CSEE JOURNAL OF POWER AND ENERGY SYSTEMS, VOL. 8, NO.

4, JULY 2022 1145

Adaptive Fault Detection Scheme Using an

Optimized Self-healing Ensemble Machine
Learning Algorithm
Levent Yavuz, Student Member, IEEE, Ahmet Soran, Member, IEEE, Ahmet Önen , Member, IEEE,
Xiangjun Li, Senior Member, IEEE, and S. M. Muyeen, Senior Member, IEEE

Abstract—This paper proposes a new cost-efficient, adaptive, investment areas, such as Microgrids or Virtual Power Plants,
and self-healing algorithm in real time that detects faults in a that surely attract people looking for possible business oppor-
short period with high accuracy, even in the situations when it is tunities.
difficult to detect. Rather than using traditional machine learning
(ML) algorithms or hybrid signal processing techniques, a new However, using a set of different renewable systems needs
framework based on an optimization enabled weighted ensemble to be coordinated and/or controlled by small units, such
method is developed that combines essential ML algorithms. as hardware supported Microgrids or pure communication
In the proposed method, the system will select and compound enabled software-based Virtual Power Plants. A microgrid,
appropriate ML algorithms based on Particle Swarm Optimiza- a localized remotely controllable and self-operating group of
tion (PSO) weights. For this purpose, power system failures are
simulated by using the PSCAD-Python co-simulation. One of energy sources, uses a type of controller that may operate in
the salient features of this study is that the proposed solution grid-connected or islanded modes to dispatch energy [1].
works on real-time raw data without using any pre-computational More importantly, increasing more renewable penetrations
techniques or pre-stored information. Therefore, the proposed reduces rotating inertia. For example, the total inertia cur-
technique will be able to work on different systems, topologies, rently available in the South Australia network is around
or data collections. The proposed fault detection technique is
validated by using PSCAD-Python co-simulation on a modified 16,200 MWs. In 2017, without the Northern Power Station
and standard IEEE-14 and standard IEEE-39 bus considering and Torrens Island “A” Power Station, the total available inertia
network faults which are difficult to detect. would reduce to around 10,000 MWs [2] due to the inclusions
of renewable sources, which is very alarming. This may cause
Index Terms—Decision tree (DT), ensemble machine learning system instability or even a blackout, in case of catastrophic
algorithm, fault detection, islanding operation, k-Nearest
events or grid fault conditions [3], [4] and faults are needed
Neighbor (kNN), linear discriminant analysis (LDA), logistic
regression (LR), Naı̈ve Bayes (NB), self-healing algorithm. to be cleared quickly. Therefore, special precautions must be
taken to avoid any greater outage and/or cascading blackouts.
Considering these, many fault detection techniques have been
developed recently and reported in power system literature [5],
I. I NTRODUCTION [6].
In the traditional power system, primarily the protective

T HE rapid growth of energy consumption throughout the

world brings new technologies together in power sys-
tems. In addition to traditional systems for energy generation,
relay receives the voltage, current, and frequency information
from transmission/distribution systems using instrumentation
transformers. That information is typically processed by the
such smaller on-site power generating platforms, (e.g., wind protective relay to take action in case of emergency or abnor-
turbines or other renewable energy systems), have emerged mal conditions for the desired tripping time. The relay logic
recently. Distributed Energy Resources (DER) arouses new algorithm takes the decision of whether to trip open or to
Manuscript received August 2, 2020; revised May 3, 2021; accepted July close the circuit breaker. In the distribution system, the pro-
21, 2021. Date of online publication September 10, 2021; date of current tective relays require a large fault current to detect the faulty
version October 11, 2021. condition. This is problematic for the modern distribution
L. Yavuz and A. Soran are with the Electrical and Computer Engineering
Department in Abdullah Gül University, Kayseri 38080, Turkey. system which connects various Distributed Generations (DGs)
A. Önen (corresponding author, e-mail: [email protected]; ORCID: ht including renewable sources, microgrids, and virtual power
tps://orcid.org/0000-0001-7086-5112) is with Department of Electrical and plants. Most of the DGs, nowadays, are equipped with inverter
Computer Engineering, College of Engineering, Sultan Qaboos University,
Al-Khoud, Muscat, 123, Oman. technology which typically contributes at two times of per unit
S. M. Muyeen is with the Department of Electrical Engineering, Qatar rated current, as rule of thumb [7]. This is troublesome because
University, Doha 2713, Qatar. the current level may not be sufficient to trigger the relaying
X. J. Li is with State Key Laboratory of Control and Operation of
Renewable Energy and Storage Systems, China Electric Power Research action and it puts the inverter equipped modern distribution
Institute, Beijing 100192, China. system at risk.
DOI: 10.17775/CSEEJPES.2020.03760 Many fault detection mechanisms have been depicted in the
2096-0042 © 2020 CSEE
1146 CSEE JOURNAL OF POWER AND ENERGY SYSTEMS, VOL. 8, NO. 4, JULY 2022

power system. The real-time data acquisition device like Fre- for the static power system topologies that do not change
quency Disturbance Recorder (FDR), optimally placed Phasor frequently. In the case of modifying the network by
Measurement Units (PMUs), and many other equipments are adding/dropping a line or adding new buses, most of the
used to detect power system faults [8]. Line outage detection proposed algorithms may fail.
using phasor angle measurements are reported in [9]. Fault de- In order to create a new intelligence-based fault detection
tection occurred in DC and AC microgrids are reported in [10]. method, we consider all these metrics. Different solutions
One of the methods could be using additional equipment(s) use only one machine-learning algorithm to reach the time
to build a protection layer for renewable source integrated goal or use complex ones that merge multi-machine learning
microgrid; however, this requires additional hardware cost [5]. algorithms to obtain higher accuracy results.
Besides, the faulty equipment is required to be fixed or The proposed method combines different ML algorithms
replaced in the case of equipment failures, which adds more which are called weak-learners. Particle Swarm Optimization
cost and labor. It is noted that equipment failures are frequent (PSO) is used for this combination process in this study.
events and blackouts are mostly triggered by these problems. Optimum weights are calculated and given a set of weak-
Apparently, hardware-based applications seem not scalable learners in order to reach higher performance together. By
and they are infeasible due to financial problems. Thus, many using this approach, each weak-learners weakest part is closed
new intelligent computer-aided prediction techniques have by others. Behind the power of the proposed method is the
been developed to deal with fault scanning [11]–[13]. ensebmling technique. The main idea is to calculate the almost
In recent years, many soft computing and data-driven ap- ideal weights for each machine learning algorithm periodically
proaches have been investigated to detect power system faults. and then combine them with soft voting during the test
The performance of these techniques can be measured by conditions. With this adaptive technique, we improve both the
(i) the rate of their accuracy, (ii) detection time, (iii) cost- accuracy rate and time while only increasing training time by
efficiency, and (iv) flexibility/adaptivity. a little bit.
i. There are various methods proposed to achieve the highest Signal processing approaches, such as decision tree and
accuracy rates, i.e., deciding islanding operation cor- random forest, are extensively utilized to detect the fault in
rectly. Primarily, hybrid techniques [14], applying suit- distribution systems connecting microgrids [23]. There are also
able threshold settings [15], and signal processing tech- some other signal processing approaches: Fourier Transform,
niques [16], [17] have been proposed. Traditional fault Wavelet Transform, S-Transform, TT- Transform, and Hilbert
detection systems, signal processing methods, threshold Huang Transform techniques have successfully been applied to
settings, and their deficiencies are detailed in Section II. detect faulty conditions [4], [24], [25]. Data-driven approaches
ii. The proposed methods should be fast enough for detecting have also received attention by power system researchers
the faults over the network. There is a tradeoff between in identifying faults [26]. The Machine learning technique
accuracy and detectiontime. In the experiments, we were requires the computer/system learn various patterns based on
challenged by IEEE 1547 standards [18], which require the given input and hence it is found applicible for power
the detection of possible faults within 2 seconds at the system fault detection [27]. For instance, the machine learning
most, and attempted to obtain better results of up to algorithms, such as the k-Nearest Neighbors (kNN) algorithm,
0.021 seconds. Computational time is almost 1% of IEEE support vector machine, etc., are adopted in identifying grid
standards. faults [28], [29].
iii. While adding new components to the DER systems, we It is noted that the inclusion of renewable sources causes fre-
need to keep the system practical and feasible at the quent voltage and frequency fluctuations [20], [21] which may
same time. If renewable applications become wide-spread, cause power system vulnerabilities, making fault detection
energy marketing would become higher than ever [19]. methods unreliable. Although there are no fault conditions, the
Therefore, the fault detection module must be cost- protection devices or fault detection algorithms may receive
efficient for the new applications. erroneous signals from those random voltage and frequency
iv. DER systems gather various renewable energy resources fluctuations [22]. Therefore, the fault detection algorithms
together and create a gigantic heterogeneous environment, should also be scalable and adaptive when renewables pen-
which requires practical solutions to manage the whole etration increases. It is also observed that many algorithms
network. This is because, especially increasing PV farms are designed for the static power system [30] and in the case
on the network, it brings voltage and frequency deviations of network alterations by adding/dropping a line or adding
together [20], [21]. These deviations make the power new buses, most of the proposed algorithms may fail, so the
system unstable, and so it must be eliminated by using systems have to be adaptive.
different techniques. Consequently, fault detection meth- Considering the aforementioned issues, a real-time fault
ods can cause a mistake because of these fluctuations. detection technique is developed in this study using an opti-
Although there are no faulty conditions, transmission sys- mization enabled weighted ensemble based Machine Learning
tem operator’s devices or some algorithms could decide Algorithm. The proposed method is simple to implement as it
wrong decisions [22]. Then, the proposed fault detection uses voltage, frequency and phase angle signals obtained from
algorithms should also be scalable and adaptive for any PMUs.
type of power systems with a different number of buses. There are different solutions that use only one Machine
We observed that many proposed algorithms are designed Learning (ML) algorithm to reach the time goal [24], or com-
YAVUZ et al.: ADAPTIVE FAULT DETECTION SCHEME USING AN OPTIMIZED SELF-HEALING ENSEMBLE MACHINE LEARNING ALGORITHM 1147

plex ones that merge multi-machine learning algorithms [25], the set of observation data and checks the neighbors of the
[26] to obtain higher accuracy results. The proposed method, new incoming items according to the value of k, the number
in this study, is a type of ensembling type blending optimiza- of neighbors. In most of the ML applications, the k value is
tion algorithm where the Particle Swarm Optimization (PSO) chosen as a default 3 or 5, and also the Minkowski distance
finds the optimum weights to eliminate the forecast errors (1) is chosen as a distance length metric. The process of the
coming from each ML algorithm. Another rigid contribution of algorithm is simple; according to the number of neighbors
this method that can be used is that the system exhibits flexible and the coordinates of the new data (x1 , y1 ), · · · , (xn , yn ),
/ adaptable behaviors as a result of any change. As com- the k’s nearest neighbors are determined by measuring the
monly known, power systems exhibit a constantly changing/ distance, e.g., Euclidean distance, Minkowski distance or
changeable structure. Algorithm-based error detection systems Manhattan distance, for all of the newcomers, and finally the
may not be able to adapt to these variations. Such problems system decides the clusters of the new nodes with the closest
are frequently encountered, especially in methods developed distance [33].
by detecting the threshold. The Cross-validation method has n
! p1
been used to eliminate the overfitting problem. When the
X
d(x, y) = |(xi − yi )p | (1)
accuracy reaches 100%, the algorithm is memorizing the data i=1
set, named overfitting. This is one of the prevalent problems
2) Linear Discriminant Analysis (LDA)
in the machine learning field.
Linear Discriminant Analysis (LDA) can be used to re-
For most of this study, we used the overfitting technique,
duce the size, improve computational efficiency, and reduce
which comes from memorization of results instead of learning,
overfitting in non-digitized models. The LDA is very similar
and causes you to unrealisticlly reach a 100% accuracy rate.
to Principal Component Analysis (PCA). The PCA tries to
This is due to ignoring cross validation techniques that are
find the orthogonal component axis of the maximum variance
taken into consideration in this study to be able to represent
in a data set; The LDA tries to find the feature subspace
results in a real world power system environment.
that optimizes class separability. The LDA and the PCA are
The rest of the paper is organized as follows. Section II
linear transformation techniques that can be used to reduce the
discusses machine learning algorithms used in fault detection
number of dimensions in a dataset.
and their mathematical algorithmic background is represented.
The present state of the dataset is used to make the
The ensemble and boosting algorithms are discussed in Sec-
data more easily separable when it is not very convenient
tion III. In Section IV, the proposed algorithm architecture is
to separate the components. To achieve this, it also takes
presented. Effectiveness and success of the proposed algorithm
advantage of the covariance matrix. In fact, it is not literally a
by comparing with individual ML algorithms and Ensemble
classification algorithm. It can be used as a pretreatment before
methods are represented in Section V. Finally, findings of the
the classification process when there is not enough differences
proposed study are summarized in Section VI.
to distinguish the classes following the feature extraction. To
distinguish between classes, LDA examines the distribution of
II. I TERATIVE M ACHINE L EARNING A LGORITHMS IN
classes and uses the difference between the average values.
FAULT D ETECTION
3) Logistic Regression (LR)
ML applications are not computer programming, like tra-
Logistic regression is a statistical method used to analyze a
ditional computer algorithms. ML creates a special algorithm
dataset with one or more independent variables that determine
for a given data/situation which exactly fits the system.
a result. The result is measured by a binary variable (there are
In the power system, researchers are faced with different
only two possible results). In logistic regression, the dependent
problems that are very close to the needed ML applications.
variables must only be binary. In nother words, final results
Power transmission and distribution problems depend on too
are only 1 (TRUE, success, etc.) or 0 (FALSE, error, etc.) as
many variable states. A faulty condition could be happening on
encoded data.
transmission and distribution lines, for example; a bird/snake
The purpose of logistic regression is to find the most
can touch the cable, or a tree can fall down on the transmis-
appropriate (yet biologically plausible) model to define the
sion/distribution line, causing a short circuit. Since it is very
relationship between a number of independent (predictive or
difficult to encounter that these typess of faulty situations on a
explanatory) variables related to the two-way characteristic
regular basis, the training data set may need to be generated via
variable (dependent variable = response or outcome variable).
simulations. Then, the ML technique could be an effective way
Logistic regression produces the coefficients of a formula to
to detect these faults. When hardware-based solutions can not
estimate the probability.
entirely handle these problems, ML creates a special algorithm
4) Naı̈ve Bayes (NB)
for a given data/situation which exactly fits the system. Some
of the ML techniques which can be used in signal processing The algorithm has been widely studied since 1950, based
and fault detection are presented as follows. on the Bayes Theorem and it uses the idea of the simple prob-
abilistic classifier. In statistic and computer science, the NB is
A. Machine Learning Algorithms represented as conditional probability [31]. Terminologically,
1) k-Nearest Neighbor (kNN) the Bayesian probability is given in (2).
The kNN is easy to apply, and is a simple and effective algo- p(Ck )p(x|C k )
p(Ck |x) = (2a)
rithm for binary or multi-classification problems. It considers p(x)
1148 CSEE JOURNAL OF POWER AND ENERGY SYSTEMS, VOL. 8, NO. 4, JULY 2022

prior x likelihood
posterior = (2b) kNN LDA LR DT NB
evidence
where p(Ck |x) is the posterior probability of class, p(Ck ) is
the prior probability of class, p(x|Ck ) is the likelihood, P (x)
Choose
is evidence. algorithm
5) Decision Tree (DT)
In Machine Learning applications, the Decision Tree (DT)
Set the
is one of the most preferred algorithms due to its simplicity. parameters
The DT gives all possible outcomes and if you have enough
data for the next future prediction, it can decide precisely.
The DT uses a math-based background that relies on Shannon
Calculate
Information Theory and entropy calculations [34]. The biggest algorithm
entropy value is the start of branches and the whole tree
NO
follows it with the same scheme as shown below:
c
X Fit training data
E(s) = −pi log 2(pi ) (3) and get scores
i=1

where E(s) is entropy and it represents the power and dom-

Highest Save
inancy of the feature frequency. Therefore, the branch starts parameter
accuracy is
from the feature which has the biggest E(s) value; pi is the achieved YES values for this
data set
probability.

B. Finding Best Decision Variable for Each ML Algorithms Fig. 1. Tuning the parameters with grid search to calculate the best decision
value for each ML Algortihm.
Machine learning algorithms require several parameters that
can affect the accuracy rate. Before bagging several ML
techniques as an ensemble algorithm, in order to understand looping process before continuing with the optimization part.
the optimum parameters of each algorithm, the proposed Since the power system structure is not always kept constant
method checks the sub-set of parameters or decision variables and requires some changes over time, algorithms need to be
given in Table I (unshaded part). The ML algorithms were run able to work with updated network structures. This ability is
iteratively with different parameters and the best combination known as the self-healing property of algorithms that ensures
is obtained for a given algorithm as demonstrated in Fig. 1. the detection and update of decision variables and weights
This code cycle is processed for just one time until the when power system structures are changed.
power system or data structure changes. The main idea of For instance, when kNN is working, the algorithm checks
this loop is to obtain the best decision variables for the PSO the k neighbor value to find the best decision variable that pro-
optimization based Ensemble algorithm shown in Section IV. vides the maximum accuracy rate for a given dataset. Although
Briefly, the PSO-based Ensemble algorithm combines different these parameters can be altered with topological changes,
iterative ML techniques with the best parameters and weights. this proposed approach finds new values before applying the
Since chosen ML techniques for the bagging process will optimization algorithm. Thus, the system becomes robust and
always use the same local parameters during the decision time adaptive with various datasets.
span, optimum parameters are obtained individually with this
III. E NSEMBLE A LGORITHMS
TABLE I Ensemble methods refer to combining weak learning al-
PARAMETERS TO C ALCULATE THE D ECISION VARIABLES FOR S ELECTED
ML AND E NSEMBLE A LGORITHMS gorithms and transforming them into a strong learner with
additional processes. Weak learners can work sequentially, and
ML and Ensemble Decision Interval
Algorithms Variables each predictor tries to fix the previous results via the boosting
kNN number of [3 to 21, incremented by 1] method. The other approach is to combine the results of weak
neighbors (Testing results gives that k must learners with the bagging method. So, ensemble techniques
be 9 for that dataset to obtain
highest accuracy result) can be applied with a feedback mechanism, such as Adaboost
LDA tolerance 0.0001 and Gradient Boosting, or it can be a vote-based method, such
LR – – as the bagging approach. So the boosting algorithms work as
DT – –
NB – – ensembling methods.
Gradient Boosting and Adaboost algorithms are also tunned for
comparison with proposed method. A. Bagging Methods
AB Estimator [1 to 100, incremented by 5]
SAMME, SAMME.R (i) Majority/Hard voting is a simple case of voting methods
number of [11 to 71, incremented by 1] with a voting algorithm which is given in (4).
GB
trees
seed [1 to 11, incremented by 1]
∀C ∈ Cn (4a)
YAVUZ et al.: ADAPTIVE FAULT DETECTION SCHEME USING AN OPTIMIZED SELF-HEALING ENSEMBLE MACHINE LEARNING ALGORITHM 1149

φ = mode{C1 (x), C2 (x), C3 (x), · · · , CN (x)} (4b) to this algorithm in order to obtain the best result. The GB has

C1 (x)
  
0 2 types of algorithms; one of them is SAMME.R and the other
 C2 (x)  0 one is SAMME. Each algorithm’s decision variable values are
 ..  =  ..  (4c) listed in Table I (shaded part).
   
 .  .
CN (x) 1
IV. P ROPOSED PSO BASED W EIGHTED
φ = mode{0, · · · , 1} = 0 or 1 (4d) E NSEMBLE A PPROACH
where φ is the final decision of total results and it uses The workflow for the proposed algorithm is shown in Fig. 3
the Python’s mode command. This is a piece of code cycle; which combines five different ML algorithms explained in
C1 , C2 , · · · , CN are the classifier’s decision results, either 0 Section III, by using a novel weighted ensemble approach
or 1. blended with the Particle Swarm Optimization (PSO) tech-
In this scheme, each ML algorithm comes up with the nique to accurately detect the faults on the power system. The
decision on the given test case, separately. The final decision proposed approach is explained by the following steps:
will be given with the agreement of the majority. 1) Collect the data from the sensors and all other parts of
(ii) Soft voting gives the average probability of the decisions the system, e.g. SCADA, to train the system periodically.
rather than counting the votes on positive or negative decisions 2) For each ML algorithm, the best decision variables of the
coming from the ML algorithms. For example, when three techniques are calculated with the Brute-Force approach
algorithms give the decisions with (0.60, 0.60, 0.15), then as shown in Section III.B.
hard voting will decide negatively, since there are 2 positives 3) The weights of ML algorithms for the bagging process,
and one negative. However, the soft voting will decide it as which are based on the soft-voting technique, are cal-
being negative due to the average of the probability which is culated by Particle Swarm Optimization module. Thus,
0.45. In this example, algorithms have the same significance, PSO will give the best set of weights for the soft-voting
however, the weights can be different if the contribution of the approach. During this process, bagging is applied with the
algorithms is not the same. PSO calculated optimized weights until the next training
time window.
B. Boosting Methods 4) The ensembling process uses a minimum 2 and maxi-
AdaBoost (AB) iteratively repeats the weak learner algo- mum 5 ML algorithms according to PSO results. In this
rithm with given instances. In each iteration, misclassified process, cross validation was applied to the dataset to
data items are re-weighted according to the information gained evaluate the predictive performance of the model results
from the previous step. With this feedback mechanism, the AB in the training dataset step. In applying this approach,
runs a classifier, changes the weights, runs another classifier, one of the well-known methods, K-Fold Cross Validation
and repeats until most of the items are classified properly. (cv), is applied as cv=5. That means, 80% of data was
Thus, there is no parallel calculation, each step must follow chosen for the training part and 20% of data was for
the previous ones, just like a chain. testing purposes.
The Bagging and boosting processes are represented in 5) A self-healing algorithm has been developed which is
Fig. 2. Not only does the bagging process preceed the boosting adaptive against the structural changes and the algorithm
type but it also is reinforced with PSO that results in faster collects the data and controls the power system in case
processing time when compared with any other ML algorithm of any faults. Based on each of the algorithm weights
or boosting method. obtained in step-4, the power system control action is
triggered, in this case, in a PSCAD simulator.
6) If structural changes occur, the proposed method will
wi=1 wi=2 wi=n wi=1 wi=2 wi=n recalculate each algorithm’s special parameters (such as
the k parameter for KNN) and recalculate the optimum
values by following the same steps to obtain the best
fi=1 (x) fi=2 (x) fi=n (x)
weights for each algorithm’s self-improved algorithm
fi=1 (x) fi=2 (x) fi=n (x)
ability. Otherwise, if the system structure stays stable,
the calculated weights will be used to detect faults. In
M
fM=sign Σ (fx)
M that case, structural changes means the training dataset
fM=sign Σ (fx)
i=1 i=1 must be changed, so the parameters of each algorithm
must be recalculated, and PSO will optimize the weights
Bagging Boosting again and again until ane accurate result is obtained.
7) If there is no structural change e.g., add/drop a new line,
Fig. 2. Bagging and boosting methods. adding new renewable sources, step-6 would remain as
indicated.
Gradient Boosting (GB) also combines multiple ML algo- Since each algorithm has pros/cons for different character-
rithms based on weights. It collects weak learners and makes a istics of datasets, all of them are combined together to obtain
new stronger learner and works as a team. A dataset is applied the powerful sides of each.
1150 CSEE JOURNAL OF POWER AND ENERGY SYSTEMS, VOL. 8, NO. 4, JULY 2022

Start PSCAD
by using python
and start
collecting datas

Model Changed?
NO Add/drop new bus YES
or tranmission lines
Training
dataset Power system
control action

Apply to e.g.islanding
kNN LDA LR DT NB PSCAD operation, opening
circuit breakers

Particle
swarm Training
optimizer dataset

Final self-healing algorithm

PSO Module kNN LDA LR DT NB

in this part some
different weights are
given to algorithms
for bagging process
Ensembling process

Fig. 3. Workflow for the PSO based weighted ensemble algorithm.

As a result, the proposed method becomes flexible and PSO(mla1 , . . . , mlaN ) = {w1 , . . . , wN } (5e)
adaptive in case of any structural changes, which is a very
where M is the set of weak machine learning algorithms
normal and frequent behavior for the power system. This
represented by mla, K represents the class labels -which is
method does not require applying any signal processing tech-
“fault” or “not fault” in the scope of this paper-, W shows the
niques or any other pre-processing method, such as feature
weights of the algorithms. θ(M, W, K) is the function of the
selection techniques, but at the same time, it can obtain very
bagging process, and PSO calculates the optimal set of weights
high accuracy even with raw data. Most of the methods use
for θ(M, W, K) that obtains the highest accuracy. See Fig. 5
feature selection techniques to analyze the training data to
for random fault conditions.
obtain high accuracy results for the given dataset. However,
The main purpose behind this idea is explained in Fig. 4
the same selected features, or parameters that are used in the
with a mock-up example. With the proposed method, each
specific algorithm, can fail on another dataset, or scenario,
algorithm will close the gaps of other collaborative algorithms
because of their different characteristics. Thus, the proposed
and try to obtain a total consensus if there is a fault or not.
methods should be dynamically adapted to the collected data,
However, this consensus should also be as fast as possible
which is rare and hard to optimize and apply in the real
because of the time limitations while working in a real-time
environment. It is noted that, with the proposed method, real-
environment. By using the best-scored classifier combinations
time processing could be possible since it has an ability to
and their calculated weights, the proposed model will be
work with unprocessed newly collected data, as one of the
able to run the ensemble algorithm in real-time for islanding
powerful benefits with the self-healing adaptive background.
detection. Please note that the weight calculation process is
In the ensemble method, the key point is that the system will
computed just once unless there is no modification on the
give 0 weight to the algorithm if it is not needed in the selected
power system topology. In case of any structural changes in
set. Each weight represents the power rate of the algorithm
the power system model, the proposed algorithm will detect
for the given data set during the soft-voting process. The PSO
that difference and continuously train itself until it reaches
finds the best bag and provides the weights for each classifier
saturation.
in the combination. The generalized approach is shown as
follows:
V. R ESULTS A ND D ISCUSSIONS
M = {mla1 , mla2 , . . . mlaN }, ML Algorithms (5a) The islanding or fault situation must be detected as fast
K = {i0 , i1 , . . . , ik }, Class Labels (5b) and as accurate as possible, and for this purpose, conventional
W = {w1 , w2 , . . . , wN }, Weights (5c) signal processing techniques have aided in obtaining great
N
X predictions with machine learning algorithms. In addition to
θ(M, W, K) = arg maxi wj ∗ p(mlaj |i) (5d) these techniques, many feature selection methods are applied
j=1 in most of the techniques in literature to deal with deficiencies
YAVUZ et al.: ADAPTIVE FAULT DETECTION SCHEME USING AN OPTIMIZED SELF-HEALING ENSEMBLE MACHINE LEARNING ALGORITHM 1151

73.12% 71.12% 97.9%

92.55% 92.35% 80.93%

Black boxes are incorrect predictions

from dataset for each estimator

DT NB
kNN LDA LR

kNearest Linear Logistic Naïve Bayes New Combined

Decision Tree
Neighbor Discriminant Regression Algorithm
Analysis

Fig. 4. PSO effects for each ML algorithm.

13
bus and IEEE 39 bus power system models are used to show
14
the effectiveness of the proposed weighted ensemble-based
12 machine learning approach for fault detection. The standard
11
10 Wind Farm power system parameters are implemented during the PSCAD
9 simulations running on an IEEE 14 bus model [35]. In addition
6 to the well-known IEEE 14 bus model, the proposed method is
G
C 7 also applied to the modified version, shown in Fig. 4, to com-
1 C
PV Park pare the performances under entirely different characteristics
5 4 8
of power systems in pointing out its adaptive and self-healing
schemes. In this scenario, the IEEE 14 bus model system has
3
2 PV Park been modified connecting renewable sources at buses 3, 6,
G C and 8 by providing intermittency and uncertainties of voltage,
phase angle and frequencies to test the proposed method’s
adaptivity. In all three cases, PSCAD/EMTDC software is
Fig. 5. Modified IEEE 14 Bus study system including renewable sources.
coupled with Python to solve the problems and obtain solu-
tions in a co-simulated platform to mimic real-time scenarios.
and system defects. However, they focus on the given dataset PSCAD is being used to simulate the power system and
and provide excellent results for the specific state. When the Python for implementing ML algorithms. The time step of
technique provides high accuracy rates with feature selection, the co-simulation is kept as 50µs. By running the simulation
the same technique could fail on a different or dynamic data 5 sec, 210.000 of voltage, frequency and phase angle data are
structure. When a ML algorithm is improved and specialized generated for analysis as a CSV file format. About 80% of the
for a power system structure, it cannot be used on different gathered data is used for training the algorithm and then 20%
structures or it can fail in cases where there is any differences of the remaining data is used for the testing of each algorithm.
in the power system structure. Thus, most of the common Cross-validation was chosen as five, which means the algo-
solutions are not applicable in the real environment. Many im- rithms test the next 20% part of the dataset and this process
proved ML techniques show very high accuracy results, which continues five times until all data was used for the test. Than
means these improved models memorize the data instead of the results are obtained by taking an average of five different
learning. Thus, it can be easily recognized, when the algorithm accuracy results. For example, kNN accuracy results are 89%,
results in 100% accuracy, when there is an overfitting problem. 93.25%, 97%, 91%, and 92.5% for each cross-validation cycle.
To overcome these situations, a weighted and self-healing The average of the five accuracy values is 92.55% as stated
ensemble technique has been proposed in this study by choos- in Table III. The overfitting problem was handled in this way
ing 5 most robust machine learning algorithms; The IEEE 14 and ensures more realistic results.
1152 CSEE JOURNAL OF POWER AND ENERGY SYSTEMS, VOL. 8, NO. 4, JULY 2022

After these steps, each ML algorithm (kNN, LDA, LR, TABLE II

D ECISION PARAMETERS AND C OMPARISON OF ACCURACIES FOR E ACH
NB, DT), boosting algorithm (AB, GB), and finally proposed C LASSIFIER ON C LASSICAL IEEE 14 B US M ODEL
PSO based ensemble method are investigated and compared Algorithm Parameters Algorithms Accuracy
in terms of effectiveness, adaptivity, and process-time under Individual Process Results %
two different case scenarios. Times (s)
K Nearest Neighbor = 9 1.7 92.55
Neighbor
A. Case 1: Fault Analysis and Comparison Linear Tolerance = 1.49 92.35
In power system applications, symmetrical and unsymmet- Discriminant 0.0001 Solver
Analysis = svd
rical faults are occurring. In that perspective, we choose one of Logistic – 2.3 80.93
the most common occurring faults, which is an unsymmetrical Regression
fault. In this study, unsymmetrical faults are applied. In Naı̈ve Bayes – 0.6 73.12
Decision Tree criterion = 2.0 71.12
practice, just 2%–5% of symmetrical faults are occurring in gini splitter =
the power system. This unsymmetrical fault type is one of best
the most difficult ones to detect, compared to severe double- AdaBoost – 0.57 93.13
Gradient – 0.29 92.85
line-to-ground (2LG) and three-line-to-ground (3LG) faults. Boosting
The system performance has really been tested in the most XgBoost [36] – 0.11 97.62
challenging situations: single-line-to-ground (1LG) faults. PSO-Ensemble Combination 0.021 97.93
of 5 ML PSO
In this study, the 1LG fault has been selected and applied Weights
randomly for 0.15 s in lines (7–8, 10–11, 9–14, 2–3, 6–12) at kNN 0.51
LDA 0.49
5 different time instants (1.6, 2.3, 2.8, 3.8, 4.4), respectively.
LR 0.24
Different fault locations are chosen to observe the effect of the DT 0.14
proposed algorithm as the voltage drops vary randomly. For NB 0.82
instance, the voltage response at bus 9 is given in Fig. 6 to
show the fault’s effect on the voltage response. Also, for each
bus, the frequency and phase values have been used for the then individual and boosting algorithms. The proposed method
prediction process. The PSCAD timed fault logic box creates is also achieving slightly better accuracy results, compared to
faults, and the impedance is selected as 100 kΩ for the default. the XgBoost algorithm [36].
Fault locations are respectively; Process time-wise, the PSO-based Ensemble algorithm
works faster than other approaches because of its bagging
• Line 7–8, 63 km, close to bus 7,
scheme and pre-computed background in which, as an evolu-
• Line 10–11, 23 km, close to bus 11
tionary algorithm, PSO particles communicate with each other
• Line 9–14, 43 km, close to bus 14
and they use their own previous best solution. Thus, it can
• Line 2–3, 35 km, close to bus 3
reach final results very quickly. Each of the ML algorithms
• Line 6–12, 35 km, close to bus 12
are strengthened by PSO, so PSO trains the algorithms to find
the best weights, however, this process only happens one time
1.05
unless the power system structures are kept constant. The same
1 collected dataset used for comparison with boosting algorithms
RMS Voltage (p.u.)

0.95 Fault at lines Fault at lines and the proposed method gives better results in terms of
Fault at lines 2-3 6-12 both process time and accuracy. The bagging application is
0.90
Fault at lines 10-11 faster than any other boosting algorithms (AB and GB) due
0.85 7-8 Fault at lines to the parallel processing structure. In the proposed method,
9-14
0.80
different powerful ML algorithms close their gaps as they work
together. That is the reason why the proposed method gives
better results.
0 1.5 2 2.5 3 3.5 4 4.5 5
Time (s)
B. Case 2: Power System Structure Adaptivity Test
Fig. 6. Voltage response at bus 9 (1 LG fault). In this case, two different rigid (IEEE-39 bus and IEEE-
14 bus modified) models are tested. The structural change is
Table II shows how the PSO-based weighted ensemble reflected in the simulation by adding 3 renewable sources at
method, which is a combination of multiple (at least 2, up buses 3, 6, 8 and this system is named as a modified IEEE
to 5) ML algorithms, can work on-fly effectively compared to 14 Bus model to test the proposed method in adaptivity and
well-known boosting algorithms for an IEEE 14 bus standard self-healing. The 1LG fault is also considered in this case.
model. Because one of the most difficult detections is a 1LG fault,
Table II also shows a comparison of the proposed PSO- its captured dataset of the voltage, frequency and phase angle
based ensemble method with other boosting algorithms (GB signals are considered for further analysis. For the IEEE-39
and AB), in terms of time and accuracy viewpoints using bus system, faults are applied randomly to locations for 0.15 s
the IEEE 14 bus model. Accuracy results show the proposed in lines (1-39, 3–4, 6–11, 9–39, 13–14, 16–19, 19–33, 23–24)
algorithm is around 5% more accurate, and also much faster at 8 different time instants (1.2, 2.4, 3.4, 4.4, 5.4, 6.4, 7.4),
YAVUZ et al.: ADAPTIVE FAULT DETECTION SCHEME USING AN OPTIMIZED SELF-HEALING ENSEMBLE MACHINE LEARNING ALGORITHM 1153

respectively. TABLE IV
D ECISION PARAMETERS AND C OMPARISON OF ACCURACIES FOR E ACH
The nature of PSO adopts structural changes immediately C LASSIFIER ON THE IEEE 39 B US M ODEL
and efficiently. The proposed method detects the structural Algorithm Parameters Algorithms Accuracy
changes, and if they happen, the best-tuned parameters of each Individual Process Results %
algorithm are re-calculated again. Then, PSO gives different Times (s)
weights to each of them again. Because structural changes K Nearest Neighbor = 9 2.23 79.59
Neighbor
affect the dataset, previous parameters and weights may not Linear Tolerance = 1.93 78.14
be useful in that case. An additional explanation can be seen Discriminant 0.0001 Solver
in Fig. 3, step 6. Analysis = svd
Logistic – 2.97 87.02
After showing the success of the proposed method, each Regression
well-known algorithm and boosting algorithm results are ob- Naı̈ve Bayes – 1.23 81.20
tained and tested to be compared with the proposed method Decision Tree criterion = 2.83 91.00
gini splitter =
using the modified IEEE 14 bus model. In this case, accuracy best
results can be seen in Table III which shows that the proposed AdaBoost – 0.59 73.17
method adaptivity is very high when compared with other Gradient – 0.89 94.15
Boosting
methods. Also, the IEEE 39 bus model has been tested under XgBoost [36] 1.37 95.89
the same conditions and the results are shown in Table IV. PSO-Ensemble Combination 1.13 96.61
of 5 ML PSO
TABLE III Weights
D ECISION PARAMETERS AND C OMPARISON OF ACCURACIES FOR E ACH kNN 0.305
C LASSIFIER ON THE M ODIFIED IEEE 14 B US M ODEL ( WITH PV) LDA 0.226
Algorithm Parameters Algorithms Accuracy LR 0.065
Individual Process Results % DT 0.139
Times (s) NB 0.772
K Nearest Neighbor = 9 1.94 92.61
Neighbor
Linear Tolerance = 1.63 87.38 • Boosting algorithms are applied for prediction, and so
Discriminant 0.0001 Solver
Analysis = svd final accuracy results show that the proposed method
Logistic – 2.45 76.3 works successfully and can obtain the highest accuracy
Regression which is more than the boosting algorithms, such as
Naı̈ve Bayes – 0.83 84.5
Decision Tree criterion = 2.13 82.38 AdaBoost and gradient boosting.
gini splitter = • The proposed Method can update by itself, in case of any
best existing topological differences.
AdaBoost – 0.57 81.23
• As an optimization algorithm, PSO, optimizes popular
Gradient – 0.29 91.35
Boosting algorithm’s weights to achieve an as accurate result as
XgBoost [36] 0.11 93.48 possible.
PSO-Ensemble Combination 0.036 96.68
• IEEE allows 2 second delays [37], but in this case,
of 5 ML PSO
Weights the experimental prediction time interval is 0.001. So
kNN 0.69 the proposal method achieved almost 2,000 times faster
LDA 0.40
LR 0.39
detection than IEEE’s allowed time delay. Also, for the
DT 0.23 different random faults, the algorithm continues to show
NB 0.85 the same success.
• Any possible voltage fluctuation or bus system violence

• There are no signal processing techniques and they don’t can mislead final results, therefore threshold setting is not
require threshold tunning. In this study, only a raw dataset a good solution so we don’t use any threshold settings for
has been used for training and prediction parts. As a result our proposed method.
of this method, any waste of time is avoided. Also, it is • Tables III and IV also show that the proposed method pro-

not necessary to have too much computational power. Be- vides almost 5% better accuracy performance compared
cause ML algorithm’s parameters and PSO optimizations to boosting algorithms and individual ML algorithms
processes are just one time applied until power system accuracy results in both modified IEEE 14 bus models.
structure changes occur. If the power system structure • In the standard topology, XgBoost and the proposed
changes, the proposed method calculates new parameters method seem very close to each other in terms of
and weights for each algorithm. accuracy performance. However, when the topology is
• Most of the popular algorithms (5 of them) are com- changed, and the system fluctuates more, the results
bined adaptively, and that mechanism obtains a real-time show that the proposed method is clearly overperforming
self-healing feature. The proposed method automatically XgBoost.
calculates the weights of each algorithm. When a faulty The results given in Table III and Table IV show that some
condition happens, or the current weights cannot give ML algorithms, such as kNN or GB in Table III, can provide
sufficient results in a timly manner, the training part re- high accuracy results in some cases, however, they may not
starts to overcome this issue, see Fig 3. be able to reach the same levels on different structures as in
1154 CSEE JOURNAL OF POWER AND ENERGY SYSTEMS, VOL. 8, NO. 4, JULY 2022

Table IV. That is the main reason to develop the bagging based R EFERENCES
ensemble algorithms to strength deficiencies and weaknesses [1] L. L. Grigsby, Power System Stability and Control, 2017.
in the ML algorithm in different scenarios and/or structures. [2] “Update to renewable energy integration in South Australia joint aemo
It is noted that Gradient boosting works more accurately and electranet report,” AEMO, New South Wales Queensland South
Australia Victoria Australian Capital Territory Tasmania Western Aus-
than Adaboost on the modified IEEE model. The main reason tralia, Important Notice Purpose, 2016.
for this is by adding renewable sources, the gathered data [3] A. Esmaeilian and M. Kezunovic, “Prevention of power grid blackouts
has more noise because of voltage/frequency fluctuations, using intentional islanding scheme,” IEEE Transactions on Industry
Applications, vol. 53, no. 1, pp. 622–629, Jan./Feb. 2017, doi: 10.1109/
and therefore, the Adaboost algorithm can be easily defeated TIA.2016.2614772.
by noise when compared with the Gradient Boosting. With [4] S. Raza, H. Mokhlis, H. Arof, J. A. Laghari, and L. Wang, “Application
respect to accuracy and process time, the proposed PSO of signal processing techniques for islanding detection of distributed
generation in distribution network: A review,” Energy Conversion and
based Ensemble method shows very high results over boost- Management, vol. 96, pp. 613–624, May 2015, doi: 10.1016/J.ENCO
ing algorithms. Also, the proposed method, compared with NMAN.2015.03.029.
the XgBoost algorithm, clearly outperforms. The results are [5] D. S. Pillai and N. Rajasekar, “A comprehensive review on protec-
tion challenges and fault diagnosis in PV systems,” Renewable and
promising for both conditions. Sustainable Energy Reviews, vol. 91, pp. 18–40, Aug. 2018, doi:
All these comparisons show that the proposed method 10.1016/j.rser.2018.03.082.
has an adaptive characteristic and it can work significantly [6] D. S. Pillai and N. Rajasekar, “Metaheuristic algorithms for PV pa-
better than any platform without changing pre-computational rameter identification: A comprehensive review with an application to
threshold setting for fault detection in PV systems,” Renewable and
techniques so that it provides a specific solution for a given Sustainable Energy Reviews, vol. 82, pp. 3503–3525, Feb. 2018, doi:
data set. Since the PSO weights are calculated dynamically, 10.1016/j.rser.2017.10.107.
whenever needed, the PSO-based Ensemble method can be [7] J. Keller and B. Kroposki, “Understanding fault characteristics of
inverter-based distributed energy resources,” Jan. 2010, doi: 10.2172/
easily adapted in different schemes and power topologies, so 971441.
that it can train and predict data at the same time. [8] H. G. Jiang, J. J. Zhang, W. Z. Gao, and Z. P. Wu, “Fault detection,
The performance of the PSO-based Ensemble method is identification, and location in smart grid based on data-driven compu-
tational methods,” IEEE Transactions on Smart Grid, vol. 5, no. 6, pp.
not significantly affected in uncertain cases in voltage and 2947–2956, Nov. 2014, doi: 10.1109/TSG.2014.2330624.
frequencies, such as adding renewable sources. This means [9] J. E. Tate and T. J. Overbye, “Line outage detection using phasor angle
that the addition of renewable sources affects the individ- measurements,” IEEE Transactions on Power Systems, vol. 23, no. 4,
pp. 1644–1652, Nov. 2008, doi: 10.1109/TPWRS.2008.2004826.
ual machine learning algorithms’ performance, however, the [10] L. Che, M. E. Khodayar, and M. Shahidehpour, “Adaptive protection
proposed method’s progress is extremely good, even in this system for microgrids: Protection practices of a functional microgrid
situation. system,” IEEE Electrification Magazine, vol. 2, no. 1, pp. 66–80, Mar.
2014, doi: 10.1109/MELE.2013.2297031.
[11] N. Senroy, G. T. Heydt, and V. Vittal, “Decision tree assisted controlled
VI. C ONCLUSION islanding,” IEEE Transactions on Power Systems, vol. 21, no. 4, pp.
1790–1797, Nov. 2006, doi: 10.1109/TPWRS.2006.882470.
The PSO-based Ensemble method is proposed in this study [12] J. A. Laghari, H. Mokhlis, M. Karimi, A. H. A. Bakar, and H. Mohamad,
to detect faults in the power system. The proposed algorithm “Computational intelligence based techniques for islanding detection
is tested on different models, such as IEEE 14, IEEE-39 bus of distributed generation in distribution network: A review,” Energy
Conversion and Management, vol. 88, pp. 139–152, Dec. 2014, doi:
systems and modified model by adding newly commissioned 10.1016/J.ENCONMAN.2014.08.024.
renewable sources, as a means to present the case of structural [13] A. Khamis, Y. Xu, Z. Y. Dong, and R. Zhang, “Faster detection of
change of the power system. microgrid islanding events using an adaptive ensemble classifier,” IEEE
Transactions on Smart Grid, vol. 9, no. 3, pp. 1889–1899, May 2018,
In the proposed method, there are no signal processing or doi: 10.1109/TSG.2016.2601656.
feature selection techniques needed, and just raw dataset (input [14] A. A. Aburomman and M. B. I. Reaz, “A novel weighted support
signal voltage, frequency and pahse angle) is used for the vector machines multiclass classifier based on differential evolution for
predictions. Thus, the proposed method is not just specialized intrusion detection systems,” Information Sciences, vol. 414, pp. 225–
246, Nov. 2017, doi: 10.1016/j.ins.2017.06.007.
for the dataset. According to the results, it is also adaptive and [15] S. R. Mohanty, P. K. Ray, N. Kishor, and B. K. Panigrahi, “Classification
flexible for any type of structure-based dataset. It is found of disturbances in hybrid DG system using modular PNN and SVM,”
that the proposed method provides much greater accurate International Journal of Electrical Power & Energy Systems, vol. 44,
no. 1, pp. 764–777, Jan. 2013, doi: 10.1016/j.ijepes.2012.08.020.
results than individual machine learning algorithms using three [16] Z. L. Gaing, “Wavelet-based neural network for power disturbance
IEEE models (14 bus-39-bus and PV added). The proposed recognition and classification,” IEEE Transactions on Power Delivery,
method’s accuracy rates are calcualted as 97.93% for the vol. 19, no. 4, pp. 1560–1568, Oct. 2004, doi: 10.1109/TPWRD.2004.8
35281.
IEEE classical model and 96.68% for the modified (PV added) [17] M. S. ElNozahy, E. F. El-Saadany, and M. M. A. Salama, “A robust
model and 96.61% for the IEEE-39 bus model. While IEEE wavelet-ANN based technique for islanding detection,” in Proceedings
1547 standards allow 2 second delays to detect any possible of 2011 IEEE Power and Energy Society General Meeting, 2011, pp.
1–8, doi: 10.1109/PES.2011.6039158.
faults, the proposed method obtains better results, up to 0.021 s [18] IEEE Standard for Interconnecting Distributed Resources with Electric
for the IEEE 14 bus and 0.036 s for the modified IEEE 14 Power Systems, IEEE Standard 1547–2003, 2003, doi: 10.1109/IEEEST
bus model. This means, the computational time is almost D.2003.94285.
[19] C. Candelise, M. Winskel, and R. J. K. Gross, “The dynamics of
1% of IEEE standards which is small enough for a power solar PV costs and prices as a challenge for technology forecasting,”
system study. Thus, the proposed method will provide higher Renewable and Sustainable Energy Reviews, vol. 26, pp. 96–107, Oct.
and faster results than the most popular machine learning 2013, doi: 10.1016/j.rser.2013.05.012.
[20] M. Nayeripour, M. Hoseintabar, and T. Niknam, “Frequency deviation
algorithms and also provides adaptivity for any structural control by coordination control of FC and double-layer capacitor in
changes. an autonomous hybrid renewable energy power generation system,”
YAVUZ et al.: ADAPTIVE FAULT DETECTION SCHEME USING AN OPTIMIZED SELF-HEALING ENSEMBLE MACHINE LEARNING ALGORITHM 1155

Renewable Energy, vol. 36, no. 6, pp. 1741–1746, Jun. 2011, doi: [35] PSCAD. (2018, May 22). IEEE 14 bus technical system[Online]. 2018.
10.1016/j.renene.2010.12.012. Available: https://www.pscad.com/knowledge-base/download/ieee 14 b
[21] W. Grycan, B. Brusilowicz, and M. Kupaj, “Photovoltaic farm impact on us technical note.pdf.
parameters of power quality and the current legislation,” Solar Energy, [36] M. H. Chen, Q. Y. Liu, S. H. Chen, Y. C. Liu, C. H. Zhang, and R. H.
vol. 165, pp. 189–198, May 2018, doi: 10.1016/J.SOLENER.2018.03. Liu, “XGBoost-based algorithm interpretation and application on post-
016. fault transient stability status prediction of power system,” IEEE Access,
[22] M. Datta, T. Senjyu, A. Yona, T. Funabashi, and C. H. Kim, “A vol. 7, pp. 13149–13158, Jan. 2019, doi: 10.1109/ACCESS.2019.2893
frequency-control approach by photovoltaic generator in a PV-diesel 448.
hybrid power system,” IEEE Transactions on Energy Conversion, vol. [37] D. F. Qiao, Q. S. Cheng, and Y. D. Hou, “Fault detection and isolation
26, no. 2, pp. 559–571, Jun. 2011, doi: 10.1109/TEC.2010.2089688. of sensor in time-delay systems based on space geometry method,” in
[23] S. Kar, S. R. Samantaray, and M. D. Zadeh, “Data-mining model based 2016 IEEE 11th Conference on Industrial Electronics and Applications,
intelligent differential microgrid protection scheme,” IEEE Systems 2016, pp. 444–449, doi: 10.1109/ICIEA.2016.7603625.
Journal, vol. 11, no. 2, pp. 1161–1169, Jun. 2017, doi: 10.1109/JSYS
T.2014.2380432.
[24] J. P. Singh and N. Bouguila, “Proportional data clustering using K-
means algorithm: A comparison of different distances,” in 2017 IEEE
International Conference on Industrial Technology, 2017, pp. 1048– Levent Yavuz graduated from the Erciyes Univer-
1052, doi: 10.1109/ICIT.2017.7915506. sity Physics and Electrical Electronic Engineering
[25] H. Shayeghi, B. Sobhani, E. Shahryari, and A. Akbarimajd, “Op- and currently studying as Research Assistant at the
timal neuro-fuzzy based islanding detection method for Distributed Abdullah Gül University.
Generation,” Neurocomputing, vol. 177, pp. 478–488, Feb. 2016, doi: He established a company, MaM High Tech Corp,
10.1016/J.NEUCOM.2015.11.056. by Tubitak Foundation as an entrepreneur.
[26] R. Aggarwal and Y. H. Song, “Artificial neural networks in power He received 5 patents. His research interests in-
systems. Part 3: Examples of applications in power systems,” Power clude machine learning, deep reinforcement learn-
Engineering Journal, vol. 12, no. 6, pp. 279–287, Dec. 1998, doi: ing, virtual power plant, microgrids, solar generation
10.1049/pe:19980609. forecasting, nuclear energy and renewable energy
[27] K. B. Debnath and M. Mourshed, “Forecasting methods in energy systems.
planning models,” Renewable and Sustainable Energy Reviews, vol. 88,
pp. 297–325, May 2018, doi: 10.1016/j.rser.2018.02.002.
[28] L. S. Moulin, A. P. A. Da Silva, M. A. El-Sharkawi, and R. J. Marks,
“Support vector machines for transient stability analysis of large-scale Ahmet Soran received the B.Sc. and M.Sc. degrees
power systems,” IEEE Transactions on Power Systems, vol. 19, no. 2, in Computer Engineering from TOBB ETÜ, Ankara,
pp. 815–825, May 2004, doi: 10.1109/TPWRS.2004.826018. Turkey, in 2009 and 2012, respectively, and the
[29] D. Zhang, X. Han, and C. Deng, Review on the research and practice of Ph.D. degree in Computer Science and Engineering
deep learning and reinforcement learning in smart grids, CSEE Journal from the University of Nevada, Reno (UNR), in
of Power and Energy Systems, vol. 4, no. 3, pp. 362–370, 2018, doi: 2017. He joined the Computer Engineering Depart-
10.17775/cseejpes.2018.00520. ment, Abdul- lah Gül University, Kayseri, Turkey,
[30] P. Sharma, S. Banerjee, S. Tandel, R. Aguiar, R Amorim, and D. in 2018. His research interests include graph theory
Pinheiro, “Enhancing network management frameworks with SDN-like applications, network communications, device-to-
control,” in 2013 IFIP/IEEE International Symposium on Integrated device protocols, network and traffic management,
Network Management, 2013, pp. 688–691. network architecture, smart grids, the Internet of
[31] O. N. Faqhruldin, E. F. El-Saadany, and H. H. Zeineldin, “Naive Things, cyber security, and privacy.
Bayesian islanding detection technique for distributed generation in
modern distribution system,” in 2012 IEEE Electrical Power and Energy
Conference, 2012, pp. 69–74, doi: 10.1109/EPEC.2012.6474982.
[32] H. L. He, W. Y. Zhang, and S. Zhang, “A novel ensemble method for
credit scoring: Adaption of different imbalance ratios,” Expert Systems Ahmet Önen received the B.Sc. degree in Electrical-
Electronics Engineering from Gaziantep University,
with Applications, vol. 98, pp. 105–117, May 2018, doi: 10.1016/j.eswa
in 2005, the M.S. degree in Electrical and Computer
.2018.01.012.
Engineering from Clemson University, in 2010, and
[33] J. N. Myhre, K. K. Ø Mikalsen, S. Løkse, and R. Jenssen, “Robust
the Ph.D. degree from the Department of Electrical
clustering using a kNN mode seeking ensemble,” Pattern Recognition,
and Computer Engineering, Virginia Tech, in 2014.
vol. 76, pp. 491–505, Apr. 2018, doi: 10.1016/j.patcog.2017.11.023.
He is currently working as an Associate Professor
[34] M. A. Hassan, A. Khalil, S. Kaseb, and M. A. Kassem, “Exploring the
with Sultan Qaboos University. His research interest
potential of tree-based ensemble methods in solar radiation modeling,”
Applied Energy, vol. 203, pp. 897–916, Oct. 2017, doi: 10.1016/j.apen includes the development and deployment of renew-
able energy along with AI.
ergy.2017.06.104.
1156 CSEE JOURNAL OF POWER AND ENERGY SYSTEMS, VOL. 8, NO. 4, JULY 2022

Xiangjun Li (M’06–SM’12) received the B.E. de- S. M. Muyeen (S’03–M’08–SM’12) received his
gree in Electrical Engineering from Shenyang Uni- B.Sc. Eng. Degree from Rajshahi University of
versity of Technology, China, in July 2001, and Engineering and Technology (RUET), Bangladesh
his M.E. and Ph.D. degrees in Electrical and Elec- formerly known as Rajshahi Institute of Technology,
tronic Engineering from Kitami Institute of Tech- in 2000 and M.Eng. and Ph.D. Degrees from Kitami
nology (KIT), Japan, in March 2004 and March Institute of Technology, Japan, in 2005 and 2008, re-
2006, respectively. From May 2006 to March 2010, spectively, all in Electrical and Electronic Engineer-
he worked as a postdoctoral research fellow at ing. At the present, he is working as a full Professor
the Korea Institute of Energy Research (KIER), in the Electrical Engineering Department of Qatar
Daejeon, Korea, and Tsinghua University, Beijing, University. His research interests are power system
China, respectively. In March 2010, he joined the stability and control, electrical machine, FACTS,
Energy Storage and Electro-technics Department (ESED) (before: Electrical energy storage system (ESS), Renewable Energy, and HVDC system. He
Engineering and New Material Department), China Electric Power Research has been a Keynote Speaker and an Invited Speaker at many international
Institute (CEPRI), Beijing, China, where he has been engaged in the topic conferences, workshops, and universities. He has published more than 250+
of integration/control/SCADA/application technologies for large-scale multi- articles in different journals and international conferences. He has published
type battery energy storage systems/stations, distributed generation systems, seven books as an author or editor. He is serving as Editor/Associate Editor
electric vehicles, and micro-grids. He is the director of energy storage system for many prestigious Journals from IEEE, IET, and other publishers including
integration & configuration technology laboratory for the ESED of CEPRI IEEE Transactions on Energy Conversion, IEEE Power Engineering Letters,
from June 2019. He is the director of large-scale energy storage technology IET Renewable Power Generation and IET Generation, Transmission &
and application laboratory of CEPRI from May 2020. His research interests Distribution, etc. Dr. Muyeen is the senior member of IEEE and Fellow of
include renewable energy power generation, electric energy saving/storage Engineers Australia.
technology, and power system engineering. He has served as the Chair of the
IEEE CIS Task Force on ADP and RL in Power and Energy Internets. He
has also served as the editor of the IEEE Transactions on Sustainable Energy,
editor of IET Renewable Power Generation, editor of Protection and Control
of Modern Power Systems (Springer journal), and guest editor of Engineering
(Elsevier journal). He has been awarded three U.S. Patents of Invention and 79
Chinese Patents of Invention. Prof. Li is a Fellow of IET, chartered engineer,
and is one of the senior members of IEEE, CSEE, CAS, and CES, etc.