [30]. However, over the past two whereby the hackers were linked to Russia [34].
decades, SCADA networks have been equipped with IoT These experiences and several reported cases have
devices that sometimes communicate over open channels showcased the immeasurable consequences of attacks on
which exposes the networks to numerous vulnerabilities and SCADA networks.
network based cyber-attacks [28]. These threats and attacks
are projected to escalate in geometric rates in the nearest 3) TRANSIENT STABILITY ASSESSMENT
future as intruders/attackers find the energy infrastructures Various incidences such as some specific blackouts in the
(arguably the most important of all CI) as a lucrative avenue United States, some European and Asian countries have
to gain attention [31]. The Industrial Control Systems Cyber shown that instabilities arising from RES, increasing energy
Emergency Response Team (ICS-CERT) announced that, out demands and disturbances demands the necessity for
of the 245 recorded cyber incidents on CI in 2014, 79 were improved tools in monitoring dynamic security of modern
targeted at the energy sector [2]. power system. According to [35], evaluating the capacity of a
Based on the motives and the cause of attacks, SCADA power system to withstand and survive a finite set of
threats and attacks can be categorized as [32]: contingencies to an acceptable steady-state condition is
1. Internal/Malicious – operators, employees or contractors termed dynamic security assessment (DSA). As shown in
with intentional motives to cause disasters to the Fig. 2, transient stability, voltage stability and frequency
SCADA network. Example is the well-publicized stability are identified as the main DSA categories.
stuxnet worm attack by a resentful engineer via a Transient stability assessment (TSA) provides system
removable drive [15]. operators analytical judgment of the power system dynamic
2. Internal/Non-malicious – operator making an accidental performance under various contingencies. As defined by
mistake that causes harm to power system James et al. [36], transient stability is the ability of all the
infrastructures. Example is the 2003 Ohio Davis-Besse generators to preserve synchronism after a severe disturbance
nuclear plant "Slammer" worm infection that led to the such as a fault or sudden loss of generator/ load or other
plant being disabled for hours [33]. components. Transient instability is among the key causes of
3. External/Opportunistic - hackers seeking a challenge or numerous power instability scenarios including islanding and
playing around. widespread blackouts experienced recently in Iran and

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

FIGURE 2. Dynamic Security Assessment Architecture [35]

Russian [37]. According to the authors in [38], system Based on the time frame, voltage instability crisis can be
operators must be able to evaluate the power system stability categorized into short term and long term [43]:
condition and, if needed, organize corrective actions to  Short term voltage stability (STVS)- STVS problems
preserve the needed stability in order to avoid the blackouts occur within a short time frame (seconds) after the fault
and failures. Traditionally, TSA mathematically corresponds clearing. The short term voltage instability is mostly
to solving a set of high-dimensional non-linear differential triggered by the dynamic characteristics of complex
algebraic equations (DAE). Conventional TSA methods such induction motor loads.
as the time domain simulation (TDS) and direct methods  Long term voltage stability- unlike the STVS, the time
(e.g. transient energy function (TEF) methods, the extended frame for long term voltage instability problems or
equal-area criterion (EEAC) and Lyapunov exponents have collapse is longer (average of 0.5-30 minutes). Long
been widely proposed in numerous literature. Generally, the term voltage instability problems are mostly experienced
main limitation of these methods is the large computational in heavily loaded power system. The long term voltage
effort to evaluate the swing curves for all the generators, for instability is mostly triggered by slow acting equipment,
different load levels, faults and clearance times. Thus, the such as tap-changing transformers and current limiters
conventional TSA methods do not meet the requirements [44].
demanded by modern power system. The consequences of the voltage instability problems that
4) VOLTAGE STABILITY ASSESSMENT were experienced in France, Tokyo and more recently in
While the attention has mostly focused on transient stability, Israel have shown the economic and social significances of
another increasingly important dynamic security topic in voltage instability, hence making voltage stability analysis
modern power system is voltage stability. Unlike transient one of the most discussed topics in the power system
stability that deals majorly with generators’ synchronism, research world. Voltage stability analysis involves
voltage stability is majorly tied to load dynamics as well as monitoring the power system reaction to continuous change
reactive power management [39]. Voltage stability denotes in generator and load dynamics.
the power system’s ability to maintain the bus voltages at
acceptable values after a disturbance from a given operating III. POWER SYSTEM STABILITY AND SECURITY
SOLUTIONS
condition [40]. The authors in [41], [42] explained that, the
In the olden days, majority of power system stability and
voltage profile of a power system is usually preserved within
security challenges were judged on visual inspections based
a stable range, however when the power system encounter a
on operators’ knowledge and experiences. However, more
significant fault or disturbance, several incidents may occur.
recently, several conventional methods have been proposed
These incidents include: the voltage profile may lose stability
to maintain the steady operation of the heterogeneous power
thereby instituting the induction motors to decelerate
system. In the literature, numerous signal processing
radically, draws high reactive current and ultimately lead to a
techniques have been proposed for PDQ studies over the
progressive and uncontrollable decline in voltage
years. Also, various mathematical formulations that focused
magnitudes. Apart from blackouts, other effects of voltage
on estimating the power system’s margin towards voltage
instability on the network includes the swift removal of
instability, time-domain simulations and direct methods have
generator(s) or transmission element(s) and low voltage
been proposed for DSA studies. Furthermore, as Intrusion
supply [8]. Thus, a blackout may ensue within few minutes.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

Detection Systems (IDSs) are adjudged as the de-facto environment, alters the state of the environment by taking
protection methodology for information technology systems, some control actions. Afterwards, they observe the effects of
several statistical and theoretical formulations have been the actions in order to maximize the notion of cumulative
applied as models for various SCADA intrusion detection reward [47]. Examples of RL algorithms include Q-learning,
schemes down the years. However, due to varieties of DQN, SARSA, DDPG, etc. [48]. From the conventional
factors, these statistical formulations are too rigid and point of view, MLTs for power system studies are generally
computationally inept for some specific scenarios. Hence divided into three major phases: dataset generation, data
they are incapable of effectively protecting the modern preprocessing (feature selection/extraction) and
heterogeneous systems [33], [45]. In addition, with the evaluation/classification. In some studies, authors deployed
deployment of PMU devices into modern power system, various optimization techniques and ensemble multiple
significant progress has been made with regards to efficient machine learning algorithms to boost the classification
measurement documentation as the PMU devices provide performances. The rest of this section explicitly discusses
time synchronized phasor measurements thereby enhancing some of the power system studies that have been done using
fast decision making and control actions which ultimately MLTs.
assist in creating a path for the successful realization of
accurate and effective DSA. However, important factors such A. MLTs for power quality disturbance classification
as the enormous data associated with PMU devices, the In any real power system, there are multiple sources and
uncertainties associated with measurement errors, intrusion types of power quality disturbances, hence the accurate
on PMU devices and most importantly, the non-linearity and detection and classifications of specific events are highly
computational complexity of the current and future power crucial [49]. Due to its massive success, various research
system operations, etc. have exposed the limitations of works in the literature have proposed several combinations of
conventional security and stability methods in effective machine learning algorithms and signal processing
mitigation. The need for proactive, well-calibrated, fast, techniques for detecting and classifying PQD events.
reliable and advanced security and stability methods for Basically, the combination follows the architectural
modern-day power system have become essential to all framework shown in Fig. 3.
energy stakeholders especially in the face of incessant
blackouts and other power system menaces. In the last three
decades, machine learning algorithms are proving extremely
efficient in power system studies. MLTs have been widely
proposed in power system studies involving the monitoring
and classification of various power system menaces.

IV. MLTs FOR POWER SYSTEM MENACES

It is becoming increasingly challenging to protect the
modern-day power system using conventional methods.
MLTs and deep learning techniques are continuously proving
to be a feasible option as they tick the various security factor
boxes i.e. high performance, high speed of execution and
FIGURE 3. PQD classification steps
efficiency. Machine learning methods have been widely
proposed in power system literature for monitoring, intrusion
detection, prediction and classification of various power 1) PQD FEATURE EXTRACTION/SIGNAL ANALYSIS
system menaces. Based on the deployed learning STAGE
mechanisms, Sharma and Wang [46] categorized existing In the literature, synthetic parametric equations of power
MLTs and deep learning techniques into supervised learning, quality disturbances based on IEEE standards are simulated
unsupervised learning and reinforcement learning. With using software such as MATLAB to generate single and
regards to supervised learning, the learning techniques multiple classes of large dimension feature dataset [23], [49]-
require the prior knowledge of a training dataset to learn a [54]. The large dimension feature data generated is not ideal
link between the input as well as the expected output. to be used as the input of the classifier as they will
Examples of supervised learning techniques include NN, significantly increase computational time and reduce
SVM, DT, etc. On the other hand, unsupervised learning classification accuracy. Hence there is usually a need for the
techniques attempt to find hidden patterns in the data without extraction of the dominant features for classification [53]. In
the need of any labelled data, training dataset and expected the literature, several signal processing techniques have been
output. Examples of unsupervised learning include K-Means, deployed for extracting the dominant features of PQD
PCA, etc. With regards to reinforcement learning techniques, waveforms. Various time domain signal processing
a learning agent observes and interacts with a system techniques such as EMD [24], [55] and frequency domain

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

techniques such as FT [56] have been successfully deployed frequency resolution as they provide a short window for high
for feature extraction in MLT based PQD studies. Despite the frequency components and long window for low frequency
fact that the methods presented good performances, the fact components. In a similar work involving two signal
that either frequency or time domain techniques cannot processing technique, Biswal et al. [64] used EMD to
analyze signals at neither time nor frequency domain separate out intrinsic mode functions and applied HT on the
respectively is a massive limitation. As better options, several intrinsic mode functions to extract instantaneous amplitude
authors have deployed time-frequency domain techniques as and frequency components.
they are capable of extracting features from both domains.
The authors in [57]-[59] deployed wavelet transform as it 2) PQD FEATURE SELECTION/OPTIMIZATION STAGE
offers good time-frequency characteristics and it is well In order to enhance the PQD event classification
known to have excellent ability in analyzing local performances, various authors proposed the deployment of
discontinuities of signals [51]. In similar studies, Alshahrani several feature selection/reduction and optimization
et al. [60] and Naik et al. [61] opted for DWT and WPT techniques. The main objective of using these techniques is
respectively as they offer additional advantage with regards to eliminate redundant features such as noise in the extracted
to fixed window size. Also, in order to boost the feature signals and to optimize the classifiers’ performance accuracy
extraction performance and speed of extraction, Abdoos et [53]. Most feature selection methods for PQD feature vectors
al. [49] deployed VMD and ST signal processing as they can be either wrapper based or filter based [49]. As explained
have few tuning parameters in comparison to several other Abdoos [49] explained that wrapper based feature selection
methods. As a widely used signal processing tool, WT have are time consuming but highly efficient while filter based are
various mother wavelet filters and decomposition levels. faster as they rank features based on intrinsic attributes. Ahila
Manimala et al. [54] explained that the choice of the mother et al. [51] deployed a PSO based wrapper selection model to
wavelet is crucial in wavelet analysis and it can affect the obtain the optimal number of hidden nodes and to select the
analysis results. The authors further explained that suitable beneficial subset of features in their PQD classification
mother wavelet can be determined based on properties such study. Despite the fact that the work in [50] did not deploy
as PQ indices calculation, orthogonality, maximum number any feature selection method, the authors acknowledged that
of vanishing moments, and compactness support. In the the use of optimization technique for feature selection creates
literature, the most frequently used mother wavelet class for better classification even though they may require huge
PQD studies is the fourth order daubechies 4 wavelet (DB4) computational resources, time and complex simulations. As a
as it possesses the described characteristics and it is known to means to obtain optimal structure coupled with reduced
have a close similarity to power disturbance signal [51], feature vector dimension, Abdoos et al. [49] deployed
Daubechies was deployed in studies such as the works in sequential forward selection (SFS) and sequential backward
[23], [50], [57], [62]. With regards to feature vector selection (SBS) as wrapper based methods and Gram–
decomposition modes, various authors have decomposed Schmidt orthogonalization (GSO) as feature selection and
generated signal samples at various resolution levels to get optimization techniques respectively. Using the
wavelet coefficients that suits their classification combinations, the authors were able to eliminate redundant
specifications. Eristi and Demir [62] applied a 8-level features, reduce the computational cost and most importantly,
wavelet decomposition to construct a nine-dimensional improved the generalization capability of the deployed
feature vector that suits their RBF kernel-based SVM classifier in their PQD study. In a related work, Bosnic et al.
classification study. Furthermore, at each decomposition [53] also deployed SFS to identify the most discriminative
level, different numbers of statistical methods such as features in the PQD feature vector in their classification
standard deviation, mean, skewness, kurtosis, RMS, Shannon study. As GSO is well known for its numerical instability
entropy, log energy entropy and norm entropy, etc. are with respect to rounding error, Liquan et al. [57] opted for
calculated for feature vector which are used as the input the highly rated particle swarm optimization (PSO) in
vector for the MLT classification. Bosnic et al. [53] deployed optimizing the SVM parameters in their PQD classification
mean, standard deviation, skewness, kurtosis, RMS, Shannon study. In another work that involves the use of heuristic
entropy, log energy entropy and norm entropy while optimization technique, Khokhar et al. [23] used ABC
Kanirajan and Kumar [63] used standard deviation, variance, algorithm to select optimal features in their PQD event
norm, median, absolute deviation and mean absolute classification study. In order to avoid combinational
deviation as it suits the specification of their proposed problems that are peculiar to some of the popular heuristic
classification models. For better performance, various optimization techniques, the authors in [65] used ant colony
authors have explored the use of multiple extraction optimization (ACO) as the feature selection/optimization
techniques whereby the advantage of one signal processing technique in their PQD study. Similarly, Manimala et al. [54]
tool is boosted by the ensemble partner. Using the used Genetic Algorithms (GA) and Simulated Annealing
combination of WT and MRA, Kanirajan and Kumar [63] (SA) optimization techniques for the selection of the most
argued that WT and MRA provides an excellent time

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

dominant features for the SVM classifier deployed in their machine learning algorithm(s) that the authors deployed in
study. the classification of PQD events are presented. As shown in
Table I, the ideas of MLT approaches to PQD events
3) CLASSIFICATION STAGE detection and classifications have been hugely successful.
In the literature, numerous machine learning and deep Furthermore, it can be observed from Table I that the most
learning algorithms such as SVM [49], [53], [55], [57], [59], adopted machine learning tools is the SVM as it guarantees
DT [64], [66], K-Means [67], various ANN types including efficiency and high accuracy. In addition, it can be observed
Probabilistic Neural Network (PNN) [23], [68]-[70], Feed that the integrations of optimization techniques have positive
Forward Neural Network (FFNN) [51], [61], Radial basis influences on the classification results.
function network (RBFNN) [63], [67] and deep neural
networks such as Convolutional Neural Network (CNN) B. MLTs for SCADA Network Vulnerabilities and
Threats
[22], [52], etc. have been employed in PQD classification.
However, despite the numerous varieties of classifiers’ The widespread cyber presence especially in the form of
proposed for PQD classification, SVM and PNN are the most advanced communication gadgets and IoT deployments in
widely deployed, owing to some of their attributes that today’s SCADA network have raised the power system’s
makes them highly suitable for PQD classifications. Being a vulnerabilities to security threats, attacks and intrusions that
derivation of Bayesian and kernel fisher discriminant conventional security measures such as whitelisting,
analysis algorithm, PNN are well known for their high encryption algorithms, authentication, antivirus programs,
accuracy and excellent classification performance in studies firewalls, traditional IDS, etc. are incapacitated as mitigation
involving signal outliers. Khokhar et al. [23] and Huang et strategies [45], [71], [72]. In the past decade, the world have
al. [69] explained that PNN is highly efficient in classifying witnessed how various forms of cyberattacks on SCADA
PQD events as they do not require initial weight settings network have geometrically increased [3], [45]. It is
compared to other neural network models such as FFNN and increasingly imperative to devise effective IDSs that can
RBFNN. Also, as a technique that is based on the Vapnik– efficiently detect attacks and intrusions in early stages. The
Chervonenkis dimension theory of Statistical Learning ability of MLTs in autonomously learning, adapting to
Theory and structural risk minimization principle, SVM is variations and acting without being pre-programmed have
highly rated as a powerful classifier. Various authors in the enhanced their reputation as credible methodologies for
literature have successfully deployed SVM in numerous intelligent and efficient IDSs in recent times. Traditionally,
power system studies. Abdoos et al. [49] used a RBF kernel MLTs approach for SCADA network IDSs involve three
SVM for classifying PQD events. In the study, the authors major stages: (1) dataset generation, (2) Data Preprocessing
used improved ant colony optimization algorithm to and (3) Classification/ Detection.
determine the parameters of the RBF kernel. Eristi and
Demir [62] also deployed RBF kernel SVM as the authors 1) DATASET GENERATION STAGE
argued that the kernel type can behave like a linear kernel or Traditionally, for an intruder to compromise a SCADA
a sigmoid kernel under different parameter settings. In a network, it is predictable that the intruder will somehow
comparative study conducted in [54], the authors compared create a footprint or disruption no matter how marginal.
the result of rbf kernel SVM and polynomial kernel SVM. SCADA network attack mitigation using MLTs involves the
The authors achieved a better classification accuracy with capture and analysis of SCADA network data traffic to build
RBF kernel compared to polynomial kernel. However, in a training and testing dataset. Even though communication
another comparative study on a developed model involving between SCADA components can be performed by different
three classifiers (DT, k-nearest neighbors (KNN) and SVM), network protocols, the most widely deployed in the literature
Singh et al. [65] adjudged DT as the most effective classifier is the MODBUS over TCP/IP [3], [28], [45]. However, due
as it presented the best result in terms of accuracy and to the unavailability of real-time SCADA dataset, researchers
classification time. In a work that involves neural network, make use of publicly available datasets such as the simulated
Kanirajan and Kumar [63] deployed Gaussian functions for Mississippi State University (MSU) SCADA laboratory gas
their RBFNN. In the study, the authors varied the weight by pipeline dataset [28], [73], [74] and KDD99 dataset [75].
updating it at every iteration so as to boost the classification Other notable public dataset include the cybergym dataset
accuracy. [3], UC Irvine machine learning repository dataset [73] and
the University of Arkansas’s National Center for Reliable
Electric Power Transmission (NCREPT) testbed dataset [72].
4) COMPARISON AND DISCUSSIONS OF MLT FOR PQD
It is worth mentioning that some authors such as Shitharth
CLASSIFICATION
[76] and Gao et al. [45] simulated SCADA testbeds in their
In Table I, we summarize some recent works that involved
classification studies. Shitharth [76] designed a SCADA
the deployment of MLTs in the detection and classification
network structures that comprises of 100 sensor nodes using
of PQD events. The feature extraction techniques,
Network Simulator 2 to generate the SCADA datasets.
optimization/feature selection technique(s) adopted and

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

TABLE I COMPARISON OF RECENTLY PROPOSED METHODS FOR PQD EVENTS CLASSIFICATION USING MLT

Machine
Feature extraction Optimization/Feature
Ref. Learning Accuracy Brief summary of the work’s methodology
technique selection technique
Algorithm achieved
[49] VMD + ST SVM SBS/SFS/ GSO 99.66% Large feature vector involving 9 classes of events. 3
decomposition modes used. Few tuning parameters used for
VMD and ST. SBS gave the best accuracy. Robust algorithm
with good result in noisy environment.
[50] DWT + MRA SVM - 94% Used small dataset, Gaussian kernel SVM for feature mapping
& classification, DB4 at 8-level of decomposition, Extracted
features of energy, entropy & standard deviation. 27
dimensions of features, Limitation- algorithm was not used on
large scale dataset and it may not yield good result with large
data
[57] WT SVM PSO 95.83% DB4 wavelet as mother wavelet, 8 feature vectors as classifier
input, PSO to optimize SVM classifier parameter. RBF kernel
to classify wavelet energy difference. The RBF kernel
function performed better compared to single kernel function
[23] WT PNN ABC 99.875% Used ABC for optimization as it converges rapidly, it has
good memory, etc. PNN was used to select dominant features
and spread constants and the selected features are evaluated
using RBF and MLPNN. Limitation- some aspect of PQD
such as inter-harmonic disturbances were not used in the study
[52] CT CNN SSA 99.52% Wrapping method used for extraction. Lag-covariance matrix
of PQD waveform was constructed using trajectory matrix
algorithm. CT and MSSA were used for waveform
decomposition into 6 different levels. 6 frequency bands are
used as features. Dropout technique to avoid overfitting, ReLu
for CNN activation. Robust algorithm with excellent results
even in noisy conditions.
[51] WT ELM PSO 97.60% Wrapping method of extraction based on PSO. Wavelet
energy criterion used. DB4 used. Sampled signals
decomposed with 13-level MRA.
[53] WT ECOC-SVM SFS 98.69% Voltage disturbance generated data used. WT used to
decompose at 6-levels. 8 statistical methods to extract
features. 200 instances for training and testing of SVM-ECOC
classifier. Robust algorithm in noisy and noiseless conditions
[54] WPT SVM GA 98.33% WPT used to decompose at 4-levels. GA and SA to select
dominant features out of 128 features for the RBF kernel
SVM. 10-Fold validation used during evaluation.
[62] WMRA SVM - 99.71% ATP/EMTP used to generate PQ events. WMRA was used to
extract features of a 3-phase voltage waveform. DB4 used.
Data preprocessing involving normalization used. 10fold cross
validation of SVM for kernel and penalty parameter were
done during classification.
[67] HHT RBFNN - 94% HT and some statistical methods were used for extraction into
real and imaginary features. During training, K-Means was
first used for clustering. Algorithm performs well in noiseless
environment
[64] EMD+HT BNT - 97.90% EMD was used to separate non-imaginary signal patterns into
IMFs and also performs sifting. HT was used to extract
amplitude and frequency feature patterns from the IMF. SD
and entropy were taken as features for classification using
BNT.
[63] WT RBFNN PSO 97.85% 20 types of PQD events deployed. DB4 and symlet used. 4-
level of decomposition used for analysis. Gaussian function
used during classification. Classification result compared with
FFML, LVQ, GRNN and PNN.
[70] EMD+HT RBFNN PSO 97.85% EMD was used to separate features into IMF. HT was applied
to the first 3 IMF to acquire amplitude and phase that were
used for feature vector construction. PNN was used for
mapping and classification.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

In a similar study, Gao et al. [45] simulated a SCADA deep neural networks including CNN [72], RNN [30], [45],
testbed using virtual host Nova as MODBUS master and etc. have been extensively deployed for SCADA network
PLC by HoneyD as slaves to generate two separate datasets. monitoring, intrusion prediction, detection and classification
In order to have a balanced dataset for the classifiers’ training studies in recent times. Logically, most SCADA IDS models
and evaluation, some intrusive actions in form of are binary classification, which makes them well suited for
cyberattacks are usually integrated into the simulated SVM models. Hence, numerous forms of SVM models have
testbeds. Typical intrusive attacks into SCADA testbeds been proposed in the literature. Jiang [32] and Maglaras [80]
including command injection, response injection and denial successfully deployed K-Means clustered OCSVM in
of service attack, man-in-the-middle attack, etc. are usually classifying data as normal and flagging the anomalies into
incorporated into the simulated testbed [3], [33], [46]. clusters. Explaining the choice of OCSVM as the classifier
Sufficient network traffic that contains both normal traffic deployed in their study, Maglaras et al. [77] argued that
and the abnormal traffic (due to the intrusion) are captured as unlike several other MLTs, OCSVM do not need any labeled
dataset for data preprocessing. data for training. Hence making it ideal in SCADA IDS
environment. In a similar study that adopted OCSVM, the
2) DATA PREPROCESSING STAGE author in [77] successfully used a RBF kernel for training the
Preprocessing processes such as feature reduction, developed OCSVM classifier. In a comparative study, Da
selection/extraction, mapping and scaling are highly Silva [81] compared the result of OCSVM with SVDD
important for efficient SCADA intrusion classification [3], models in successfully detecting SCADA system
[74], [77]. The main purpose of deploying these cyberattacks. The authors revealed that both algorithms are
preprocessing techniques is to have a well-organized dataset well suited for the classification task. Another prominent
that can be used for efficient training, testing and validation machine learning algorithm that is well recognized in
at the classification stage. According to Ullah and Mahmoud SCADA network IDS literature is the neural network
[74], the preprocessing procedures assist in removing models. Kalech [3] described the efficacy of SOM by
irrelevant and redundant features which can cause overfitting, combining it with HMM in successfully detecting temporal
skew predictions and misclassification. The preprocessing patterns on two different MODBUS datasets. In a related
techniques also have positive influences on the approach, Shitharith [76] deployed IWP-CSO in optimizing
classifications’ computation time. Various preprocessing the input features for a Hierarchical neuron architecture
techniques such as categorical labelling, cleaning, scaling, based neural network (HNA-NN) classification model.
extraction, selection, standardization and normalization are Several authors in the literature ensemble multiple machine
typically deployed on raw collected SCADA network learning algorithms as the limitation of one algorithm can be
dataset. To reduce the dimension of dataset, Khan et al. [78] boosted by its partner thereby creating better classification
used PCA, CCA and ICA. To improve on the reduced data algorithms that offer improved performances and robustness.
features, Khan et al. [78] further used bloom filter and Nader et al. [33] ensemble SVDD and KPCA for classifying
AllKNN for balancing and re-sampling of the dataset intrusion on the MSU dataset. In a related work, the authors
deployed in their study. In a similar study, Mansouri et al. in [82] successfully deployed CNN and RNN for
[73] deployed, PCA, ICA, GHA, SVD and SOM for cyberattacks detection using a SWAT dataset. Also, Gao [45]
dimension reduction to boost the NN classification accuracy ensemble LSTM-RNN and FNN for detecting cyberattacks
in their SCADA security study. With regards to feature on a simulated SCADA testbed.
extraction procedures, Kalech [3] deployed five feature In order to identify which machine learning algorithm is
extraction methods including feature extraction based on better suited for a particular dataset, several authors explored
function code and feature extraction based on time factor, different machine learning algorithms on the same dataset.
etc. In a related work, Wang et al. [75] deployed information Beaver et al. [83] explored the feasibility of different
gain for ranking dataset feature and for selecting dominant algorithms including NB, SVM, RF, OneR, J48, NNge in
features before utilizing SVD to obtain a low-dimensional detecting command and data attack injections into SCADA
approximation of the original feature. Ullah and Mahmoud network. From the achieved result, the authors acknowledged
[74] also used information gain for ranking features, NNge and RF as the best performers compared to the other
InfoGainAtributeEval as evaluator and ranker as search models. In a similar study, Qu [84] deployed some
method and filter based approach for feature selection. In supervised learning models including DT, KNN, SVM, RF,
another related work, Perez et al. [28] used Gaussian Mixture LR and OCSVM for SCADA network traffic IDS. From the
Model (GMM), K-means cluster, Zero imputation & result, the authors acknowledged OCSVM as the best
indicators and forward-filling techniques as data cleaning performer. In a similar comparative study, Perez et al. [28]
methods in their SCADA IDS study. compared the performance of SVM and RF using Mean,
3) DETECTION AND CLASSIFICATION STAGE
standard deviation and min-max approaches in classifying
Various MLTs such as SVM [28], [32], [77], K-Nearest both binary and a 7-class dataset that contains 20 features of
Neighbour [78], [79], LR [75], various neural networks and MODBUS packet datasets.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

TABLE II COMPARISON OF RECENTLY PROPOSED MLT BASED IDS METHODS FOR SCADA NETWORK PROTECTION

Data preprocessing/
Dataset Machine learning Accuracy
Ref. Optimization Brief summary of the work’s methodology
deployed algorithm achieved
techniques
[45] Simulated FE FNN +LSTM 99.76% Simulated SCADA testbed to generate 2 MODBUS datasets using
Testbed virtual host Kali and PLC by HoneyD. Defense Wall was used to
extract 19 features. Feature scaling was done using mean and standard
deviation. Keras Tensorflow for MLT implementation. AdamOptimizer
was used for training and softmax cross entropy was used as the loss
function.
[28] MSU data FE SVM-RF 99.58% MSU gas pipeline data containing 274,628 instances with 20 features
of MODBUS packets was used. Data cleaning done using GMM, K-
Means, zero imputation & Indicators and Keep prior value. Mean,
standard deviation and min-max methods were used as data
transformation methods. Keras Tensorflow and Scikit Learn was used
for classifier implementation.
[77] Simulated RBF Kernel OCSVM 98.87% Model trained offline using network traces after data were extracted
data from a network dataset. Data scaling and mapping into numerical
variables were done as preprocessing methods. RBF-kernel used.

[3] Cyber- FE HMM-ANN - Five feature extraction methods including function code, time factor
gym data etc., used. The SOM-ANN algorithm was built using 100 neurons. The
and BGU model showed excellent result on both datasets.
data
[76] KDD 99 IWP-CSO HNA-NN 93.1% Designed network structure using NS2. The model comprises of 100
sensor nodes to generate packet data. IWP-CSO was used for feature
optimization of the HNA-NN classifier. MATLAB was used for the
classifier’s implementation.
[33] MSU data FE SVDD/KPCA - MSU gas pipeline data involving 6 different cyberattacks scenarios. 5
fold cross validation was used for the optimization of the parameters of
the classifier.
[78] MSU data PCA, CCA, ICA, Boom filter + 97% MSU gas pipeline data involving 35 different cyberattacks scenarios.
AllKNN KNN Categorical labelling, data normalization, dimension reduction
balancing and re-sampling of data were all performed during
preprocessing. RF, AdaBoost, MLPNN and QDA were used for
training dataset. 10-fold cross validation used. Proposed method gave
the best result in comparison with other methods.
[74] MSU data InfoGainAttributeE Bayes-Net 99.5% MSI gas pipeline data used. InfoGainAttributeEval was used as
val +J48 evaluator and ranker for feature processing. 5-fold validation was used
for evaluation.
[73] MSU data PCA, ICA, GHA, MLPNN 97.5% MSI gas pipeline dataset containing 3 categories of features used. Data
SVD, SOM contains 97019 samples out of which 96019 samples was used for
training and the rest was used for testing. Feature reduction was done
using PCA, ICA, GHA and SOM. Also, Greywolf algorithm was used
for training the NN model. 10-fold cross validation was done. The
result of the proposed model was compared with 24 other ML approach
and the proposed model gave the best result.

4) COMPARISON AND DISCUSSIONS OF MLT

classification studies have been successful. From Table II,
APPROACH FOR SCADA NETWORK IDS the accuracy achieved explained that most of the attacks that
Table II presents the summary of some recent research works were integrated into the various systems have been reliably
that deployed the use of MLTs for SCADA network detected or classified. However, as most research works
intrusion detection schemes. In Table II, we present a depend solely on simulated testbeds and open source
summarized description of the dataset simulation/deployed, datasets, most of the proposed schemes cannot be used to
the data preprocessing/ optimization technique(s) adopted efficiently evaluate the potency of the model in practical real
and the machine learning algorithm(s) that some authors in life systems.
recent literature deployed in the classification of SCADA
dataset. C. MLTs for TSA
As shown in Table II, the deployment of machine learning Power system stability monitoring and assessment with
algorithms for SCADA network intrusion detection and regards to its operation positioning neat stability margins is
highly important for the efficient operation of the

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

infrastructure [85]. The conventional TSA methods are proposed prediction algorithm in their TSA study. To
computationally incompetent in handling large scale modern generate input data for their algorithm, Mosavi et al. [85]
power system due to the enormous data involved. A report used a set of size and type independent trajectory features
stated that an estimated 2500 PMU devices was installed in (s&tIFs) that measures suitable awareness level of the
Chinese North Interconnection and each of them records network status and its distance from instability.
more than 30 features every 20 milliseconds [84]. With such
huge amount of data collected at high speed, using 2) FEATURE PREPROCESSING AND OPTIMIZATION
conventional TSA methods including time domain In order to remove redundancies and improve classification
simulation and transient energy function alone are and prediction of transient instability, various authors have
computationally intensive and may not meet the requirements proposed numerous feature reduction, selection and
of modern-day real-time TSA. Researchers in recent times optimization techniques that fits the type of datasets and
have therefore turned to MLT, RL, DL and DRL approaches. classification algorithms they proposed. In the TSA study
These learning techniques have been heavily deployed in conducted by Li et al. [90] using PMU post fault data, the
recent studies peculiar to power system control problems authors were able to reduce the features of their dataset to
including TSA studies. Different from conventional methods, one-third using binary jaya feature selection technique. In a
these techniques have the capacity to process large amounts similar study, Li et al. [94] deployed a feature selection
of PMU data, analyze them and classify the corresponding method based on kernelized fuzzy rough sets (KFRS) and
stability state of a system accordingly [86]. Malbasa et al. memetic algorithm. Also, the authors in [98] chose
[87] explained that machine learning approach has the Sequential Forward Selection (SFS) as feature selection
generalization ability whereby properly trained data-based method for their proposed TSA algorithm. In another related
model can make accurate stability predictions. The rest of work, Zhang et al. [99] argued that PSO has better
this subsection first described the steps of MLTs approach to optimization ability and better searching efficiency compared
TSA. The subsection also briefly described a review of RL, to other conventional optimization techniques, hence the
DL and DRL approaches to TSA. With regards to MLTs authors deployed PSO in the ELM prediction based TSA
approach, the steps for conventional methodologies are: study.
feature generation, preprocessing and
classification/prediction.
3) CLASSIFICATION/PREDICTION
Usually, MLT based TSA studies in the literature usually
1) FEATURE GENERATION
adopt the ‘offline training, online application’ model [95],
The generation of input vector sets is identified as the first [99] whereby the training model is performed offline and the
and the most important step in establishing reliable TSA TSA testing is done online. In order to minimize
model analysis [88]. Widespread use of phasor measurement misclassification cost, He et al. [100] deployed boosting
units (PMU)- based wide area measurement system (WAMS) algorithms to build the classification model as a weighted
have assisted in the acquisition of synchronized voting of multiple DTs. In another study, Wang et al. [86]
measurements thus allowing the possibility of implementing used core vector machines for its offline training procedure.
advanced wide-area protection, decision making and control In a similar study, Mahdi and Genc [101] used a generated
operations. Synchronously sampled power system variables dataset to train a multilayer perceptron offline before it was
provided by phasor measurement units (PMUs) collected deployed for online TSA testing. In most recent studies in the
before and/or immediately after clearing a fault have been literature, the testing of the various proposed schemes for
utilized as data samples for TSA studies in numerous transient stability studies are conducted using the IEEE 39-
literature [89]-[92]. As the selection of the appropriate bus test system popularly known as the New England test
features is an important criteria for TSA, most studies system [5], [89]-[91], [96]. Various generators and loads
involving MLT approach to TSA usually resort to the variation are usually modelled, with various considered
generation/extraction of these feature sample data through contingencies which include three-phase to ground short-
TDS processes [89], [91], [93], [94]. As explained by Mosavi circuit faults etc. and fault clearing times are varied in cycles.
et al. [85], one major concern in TSA studies is the selection Despite the numerous success that have been achieved using
of proper trajectory features. In numerous studies, different the ‘offline training, online application’ model’, Li et al. [90]
trajectories such as terminal voltage amplitudes, rotor argued that the ‘offline training, online application’ is
variables including angles and rotor speed [91], [95], [96] are inapplicable in the real world as offline generated training
used as predictors to judge whether the system is stable or sets cannot exhibit all the attributes expected of the time-
not. Using a different approach to TDS, He et al. [97] utilized varying modern power system. Furthermore, the authors
pattern identification strategy using dominant instability explained that when a model is not satisfied with some of the
generator grouping based on shortest path algorithm for sample data during the online application, the offline training
selecting input features. Similarly, Zhou et al. [5] used is conducted all over.
bootstrapping technique to generate dataset samples for the

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

Historically, since Sobajic and Pao [102] employed ANN for own experience in the environment. Various power system
critical fault clearing 30 years ago, numerous authors have security and stability devices such as power system stabilizer
deployed varieties of ANN for successful TSA classifications have been modelled based on RL algorithms. Similarly,
[96], [103]. Notably, some other MLTs such as SVM [5], Glavic [111] designed a resistive brake controller that is
[95], KNN [97], DT [92], Bayesian [91], etc. have also been based on RL algorithm. According to the authors in [47], the
successfully deployed for TSA classifications in the architectural structure of the application of RL in power
literature. Owing to its capability to map nonlinear system stability and control studies involves two stages
relationship between inputs and outputs, Lin [103] claimed namely: learning and execution. While the learning stage
that neural networks are superb classifiers for transient refers to the usual RL implementation, the execution stage
stability studies as their outcomes can be continuous such deploy the knowledge acquired from the learning stage for
that the margins and boundaries for transient stability can decision making. As TSA crisis can be considered as a wide-
appear smoother. Conversely, He et al.[100] argued that DTs area control systems’ (WAC) crisis, Druet et al. [112]
are excellent choices for building online DSA classifier as investigated the deployment of RL using Monte Carlo
they have good interpretability which makes them well suited control to define the switching control law for tripping
for TSA studies. Similarly, Kamwa et al. [104] also generators in order to avoid loss of synchronism.
explained that DTs are good classifiers as they perform well However, due to scalability challenges, traditional RL
with regards to cluster problems. As a viable alternative, the algorithms struggle especially with regards to large scale
application of SVM for TSA studies have been mentioned power system. As viable alternatives, RL have been
repeatedly in the literature. Supporting the motion that SVM combined with DL to form DRL algorithms which can
are excellent classifiers for TSA, Zhou et al. [5] explained implement varieties of tasks requiring high dimensional raw
that apart from the good prediction guaranteed using SVM, input and policy control [48]. As explained in [113]. the rise
SVM can calculate the “distance” between an instance and of DRL is linked to the evolution of the powerful deep neural
stable boundary, which can be further used to define the networks. The authors in [114]-[116] designed various Wide
confidence index. However, Tian et al. [105] clarified that Area Control (WAC) strategies to boost transient stability
despite the numerous advantages that SVM brings, their using various RL and DRL methods. Using a modified IEEE
deficiency lies with the parameter selection and a wrong 68 bus as test system, Yousefian et al. [116] proposed a
choice will result in poor classification. On the other hand, WAC design using RL and NN, which optimizes the closed-
Random forest based TSA is another area that is showing a loop performance of a wind integrated power grid through
lot of promise as RF models can calculate feature weights Approximate Dynamic Programming (ADP). From the
and sort features in ranking order according to the weights simulation result achieved, the authors were able to provide
[93]. Compared to other learning algorithms, the authors in transient stability index which guarantees the system’s
[106] and [107] explained that ELM has superb convergence to post-fault equilibrium. Similarly, Zhang et al.
generalization performance with quicker learning speed. [89] used RBM to extract trajectory cluster features which as
Also, the authors in [36] used LSTM in their TSA study. set as inputs for a DBN classifier.
As ensemble paradigm are known to present more accurate
classification models compared to single MLTs, Zhou et al. 5) COMPARISON AND DISCUSSIONS OF MLT
[5] and Yuanhang et al. [108] presented ensemble classifiers APPROACH FOR TRANSIENT STABILITY ASSESSMENT
using multiple SVMs. In a related TSA work that involves Table III presents the summary of some recent research
ensemble approach, Xu et al. [109] used an ensemble works that deployed the use of MLTs for TSA. In Table III,
structures of multiple ELMs for TSA training and we present a summarized description of the test system, the
classification. Deploying a different approach. Baltas et al. preprocessing/optimization technique(s) adopted and
[88] presented a comparative study using three different machine learning algorithm(s) that some authors in recent
algorithms (DT, SVM and ANN) with the aim of suggesting literature deployed in TSA studies.
which algorithms is more suited for the deployed data.
D. MLTs for Voltage Stability Assessment
4) REVIEW OF RL AND DRL APPROACHES TO TSA Various blackout events have shown how crucial the
The past few years have seen increasing efforts concentrated management of reactive power and more importantly the
on the application of RL and DRL in various decision monitoring and evaluation of voltage stability status is a key
making and control problems such as power system studies issue for maintaining global stability of modern power
most especially with regards to TSA and emergency control system [117], [118]. Conventional methods of analyzing
[47], [99]. As explained by Yang [110], the modelling of RL voltage stability such as the calculation of the P-V and Q-V
is synonymous to the process of human learning knowledge. curves at selected load buses using numerous numbers of
Thus, RL algorithms do not necessarily require the complete load flows in traditional models have shown worrying
dynamics of an environment in order to learn, they can learn limitations [43].
self-improvement sorely by judging the feedback from its

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

TABLE III COMPARISON OF RECENTLY PROPOSED MLT BASED APPROACHES FOR TSA

Preprocessing and
Test Machine learning
Ref. Optimization Accuracy Brief summary of the work’s methodology
system algorithm
Techniques achieved
[5] New Min-max Ensemble of 100% Feature generation using bootstrap sampling. 12900 samples containing
England normalization SVM 6739 stable and 6161 unstable sample generated. Min-Max
68 bus normalization as preprocessing method. 5-fold cross validation of the
ensemble classifiers involving multiple SVM.
[94] New KFRS, Memeric ELM 95.2% PMU data used. Feature selection was done using KFRS and memeric
England algorithm algorithm. The algorithm was tested using New England 39-bus system
39 bus and the Southern Hubei province power system.

[85] IEEE 24 - TSVMNN 86.27% Size and type independent features trajectory was used as dataset. The
bus features were used to generate 414 samples for training and testing. The
deep neural network based TSVMNN was used to eliminate
computational complexity of kernel tricks. Contingency samples of
SAVNW, IEEE 24 and Brazilian test system were used for training
while new England 39-bus system was used for testing.
[91] New - Bayesian 98.18% PMU data used. TDS simulation was performed to generate feature
England Multiple Kernel samples. Polynomial and Gaussian Kernel functions are used and
39 bus learning compared. Results showed that the proposed method outperformed
other ML models.
[90] New BinJaya FS OS/ELM 98.24% Post fault PMU data generated for study. Angle modulation for binary
England jaya feature selection deployed. The feature reduction reduced the
39 bus dataset to one third of initial volume. Sigmoid function was used as
and activation function for the classifier.
Liaoning
province
system
[97] West East - KNN 90% Deployed a steady state information such as topology and operating
Lines states as features. DIGG strategy was done to select input features
based on shortest path algorithm. The model has an advantage of
simple training process. KNN was used for clustering analysis.
[92] New - DT 95.1% PMU dataset involving post fault disturbance used. DT was used for
England classification. Results showed the robustness of the model
39 bus
[89] New RBM DBN 98.83% PMU data used. 27 trajectory cluster features were extracted using
England RBM and they are used as model input. 4-fold validation was used
39 bus during evaluation. Sigmoid activation function used. 4 layer of DBN
deployed. Results showed the robustness of the model

Sajan et al. [119] and Velayati et al. [117] explained that, Similar to TSA, the input vectors are usually used to train the
apart from the fact that some of the traditional methods classifier’s algorithms mostly through offline training and the
disregard the dynamic behaviors of modern power system, output vector is the Voltage Stability Margin Index (VSMI)
the methods require comparatively large computations [120] [119], [120], [122]. Diao et al. [123] pre-trained a DT
and they do not provide a detailed practical information on algorithm offline using a voltage security analysis conducted
the stability problems [43]. In recent years, the use of MLTs using the past representative and forecasted daily operating
have been identified as a promising alternative for conditions that involves 29 different scenarios for an
overcoming the various shortfalls of the conventional voltage American Electric Power (AEP) test system. In the study,
stability methods [118], [121]. The process of MLT approach synchronized feature attributes are obtained in real time using
to VSA is divided into two main stages. PMU devices and compared with the offline thresholds
determined by the DTs to assess stability status. As voltage
1) FEATURE GENERATION AND SELECTION PROCESS stability monitoring models are highly nonlinear complex
As explained by Goh et al. [43], power system can be models with large volume of dataset involved, the need for
classified as being in the voltage stability region if it can feature selection and reduction is highly important.
maintain steady acceptable voltages at all buses in the system Mohammadi and Dehghani [124] explained that the large
under normal operating conditions and after being subjected quantities of power system attributes are not appropriate to
to a disturbance. Conventionally, the typical input vectors for be used directly as classifier’s inputs. Hence, several feature
MLT based VSA are retrieved as voltage phasor from PMUs. extraction methods have been proposed in the literature. In an
effort to boost the efficiency and accuracy of a developed

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

SVM based VSA approach, the authors in [125] deployed prospect of SVM for VSA studies. By using the same
multi-objective optimization for the selection of features for dataset, Sajan et al. [126] compared the results of an
the SVM training. Similarly, the authors in [124] deployed optimized SVM with two modelled neural network and
both PCA and correlation analysis techniques as feature conveniently achieved better result from the optimized SVM
reduction and feature selection technique respectively in their approach. In another comparative study involving SVM,
DT based voltage stability study. In another related study, neural network and Adaptive Neural Fuzzy Inference System
Bahmanyar [6] reduced a developed ANN inputs (ANFIS) models in a load-ability margin estimation study,
significantly using GSO. In a similar study, Xu et al. [121] Suganyadevi and Babulal [128] achieved the best result from
chose RELIEF algorithm as the feature selection technique the developed SVM model.
for a developed NNRW training.
Furthermore, to improve the accuracy, reduce generalization 3) COMPARISON AND DISCUSSIONS OF MLT
error and training time, various optimization techniques have APPROACH FOR VOLTAGE STABILITY ASSESSMENT
been deployed to optimize machine learning parameters in Table IV presents the summary of some recent research
VSA studies. Owing to its excellent performance with works that deployed the use of MLTs for the prediction,
regards to generalization errors, GA has been widely monitoring and analysis of voltage stability assessment.
deployed in the literature for VSA. While Sajan et al. [126] Table IV presents a summarized description of the test
deployed GA to boost the performance of a modelled SVM system employed, the data preprocessing/optimization
parameter, the authors in [120] used GA to improve the technique(s) adopted, machine learning algorithm(s) that
parameter tuning of a modelled ANN. As an alternative to some authors in recent literature deployed in VSA studies.
GA, the author in [119] opted for ICA for tuning the ANN
meta-parameters in their VSA study. In another work, V. RESEARCH GAPS AND FUTURE DIRECTIONS
Jayasankar et al. [8] deployed linear optimization for the Despite the astonishing accomplishments that have been
modelled FFBPN. achieved in the application of MLTs for power system
security and stability studies, a number of challenges still
2) CLASSIFICATION OF VOLTAGE STABILITY MARGIN remain unsolved. The prediction and detection accuracy of
INDEX MLTs are known to depend majorly on the quality and
Prominent machine learning algorithms such as NN [39], quantity of the dataset and test systems employed. However,
[119], [120], [122], [127] and DT [123], [124] have been due to the non-availability and inadequacy of realistic power
deployed in various voltage stability studies most especially system data from real power stations and field devices,
in predicting the voltage stability margin index (VSMI). scholars and researchers have been restricted to the use of
Zhang et al. [41] explained that, by learning from a voltage- simulated datasets, open source datasets and the development
stable database, the nonlinear mapping relationship between of scalable testbeds which have shown inconsistency in
the power system operating parameters (input) and the predictions and classifications. Also, apart from the number
voltage stability margin (output) can be mined and of input dataset, another important factor that is peculiar to
reformulated efficiently using neural networks. Also, machine learning applications is the tuning of the parameters.
Jayasankar et al. [8] described the computation time for The rigorous events performed in tuning the parameters so as
ANN as very small and it gives incredible VSMI accuracy to achieve desired results means the MLT approaches
values. However, ANN have limitations, Zhang et al. [41] requires a high level of expert interaction and they can
argued that traditional neural networks learning algorithms sometimes be time consuming. Furthermore, most articles in
usually suffer from excessive training and the parameter the literature usually assume that PMU data are complete,
tuning can be worrisome, then leading to substandard trustworthy and available for online use. In the real world,
generalization performance. As a viable alternative, ELM the measurements may not always be available due to
have been deployed by some authors as they can learn faster jamming, malfunctioning or even attacks.
and predict VSMI more accurately. Velayati et al. [117] and With regards to MLT approaches to SCADA network
Zhang et al. [41] successfully deployed ELM in classifying intrusion mitigation, most research works typically train the
VSM. In a related study, while deploying NNRW for short developed algorithms using network traffic from open source
term voltage instability problem, Xu et al. [121] argued that datasets which mostly are outdated and are no longer relevant
NNRW are better options compared to traditional NN as they with regards to new cyberattack trends. With the high rate of
have efficient tuning mechanisms which makes them highly newly developed sophisticated cyberattacks being witnessed
suited for VSM classification. As alternatives to various worldwide, creating a solution to an outdated problem can be
neural networks methods, numerous studies in the literature irrelevant.
has successfully deployed DT for VSA studies. Mohammadi
[124] clarified that, with regards to less dataset samples, DT
has simpler splitting rules and performs excellently in online
voltage assessment classification. Another possibility is the

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

TABLE IV COMPARISON OF RECENTLY PROPOSED MLT BASED APPROACHES FOR VOLTAGE STABILITY ASSESSMENT

Preprocessing and
Test Machine learning
Ref. Optimization Accuracy Brief summary of the work’s methodology
system algorithm
Techniques achieved
[43] IEEE 9 - FFBP 99.4% System variable based VSI was used as it requires less computational
bus and time. The 6 line VSI implemented are Lmn index, LQP index, FVSI
IEEE 14 index, VCPI (power), VCPI (Loss) and LCPI index. The classifier
made use of 2 layer FFNN with error back propagation.
[87] WECC - FFBP 90.01% Synthetic data obtained using a simulated power system model was
system used. Three class data containing individual samples of 5078, 2540 and
2529 labelled OC was used to generate a total of 10147 samples was
deployed for classification. 1000 sample each of the individual OC was
taken as testing data. SVM, FFBP and RF was used for classification.
RF gave the best result.

[123] AEP - DT 95.45% PMU data used. 29 Operating Condition (OC) generated which
system involves 163 contingencies to create 9454 sample set. Hourly updated
data DT are initially trained offline using features of past representative and
forecasted 24hour OCs.
[121] New Relief algorithm NNRW 99.2% In a study focusing on STVS. various dynamic load components were
England modelled using PSS/E. PMU data used. Offline training was conducted
39 bus using forecasted OC (a day ahead method) and the corresponding
voltage collapse status for VSI was used for online testing. Relief
algorithm was used to select the dominant 136 features. During online
testing, classifiers are updated using latest OCs. 100 individual NNRW
was used for the ensemble classifier model as it presents fast training
speed.

[118] IEEE 14 -K-Means SVM 99% Instead of using large number of OCs, offline identification of Voltage
&30 bus stability pattern was done using K-Means. Parameter choosing of
online training of SVM was also done offline. Testing was done using
updated SVM. CBGA and 5fold validation was done during
evaluation.
[129] IEEE 30 CBR PFDT 94% IEEE 30 bus used for evaluation. 300 different load variations at each
bus buses and the varying of real and reactive power were considered as
contingencies. Probabilistic fuzzy decision tree and case based
reasoning were used for predicting voltage stability

Furthermore, the dependency on traditional schemes such as Future research work on MLT based approach to power
TDS for feature extractions in MLT based TSA is recognized system security and stability menaces should focus on
as a strong determinant in TSA accuracy. However, being a detailed validation of the approaches using large scale test
conventional method, TDS is well-known for its system which have similar characteristics as modern power
computational complexity especially with regards to large system.
scale power system. Better and faster means of feature
extractions for machine learning approach to stability studies VI. CONCLUSION
can be a focus for future research works. In recent times, power system security and stability has been
Also, the issue of offline training and online testing that is a major concern to all energy stakeholders especially the
peculiar with MLT based stability approach can be a huge operators. Operators must be well equipped in timely
liability. Conventionally, most research works deploy static recognition of potential intrusion, attacks, disturbances, and
post faults power flows for classifier training offline. situational awareness. The deployment of conventional
However, it is well known that modern power system is methods has shown flaws especially in terms of resiliency
robust and they have various measures in place to control, and adaptation to the trends of current and future power
protect and restore the systems especially after suffering a system. To address these problems, this paper presents a
disturbance. Therefore, class imbalance crisis cannot be comprehensive review of the most recent MLT based
overlooked as it is unrealistic to depend on only the static approaches to addressing the dominant power system
post fault data in evaluating the stability of the systems. The menaces: power quality disturbance, SCADA network
training samples generated during offline simulations may vulnerability and threats, transient stability assessment and
not be a relevant representation of the current or future status voltage stability assessment. Unlike many of the previously
of the power system. Hence, deploying the offline training published works, the paper addresses the methodologies
process may inevitably lead to a poor applicability of the applied, the limitations, drawbacks and future directions of
trained model when it is deployed for online TSA testing.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

the most recent trends in MLT applications to power system [3] M. Kalech, "Cyber-attack detection in SCADA systems using temporal
pattern recognition techniques," Comput. Secur., vol. 84, pp. 225-238,
security and stability solutions. 2019.
[4] O. A. Alimi and K. Ouahada, "Security assessment of the smart grid: A
APPENDIX review focusing on the NAN architecture," in 2018 IEEE 7th International
Table V presents all the acronyms used in the paper. Conference on Adaptive Science & Technology (ICAST), 2018, pp. 1-8.
[5] Y. Zhou et al, "A hierarchical method for transient stability prediction
of power systems using the confidence of a SVM-based ensemble
TABLE V classifier," Energies, vol. 9, (10), pp. 778, 2016.
DEFINITIONS OF ALL ACRONYMS MENTIONED IN THE PAPER [6] A. Bahmanyar and A. Karami, "Power system voltage stability
Abbreviations monitoring using artificial neural networks with a reduced set of inputs,"
DBN Deep Belief Network International Journal of Electrical Power & Energy Systems, vol. 58, pp.
DWT Discrete Wavelet Transform 246-256, 2014.
MRA Multiresolution Analysis [7] T. Chakravorti et al, "A new robust kernel ridge regression classifier
WT Wavelet Transform for islanding and power quality disturbances in a multi distributed
CT Curvelet Transform generation based microgrid," Renewable Energy Focus, vol. 28, pp. 78-99,
2019.
HHT Hilbert-Huang Transform
[8] V. Jayasankar, N. Kamaraj and N. Vanaja, "Estimation of voltage
HT Hilbert Transform
stability index for power system employing artificial neural network
ABC Artificial Bee Colony technique and TCSC placement," Neurocomputing, vol. 73, (16-18), pp.
SSA Static Single Assignment 3005-3011, 2010.
FFS Forward Feature Selection [9] C. Sun, A. Hahn and C. Liu, "Cyber security of a power grid: State-of-
GA Genetic Algorithm the-art," International Journal of Electrical Power & Energy Systems, vol.
ELM Extreme Learning Machine 99, pp. 45-56, 2018.
CCA Canonical Correlation Analysis [10] G. Liang et al, "A review of false data injection attacks against
modern power systems," IEEE Transactions on Smart Grid, vol. 8, (4), pp.
ICA Independent Component Analysis
1630-1638, 2016.
TSVMNN Twin convolutional support vector machine [11] R. Deng et al, "False data injection on state estimation in power
SVD Singular value decomposition systems—Attacks, impacts, and defense: A survey," IEEE Transactions on
SOM Self-Organizing Map Neural Network Industrial Informatics, vol. 13, (2), pp. 411-423, 2016.
OCSVM One class support vector machine [12] X. Wang et al, "Distributed detection and isolation of false data
ECOC-SVM Error-correcting output code-support vector machine injection attacks in smart grids via nonlinear unknown input observers,"
SWAT Secure Water Treatment testbed International Journal of Electrical Power & Energy Systems, vol. 110, pp.
LSTM Long term short memory 208-222, 2019.
[13] J. Zhang and X. Wang, "Quickest Detection of Time-Varying False
IWP-CSO Intrusion weighted Particle based Cuckoo Search
Data Injection Attacks in Dynamic Linear Regression Models," arXiv
Optimization
Preprint arXiv:1811.05423, 2018.
BNT Balanced Neural Tree [14] Y. Huang et al, "Defending false data injection attack on smart grid
NNRW Neural Network with Random Weight network using adaptive CUSUM test," in 2011 45th Annual Conference on
WPT Wavelet Packet Transform Information Sciences and Systems, 2011, pp. 1-6.
VMD Variational Mode Decomposition [15] S. Nazir, S. Patel and D. Patel, "Assessing and augmenting SCADA
ST S-Transform cyber security: A survey of techniques," Comput. Secur., vol. 70, pp. 436-
PMU Phasor Measurement Unit 454, 2017.
[16] P. Glauner et al, "The challenge of non-technical loss detection using
EPRI Electrical Power Research Institute
artificial intelligence: A survey," arXiv Preprint arXiv:1606.00626, 2016.
HNA-NN Hierarchical Neuron Architecture Based Neural Network [17] K. S. Perera, Z. Aung and W. L. Woon, "Machine learning techniques
DQN Deep Q Net for supporting renewable energy generation and integration: A survey," in
LR Logistic Regression International Workshop on Data Analytics for Renewable Energy
SVDD Support Vector Data Description Integration, 2014, pp. 81-96.
KPCA Kernel principal component analysis [18] M. K. Saini and R. Kapoor, "Classification of power quality events–a
IoT Internet of Things review," International Journal of Electrical Power & Energy Systems, vol.
43, (1), pp. 11-19, 2012.
CI Critical Infrastructure
[19] P. Kundur, N. J. Balu and M. G. Lauby, Power System Stability and
TDS Time domain simulation Control. McGraw-hill New York, 19947.
DDPG Deep Deterministic Policy Gradients [20] P. Jokar, N. Arianpoo and V. C. Leung, "A survey on security issues
IM Induction Motor in smart grids," Security and Communication Networks, vol. 9, (3), pp.
EMD Empirical mode decomposition 262-273, 2016.
FT Fourier Transform [21] G. Wu, J. Sun and J. Chen, "A survey on the security of cyber-
SARSA State–Action–Reward–State–Action physical systems," Control Theory and Technology, vol. 14, (1), pp. 2-10,
FFBP FeedForward BackPropagation 2016.
[22] S. Wang and H. Chen, "A novel deep learning method for the
classification of power quality disturbances using deep convolutional
neural network," Appl. Energy, vol. 235, pp. 1126-1140, 2019.
REFERENCES [23] S. Khokhar et al, "A new optimal feature selection algorithm for
[1] G. N. Baltas et al, "A novel ensemble approach for solving the classification of power quality disturbances using discrete wavelet
transient stability classification problem," in 2018 7th International transform and probabilistic neural network," Measurement, vol. 95, pp.
Conference on Renewable Energy Research and Applications (ICRERA), 246-259, 2017.
2018, pp. 1282-1286. [24] M. Lopez-Ramirez et al, "EMD-based feature extraction for power
[2] O. A. Alimi, K. Ouahada and A. M. Abu-Mahfouz, "Real time security quality disturbance classification using moments," Energies, vol. 9, (7),
assessment of the power system using a hybrid support vector machine and pp. 565, 2016.
multilayer perceptron neural network algorithms," Sustainability, vol. 11,
(13), pp. 3586, 2019.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

[25] M. Valtierra-Rodriguez et al, "Detection and classification of single [47] M. Glavic, R. Fonteneau and D. Ernst, "Reinforcement learning for
and combined power quality disturbances using neural networks," IEEE electric power system decision and control: Past considerations and
Trans. Ind. Electron., vol. 61, (5), pp. 2473-2482, 2013. perspectives," IFAC-PapersOnLine, vol. 50, (1), pp. 6918-6927, 2017.
[26] M. Teixeira et al, "SCADA system testbed for cybersecurity research [48] D. Zhang, X. Han and C. Deng, "Review on the research and practice
using machine learning approach," Future Internet, vol. 10, (8), pp. 76, of deep learning and reinforcement learning in smart grids," CSEE Journal
2018. of Power and Energy Systems, vol. 4, (3), pp. 362-370, 2018.
[27] W. Gao et al, "On SCADA control system command and response [49] A. A. Abdoos, P. K. Mianaei and M. R. Ghadikolaei, "Combined
injection and intrusion detection," in 2010 eCrime Researchers Summit, VMD-SVM based feature selection method for classification of power
2010, pp. 1-9. quality events," Applied Soft Computing, vol. 38, pp. 637-646, 2016.
[28] R. L. Perez et al, "Machine learning for reliable network attack [50] F. Jandan et al, "Recognition and classification of power quality
detection in SCADA systems," in 2018 17th IEEE International disturbances by DWT-MRA and SVM classifier," International Journal of
Conference on Trust, Security and Privacy in Computing and Advanced Computer Science and Applications, vol. 10, (3), pp. 368-377,
Communications/12th IEEE International Conference on Big Data Science 2019.
and Engineering (TrustCom/BigDataSE), 2018, pp. 633-638. [51] R. Ahila, V. Sadasivam and K. Manimala, "An integrated PSO for
[29] E. Chikuni and M. Dondo, "Investigating the security of electrical parameter determination and feature selection of ELM and its application
power systems SCADA," in Africon 2007, 2007, pp. 1-7. in classification of power system disturbances," Applied Soft Computing,
[30] C. Feng, T. Li and D. Chana, "Multi-level anomaly detection in vol. 32, pp. 23-37, 2015.
industrial control systems via package signatures and lstm networks," in [52] H. Liu et al, "Complex power quality disturbances classification via
2017 47th Annual IEEE/IFIP International Conference on Dependable curvelet transform and deep learning," Electr. Power Syst. Res., vol. 163,
Systems and Networks (DSN), 2017, pp. 261-272. pp. 1-9, 2018.
[31] S. Parthasarathy and D. Kundur, "Bloom filter based intrusion [53] J. Bosnic et al, "Power quality disturbance classification based on
detection for smart grid SCADA," in 2012 25th IEEE Canadian wavelet transform and support vector machine," in 2017 11th International
Conference on Electrical and Computer Engineering (CCECE), 2012, pp. Conference on Measurement, 2017, pp. 9-13.
1-6. [54] K. Manimala, K. Selvi and R. Ahila, "Optimization techniques for
[32] J. Jiang and L. Yasakethu, "Anomaly detection via one class svm for improving power quality data mining using wavelet packet based support
protection of scada systems," in 2013 International Conference on Cyber- vector machine," Neurocomputing, vol. 77, (1), pp. 36-47, 2012.
Enabled Distributed Computing and Knowledge Discovery, 2013, pp. 82- [55] Z. Liu, Y. Cui and W. Li, "A classification method for complex power
88. quality disturbances using EEMD and rank wavelet SVM," IEEE
[33] P. Nader, P. Honeine and P. Beauseroy, "Intrusion detection in Transactions on Smart Grid, vol. 6, (4), pp. 1678-1685, 2015.
SCADA systems using one-class classification," in 21st European Signal [56] U. Singh and S. N. Singh, "Application of fractional Fourier
Processing Conference (EUSIPCO 2013), 2013, pp. 1-5. transform for classification of power quality disturbances," IET Science,
[34] R. M. Lee, M. J. Assante and T. Conway, "TLP: White Analysis of Measurement & Technology, vol. 11, (1), pp. 67-76, 2017.
the Cyber Attack on the Ukrainian Power Grid," E-ISAC, Tech.Rep, 2016. [57] Z. Liquan, G. Meijiao and W. Lin, "Classification of multiple power
[35] A. Sharifian and S. Sharifian, "A new power system transient stability quality disturbances based on the improved SVM," in 2017 International
assessment method based on Type-2 fuzzy neural network estimation," Conference on Wireless Communications, Signal Processing and
International Journal of Electrical Power & Energy Systems, vol. 64, pp. Networking (WiSPNET), 2017, pp. 2625-2628.
71-87, 2015. [58] C. Chunling et al, "Power quality disturbances classification based on
[36] J. James et al, "Intelligent time-adaptive transient stability assessment multi-class classification SVM," in 2009 2nd International Conference on
system," IEEE Trans. Power Syst., vol. 33, (1), pp. 1049-1058, 2017. Power Electronics and Intelligent Transportation System (PEITS), 2009,
[37] N. Amjady and S. F. Majedi, "Transient stability prediction by a pp. 290-294.
hybrid intelligent system," IEEE Trans. Power Syst., vol. 22, (3), pp. 1275- [59] D. De Yong, S. Bhowmik and F. Magnago, "An effective power
1283, 2007. quality classifier using wavelet transform and support vector machines,"
[38] J. James et al, "Delay aware intelligent transient stability assessment Expert Syst. Appl., vol. 42, (15-16), pp. 6075-6081, 2015.
system," IEEE Access, vol. 5, pp. 17230-17239, 2017. [60] S. Alshahrani et al, "Evaluation and classification of power quality
[39] S. Suliman and T. A. Rahman, "Artificial immune system based disturbances based on discrete wavelet transform and artificial neural
machine learning for voltage stability prediction in power system," in 2010 networks," in 2015 50th International Universities Power Engineering
4th International Power Engineering and Optimization Conference Conference (UPEC), 2015, pp. 1-5.
(PEOCO), 2010, pp. 53-58. [61] C. Naik et al, "Classification of power quality events using wavelet
[40] Z. Nie et al, "A PMU-based voltage security assessment framework packet transform and extreme learning machine," in 2016 IEEE 2nd
using hoeffding-tree-based learning," in 2017 19th International Annual Southern Power Electronics Conference (SPEC), 2016, pp. 1-6.
Conference on Intelligent System Application to Power Systems (ISAP), [62] H. Erişti and Y. Demir, "A new algorithm for automatic classification
2017, pp. 1-6. of power quality events based on wavelet transform and SVM," Expert
[41] R. Zhang et al, "Voltage stability margin prediction by ensemble Syst. Appl., vol. 37, (6), pp. 4094-4102, 2010.
based extreme learning machine," in 2013 IEEE Power & Energy Society [63] P. Kanirajan and V. S. Kumar, "Power quality disturbance detection
General Meeting, 2013, pp. 1-5. and classification using wavelet and RBFNN," Applied Soft Computing,
[42] O. B. Adewuyi et al, "Static voltage stability improvement with vol. 35, pp. 470-481, 2015.
battery energy storage considering optimal control of active and reactive [64] B. Biswal et al, "Automatic classification of power quality events
power injection," Electr. Power Syst. Res., vol. 172, pp. 303-312, 2019. using balanced neural tree," IEEE Trans. Ind. Electron., vol. 61, (1), pp.
[43] H. H. Goh et al, "Evaluation for voltage stability indices in power 521-530, 2013.
system using artificial neural network," Procedia Engineering, vol. 118, [65] U. Singh and S. N. Singh, "A new optimal feature selection scheme
pp. 1127-1136, 2015. for classification of power quality disturbances based on ant colony
[44] S. Johansson, Long-Term Voltage Stability in Power Systems- framework," Applied Soft Computing, vol. 74, pp. 216-225, 2019.
Alleviating the Impact of Generator Current Limiters. Chalmers [66] M. Biswal and P. K. Dash, "Detection and characterization of
University of Technology, 1998. multiple power quality disturbances with a fast S-transform and decision
[45] J. Gao et al, "Omni SCADA Intrusion Detection Using Deep tree based classifier," Digital Signal Processing, vol. 23, (4), pp. 1071-
Learning Algorithms," arXiv Preprint arXiv:1908.01974, 2019. 1083, 2013.
[46] S. K. Sharma and X. Wang, "Towards massive machine type [67] T. Jayasree, D. Devaraj and R. Sukanesh, "Power quality disturbance
communications in ultra-dense cellular IoT networks: Current issues and classification using Hilbert transform and RBF networks,"
machine learning-assisted solutions," IEEE Communications Surveys & Neurocomputing, vol. 73, (7-9), pp. 1451-1456, 2010.
Tutorials, 2019. [68] R. Kumar, B. Singh and D. T. Shahani, "Recognition of single-stage
and multiple power quality events using Hilbert–Huang transform and

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

probabilistic neural network," Electric Power Components and Systems, [91] X. Gu and Y. Li, "Bayesian multiple kernels learning-based transient
vol. 43, (6), pp. 607-619, 2015. stability assessment of power systems using synchronized measurements,"
[69] N. Huang et al, "Power quality disturbances classification based on S- in 2013 IEEE Power & Energy Society General Meeting, 2013, pp. 1-5.
transform and probabilistic neural network," Neurocomputing, vol. 98, pp. [92] Y. Yang et al, "Measurement-based cell-DT method for power system
12-23, 2012. transient stability classification," CSEE Journal of Power and Energy
[70] S. Shukla, S. Mishra and B. Singh, "Empirical-mode decomposition Systems, vol. 3, (3), pp. 278-285, 2017.
with Hilbert transform for power-quality assessment," IEEE Trans. Power [93] C. Zhang et al, "Feature selection of power system transient stability
Del., vol. 24, (4), pp. 2159-2165, 2009. assessment based on random forest and recursive feature elimination," in
[71] L. Zhou et al, "Automatic fine-grained access control in SCADA by 2016 IEEE PES Asia-Pacific Power and Energy Engineering Conference
machine learning," Future Generation Comput. Syst., vol. 93, pp. 548-559, (APPEEC), 2016, pp. 1264-1268.
2019. [94] Y. Li, G. Li and Z. Wang, "Rule extraction based on extreme learning
[72] H. Yang, L. Cheng and M. C. Chuah, "Deep-learning-based network machine and an improved ant-miner algorithm for transient stability
intrusion detection for SCADA systems," in 2019 IEEE Conference on assessment," PloS One, vol. 10, (6), pp. e0130814, 2015.
Communications and Network Security (CNS), 2019, pp. 1-7. [95] H. Wei et al, "Real-time transient stability assessment in power
[73] A. Mansouri, B. Majidi and A. Shamisa, "Anomaly detection in system based on improved SVM," Journal of Modern Power Systems and
industrial control systems using evolutionary-based optimization of neural Clean Energy, vol. 7, (1), pp. 26-37, 2019.
networks," 2017. [96] E. A. Frimpong, P. Y. Okyere and J. Asumadu, "On-line
[74] I. Ullah and Q. H. Mahmoud, "A hybrid model for anomaly-based determination of transient stability status using MLPNN," in 2017 IEEE
intrusion detection in SCADA networks," in 2017 IEEE International PES PowerAfrica, 2017, pp. 23-27.
Conference on Big Data (Big Data), 2017, pp. 2160-2167. [97] C. He, L. Guan and W. Mo, "A method for transient stability
[75] H. Wang et al, "Hierarchical Online Intrusion Detection for SCADA assessment based on pattern recognition," in 2016 International
Networks," arXiv Preprint arXiv:1611.09418, 2016. Conference on Smart Grid and Clean Energy Technologies (ICSGCE),
[76] S. Shitharth, "An enhanced optimization based algorithm for intrusion 2016, pp. 343-347.
detection in SCADA network," Comput. Secur., vol. 70, pp. 16-26, 2017. [98] S. Kalyani and K. S. Swarup, "Binary svm approach for security
[77] L. A. Maglaras and J. Jiang, "Intrusion detection in SCADA systems assessment and classification in power systems," in 2009 Annual IEEE
using machine learning techniques," in 2014 Science and Information India Conference, 2009, pp. 1-4.
Conference, 2014, pp. 626-631. [99] Y. Zhang et al, "Optimized extreme learning machine for power
[78] I. A. Khan et al, "HML-IDS: A Hybrid-Multilevel Anomaly system transient stability prediction using synchrophasors," Mathematical
Prediction Approach for Intrusion Detection in SCADA Systems," IEEE Problems in Engineering, vol. 2015, 2015.
Access, vol. 7, pp. 89507-89521, 2019. [100] M. He, J. Zhang and V. Vittal, "Robust online dynamic security
[79] A. Almalawi et al, "An unsupervised anomaly-based detection assessment using adaptive ensemble decision-tree learning," IEEE Trans.
approach for integrity attacks on SCADA systems," Comput. Secur., vol. Power Syst., vol. 28, (4), pp. 4089-4098, 2013.
46, pp. 94-110, 2014. [101] M. Mahdi and V. I. Genc, "Artificial neural network based algorithm
[80] L. A. Maglaras and J. Jiang, "Ocsvm model combined with k-means for early prediction of transient stability using wide area measurements,"
recursive clustering for intrusion detection in scada systems," in 10th in 2017 5th International Istanbul Smart Grid and Cities Congress and
International Conference on Heterogeneous Networking for Quality, Fair (ICSG), 2017, pp. 17-21.
Reliability, Security and Robustness, 2014, pp. 133-134. [102] D. J. Sobajic and Y. Pao, "Artificial neural-net based dynamic
[81] E. G. da Silva et al, "A one-class NIDS for SDN-based SCADA security assessment for electric power systems," IEEE Trans. Power Syst.,
systems," in 2016 IEEE 40th Annual Computer Software and Applications vol. 4, (1), pp. 220-228, 1989.
Conference (COMPSAC), 2016, pp. 303-312. [103] Y. Lin, "Comparison of CART-and MLP-based power system
[82] M. Kravchik and A. Shabtai, "Detecting cyber-attacks in industrial transient stability preventive control," International Journal of Electrical
control systems using convolutional neural networks," in Proceedings of Power & Energy Systems, vol. 45, (1), pp. 129-136, 2013.
the 2018 Workshop on Cyber-Physical Systems Security and PrivaCy, [104] I. Kamwa, S. Samantaray and G. Joos, "Catastrophe predictors from
2018, pp. 72-83. ensemble decision-tree learning of wide-area severity indices," IEEE
[83] J. M. Beaver, R. C. Borges-Hink and M. A. Buckner, "An evaluation Transactions on Smart Grid, vol. 1, (2), pp. 144-158, 2010.
of machine learning methods to detect malicious SCADA [105] F. Tian et al, "A preventive transient stability control method based
communications," in 2013 12th International Conference on Machine on support vector machine," Electr. Power Syst. Res., vol. 170, pp. 286-
Learning and Applications, 2013, pp. 54-59. 293, 2019.
[84] H. Qu et al, "Intrusion detection in SCADA/modbus network based [106] Y. Li and X. Gu, "Power system transient stability assessment based
on machine learning," in International Conference on Machine Learning on online sequential extreme learning machine," in 2013 IEEE PES Asia-
and Intelligent Communications, 2017, pp. 437-454. Pacific Power and Energy Engineering Conference (APPEEC), 2013, pp.
[85] A. B. Mosavi, A. Amiri and H. Hosseini, "A learning framework for 1-4.
size and type independent transient stability prediction of power system [107] I. B. Sulistiawati et al, "Critical clearing time prediction within
using twin convolutional support vector machine," IEEE Access, vol. 6, various loads for transient stability assessment by means of the extreme
pp. 69937-69947, 2018. learning machine method," International Journal of Electrical Power &
[86] B. Wang et al, "Power system transient stability assessment based on Energy Systems, vol. 77, pp. 345-352, 2016.
big data and the core vector machine," IEEE Transactions on Smart Grid, [108] D. Yuanhang et al, "Multi-support vector machine power system
vol. 7, (5), pp. 2561-2570, 2016. transient stability assessment based on relief algorithm," in 2015 IEEE
[87] V. Malbasa et al, "Voltage stability prediction using active machine PES Asia-Pacific Power and Energy Engineering Conference (APPEEC),
learning," IEEE Transactions on Smart Grid, vol. 8, (6), pp. 3117-3124, 2015, pp. 1-5.
2017. [109] Y. Xu et al, "A reliable intelligent system for real-time dynamic
[88] N. G. Baltas et al, "A comparative analysis of decision trees, support security assessment of power systems," IEEE Trans. Power Syst., vol. 27,
vector machines and artificial neural networks for on-line transient (3), pp. 1253-1263, 2012.
stability assessment," in 2018 International Conference on Smart Energy [110] T. Yang et al, "Reinforcement learning in sustainable energy and
Systems and Technologies (SEST), 2018, pp. 1-6. electric systems: a survey," Annual Reviews in Control, 2020.
[89] R. Zhang et al, "Transient stability prediction of power systems based [111] M. Glavic, "Design of a resistive brake controller for power system
on deep belief networks," in 2018 2nd IEEE Conference on Energy stability enhancement using reinforcement learning," IEEE Trans. Control
Internet and Energy System Integration (EI2), 2018, pp. 1-6. Syst. Technol., vol. 13, (5), pp. 743-751, 2005.
[90] Y. Li and Z. Yang, "Application of EOS-ELM with binary Jaya-based [112] C. Druet, D. Ernst and L. Wehenkel, "Application of reinforcement
feature selection to real-time transient stability assessment using PMU learning to electrical power system closed-loop emergency control," in
data," IEEE Access, vol. 5, pp. 23092-23101, 2017. European Conference on Principles of Data Mining and Knowledge
Discovery, 2000, pp. 86-95.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3003568, IEEE Access

[113] M. Glavic, "(Deep) Reinforcement learning for electric power member of remarkable professional organizations, which include the South
system control and related problems: A short review and perspectives," African Institute of Electrical Engineers (SAIEE) and Engineering Council
Annual Reviews in Control, 2019. of South Africa (ECSA). He has published some papers in reputable
[114] R. Yousefian and S. Kamalasadan, "Energy function inspired value journals and conference proceedings. He is currently undergoing his
priority based global wide-area control of power grid," IEEE Transactions doctoral program at the University of Johannesburg. His current research
on Smart Grid, vol. 9, (2), pp. 552-563, 2016. interests include power systems security, energy management and
[115] R. Yousefian and S. Kamalasadan, "A Lyapunov function based optimization, smart grid, machine learning and data science, data privacy
optimal hybrid power system controller for improved transient stability," and security.
Electr. Power Syst. Res., vol. 137, pp. 6-15, 2016.
[116] R. Yousefian, R. Bhattarai and S. Kamalasadan, "Transient stability
enhancement of power grid with integrated wide area control of wind
farms and synchronous generators," IEEE Trans. Power Syst., vol. 32, (6),
pp. 4818-4831, 2017.
[117] M. H. Velayati, N. Amjady and I. Khajevandi, "Prediction of
dynamic voltage stability status based on Hopf and limit induced
bifurcations using extreme learning machine," International Journal of
Electrical Power & Energy Systems, vol. 69, pp. 150-159, 2015.
[118] S. Pérez-Londoño, G. Olivar-Tost and J. Mora-Florez, "Online
determination of voltage stability weak areas for situational awareness
improvement," Electr. Power Syst. Res., vol. 145, pp. 112-121, 2017. Khmaies Ouahada is a Professor at the
[119] K. Sajan, V. Kumar and B. Tyagi, "ICA based artificial neural University of Johannesburg, South Africa, where he received his M. Eng.,
network model for voltage stability monitoring," in TENCON 2015-2015 with distinction, and his D. Eng. Degrees in 2002 and 2009 respectively.
IEEE Region 10 Conference, 2015, pp. 1-3. He has received his B. Eng. Degree in 1995 from the University of
[120] K. Sajan, B. Tyagi and V. Kumar, "Genetic algorithm based artificial Khartoum in Sudan. He has worked at Sudatel, the Sudanese National
neural network model for voltage stability monitoring," in 2014 Eighteenth Communications company. He is the founder and Chairman of the centre
National Power Systems Conference (NPSC), 2014, pp. 1-5. for smart communications systems at the faculty of Engineering and the
[121] Y. Xu et al, "Assessing short-term voltage stability of electric power Built Environment, University of Johannesburg. Prof Khmaies Ouahada is
systems by a hierarchical intelligent system," IEEE Transactions on a rated researcher at the National Research Foundation of South Africa. He
Neural Networks and Learning Systems, vol. 27, (8), pp. 1686-1696, 2015. is a senior member of the IEEE Information Theory and Communications
[122] S. Kamalasadan, D. Thukaram and A. Srivastava, "A new intelligent societies. He is also a member of the IEEE South Africa Information
algorithm for online voltage stability assessment and monitoring," Theory Society Chapter. He is a senior member of the SAIEE society.
International Journal of Electrical Power & Energy Systems, vol. 31, (2- His research interests are information theory, coding techniques, power-
3), pp. 100-110, 2009. line communications, visible light communications, smart grid, energy
[123] R. Diao et al, "Decision tree-based online voltage security demand management, renewable energy, wireless sensor networks, reverse
assessment using PMU measurements," IEEE Trans. Power Syst., vol. 24, engineering and engineering education.
(2), pp. 832-839, 2009.
[124] H. Mohammadi and M. Dehghani, "PMU based voltage security
assessment of power systems exploiting principal component analysis and
decision trees," International Journal of Electrical Power & Energy
Systems, vol. 64, pp. 655-663, 2015.
[125] H. Mohammadi et al, "Voltage stability assessment using multi-
objective biogeography-based subset selection," International Journal of
Electrical Power & Energy Systems, vol. 103, pp. 525-536, 2018.
[126] K. Sajan, V. Kumar and B. Tyagi, "Genetic algorithm based support
vector machine for on-line voltage stability monitoring," International
Journal of Electrical Power & Energy Systems, vol. 73, pp. 200-208, Adnan M. Abu-Mahfouz (M12SM17)
2015. received his MEng and PhD degrees in computer engineering from the
[127] A. Maiorano and M. Trovato, "A neural network-based tool for University of Pretoria. He is currently a Principal Researcher at the
preventive control of voltage stability in multi-area power systems," Council for Scientific and Industrial Research (CSIR), Professor
Neurocomputing, vol. 23, (1-3), pp. 161-176, 1998. Extraordinaire at Tshwane University of Technology, Visiting Professor at
[128] M. Suganyadevi and C. Babulal, "Fast Assessment of Voltage University of Johannesburg and Extraordinary faculty member at
Stability Margin of a Power System." Journal of Electrical Systems, vol. University of Pretoria. His research interests are wireless sensor and
10, (3), 2014. actuator network, low power wide area networks, software defined
[129] S. Nandanwar et al, "Voltage Security Assessment by Using PFDT wireless sensor network, cognitive radio, network security, network
and CBR Methods in Emerging Power System," Energy Procedia, vol. management, sensor/actuator node development. He is an associate editor
144, pp. 170-181, 2018. at IEEE Access, IEEE Internet of Things and IEEE Transaction on
Industrial Informatics, Senior Member of the IEEE and Member of many
IEEE Technical Communities. He participated in the formulation of many
large and multidisciplinary R&D successful proposals (as Principal
Investigator or main author/contributor). He is the founder of the Smart
Networks collaboration initiative that aims to develop efficient and secure
networks for the future smart systems, such as smart cities, smart grid and
smart water grid.

Oyeniyi Akeem ALIMI received his Master of

Technology degree in Electrical Engineering from the University of
Johannesburg, South Africa in 2017. Previously, he obtained Bachelor of
Technology (Honours) degrees in Electronics and Electrical Engineering
from Ladoke Akintola University of Technology, Nigeria in 2011. He is a

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.