IEEE Access

Dynamic Power Quality Disturbance Classification in Grid-

Integrated PV Systems: Leveraging Clark Transformed
Modal Voltage and Subspace Weighted KNN

Journal: IEEE Access

Manuscript ID Access-2024-12408

Manuscript Type: Society Section: IEEE Power & Energy Society Section

Date Submitted by the

29-Mar-2024
Author:

Complete List of Authors: Mishra, Sairam; Siksha O Anusandhan University Institute of Technical
Education and Research
Mallick, Ranjan kumar; Siksha O Anusandhan University Institute of
Technical Education and Research
Nayak, Pravati; Siksha O Anusandhan University Institute of Technical
Education and Research
R.T., Naayagi; Newcastle University International Singapore
Panda, Gayadhar; National Institute of Technology Meghalaya

Keywords: <b>Please choose

keywords carefully as they Machine learning, Microgrids, Power quality, Solar energy, Wavelet
help us find the most suitable transforms
Editor to review</b>:

Subject Category<br>Please
select at least two subject Computational and artificial intelligence, Power engineering and energy,
categories that best reflect Signal processing
the scope of your manuscript:

Additional Manuscript
Keywords:

For Review Only

Page 1 of 13 IEEE Access

1
2
3
4
5
6
7
AUTHOR RESPONSES TO IEEE ACCESS
8 SUBMISSION QUESTIONS
9 Author chosen
10 Research Article
manuscript type:
11
Author explanation
12
/justification for
13 CUST_JUSTIFICATION :No data available.
14
choosing this
15 manuscript type:
16 Author description of
17 how this manuscript fits
CUST_SCOPE :No data available.
18 within the scope of IEEE
19 Access:
20 Author description
21 detailing the unique
22 contribution of the CUST_UNIQUE_CONTRIBUTION :No data available.
23 manuscript related to
24 existing literature:
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60 For Review Only
 IEEE Access Page 2 of 13

1
2 Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
3 Digital Object Identifier 10.1109/ACCESS.2017.Doi Number
4
5
6
7
8
Dynamic Power Quality Disturbance
9
10
Classification in Grid-Integrated PV Systems:
11
12
Leveraging Clark Transformed Modal Voltage
13
14 and Subspace Weighted KNN
15
16 Sairam Mishra1, Ranjan K. Mallick2, Pravati Nayak1, Naayagi Thaiyal Ramasamy3, Senior
17 Member, IEEE, Gayadhar Panda4, Senior Member, IEEE
1
18 Department of Electrical Engineering, Siksha ‘O’ Anushandhan deemed to be University, Bhubaneswar, Odisha, PIN -751030, India
2
Department of Electrical and Electronics Engineering, Siksha ‘O’ Anushandhan deemed to be University, Bhubaneswar, Odisha, PI N-751030, India
19 3
Newcastle University, 172A Ang Mo Kio Avenue 8 #05-01, SIT Building @ Nanyang Polytechnic, Singapore-567739
20 4
National Institute of Technology Meghalaya, Bijni Complex, Laitumkhrah, Shillong, Meghalaya – 793003
21
22 Corresponding author: Naayagi Ramasamy (e-mail: [email protected]), Ranjan Kumar Mallick ([email protected])
23 This work is carried out in the Department of Electrical Engineering, SOA University, India in collaboration with Department of Electrical Power Engineering,
24 Newcastle University, Singapore and funding support is provided by Newcastle University.
25
26
27 ABSTRACT This study, emphasizing on Power Quality Disturbance Event (PQDE) detection of microgrids
28 plugged in with a Solar Energy Conversion System (SECS). A novel signal reduction technique called Clark
29 Transformed Modal is suggested in this work to reduce the three-phase voltage to a unit signal so as to ensure
30 better memory utilization and reduced computational burden in the feature extraction stage. Further, the
31 feature extracted from the unit signal with multi resolution analysis of Maximum Overlap Discrete Wavelet
32 Transform is fed into a novel subspace weighted kth nearest Neighbour (SWKNN) classifier. Several
33 disturbances such as sag, swell. transients, notches, flicker etc are intentionally simulated in a PV-grid tied
34 MATLAB/Simulink model. The outcome of the study uncovers the powerful classification ability of the
35 proposed MODWT-SWKNN classifier in terms of different performance indices and found to be superior
36 while comparing with several other competitive classifier models. Moreover, the disturban ce detection
37 system as a whole is verified under ideal as well as noisy conditions and found to be proficient in terms of
38 optimum detection time and accuracy in OPAL-RT real time environment.
39
40
INDEX TERMS Power Quality, Clark Transform; kth Nearest Neighbour; Maximum Overlap Discrete
41
Wavelet Transform; Microgrid
42
43
44
I. INTRODUCTION power quality issues etc [2]. Therefore, ensuring the secure
45
The escalating demand for electricity and a substantial and dependable operation of microgrids has become
46
decline in fossil fuel availability in recent decades have paramount. On top of that, the rapid expansion of global
47
prompted the exploration of alternative energy sources to fulfil industrialization has led to an increased utilization of nonlinear
48
energy needs. Integrating distributed energy sources, such as components, sensitive equipment, and relay protection
49
Renewable Sources of Energy (RSEs) or traditional devices. This trend has resulted in unwanted fluctuations in
50
synchronous generators, into distribution networks to ensure power signals. These deviations from standard values in
51
high-quality power supply has given rise to the concept of voltage, current, or frequency are collectively known as Power
52
micro-grids. Among various RSEs, Solar Energy Conversion Quality Disturbance Events (PQDEs). Common PQDEs
53
System (SECS) has gained notable traction worldwide due to include transients, harmonics, voltage sag/swell, unbalanced
54
its natural abundance, light weight, and cost-effectiveness [1]. voltage/current, and interruptions [3]. These disturbances can
55
However, SECS also brings a series of challenges such as have serious consequences, including sympathetic tripping,
56
intermittency, inverter interaction, on/off grid control aspects protection and control device malfunctions, damage to
57
58
59 For Review Only
VOLUME XX, 2017 1
60
Page 3 of 13 IEEE Access

1
2
3 sensitive electronic equipment, memory failures in monitoring Prakash et al. [16] proposed Independent Component Analysis
4 systems, loss of productivity, and more. These repercussions (ICA) for feature extraction combined with SVM for PQD
5 affect industries, utilities, and consumers, with end users classification, outperforming wavelet-based SVM. Diego et
6 ultimately bearing the brunt of the consequences. This al. [17] used wavelet packet signal processing with fuzzy
7 necessitates the development of a robust protection scheme classifier, demonstrating effectiveness in detecting PQ factors
8 capable of identifying and categorizing various types of Power like harmonics, power factor, and voltage variation in AC
9 Quality Disturbance Events. These events may occur due any MGs under diverse conditions. In [18] a deep Convolutional
10 intentional or unintentional operations in the microgrid which Neural Network (CNN) a pproach is presented, known for
11 needed to be monitored as per IEEE PQ monitoring standard noise immunity, accuracy, and speed in detecting single and
12 1159-2009 [4]. multiple PQDs. Apart from that a number of single classifier-
13 To counter the aforementioned issue, a number of PQ based PQ detection schemes are reported in literature [19-22],
14 detection techniques are reported in literatures that make use those may suffer from under/overfitting, limited
15 of the thresholds of system parameters [5-8]. While these generalization ability, difficulty in handling higher dimension
16 methods exhibit proficiency in detecting significant Power feature set etc. Ensemble classifiers [23-25] can be a solution,
17 Quality (PQ) events, establishing appropriate detection where a number of classifiers are trained to provide the result
18 thresholds involves balancing the rates of false alarms and in voting approach. Therefore, the work emphasis on an
19 accurate detections [9]. Furthermore, these techniques could ensemble machine learning technique (MLT) and chosen a
20 exhibit susceptibility to factors like harmonics, sampling base classifier called Kth Nearest Neighbour (KNN) due to its
21 frequency, and various user-defined parameters. To overcome unmatched ability of handling unbalanced data, non-
22 this particular issue, signal processing along with machine parametric nature, data adaptability, ease of feature and class
23 intelligence technologies are suggested by many authors [10] scaling etc. Moreover, KNN can be ensembled by sub-spacing
24 for automatic detection of PQDEs using the common principle technique. Yungi Wang et al. [26] introduced a new feature
25 displayed in Fig.1. extraction algorithm, normalized Renyi entropy, with KNN
26 and SVM classifiers to predict PQ events in islanded MGs. In
27 [10], various signal processing methods and classifiers were
28 applied for detecting and classifying single and multiple PQDs
29 in MGs where VMD with KNN showed higher classification
30 accuracy.
31 The research work deals with insightful feature extraction
32 for performing effective PQDE classification and detection.
33 FIGURE 1. Power Quality Event Detection with signal processing and This is possible if and only if the hidden component of the
machine learner disturbance signal is uncovered through a powerful signal pre-
34 Various Signal Processing Techniques (SPTs) have been
35 processor. In addition to that, the performance of the classifier
proposed for feature extraction in pattern recognition. These can be bolstered by effective ensemble applications to the
36 techniques utilize time-frequency transforms to differentiate
37 existing classifiers. However, the three-phase quantity may or
disturbances. While the (Fast Fourier Transform) FFT suffers may not have any disturbance in all phases and might have in
38 from spectral leakage and loss of time information [11], the
39 one or two phases. The study has taken all these problem
fixed window length of (Short Time Fourier Transform) STFT statements in to account and the respective solutions are
40 makes it unsuitable for non-stationary transient signals [12].
41 suggested as follows:
To address this, Wavelet Transform (WT) methods have
42
gained traction for analyzing such signals containing noise, ➢ Initially, three-phase voltage signal is collected
43
fluctuations and sinusoidal components [13]. Further, out of from the point of common coupling (PCC) of the
44
discrete and continuous WTs, the Discrete Wavelet Transform microgrid model.
45
(DWT) is preferred due to its adaptive window size and ➢ A novel Clarke’s Transformation based signal
46
predefined filter design, making it suitable for preprocessing reduction approach is adapted to create a unit
47
in classification tasks [14]. In Microgrid (MG) network, signal which will facilitate the consideration of all
48
researchers have employed various intelligent classifiers for three phases in to one signal.
49
Power Quality Disturbances (PQDs) classification. Tatiana ➢ The signal preprocessor namely, Maximum
50
Chakravorti et al. [3] used Reduced Kernel Ridge Regression Overlap Discrete Wavelet Transform (MODWT)
51
(RKRR) combined with Variational Mode Decomposition is implemented to compute the multi-level
52
(VMD) to detect islanding, faults, and PQ events in grid- components and the most efficient mother wavelet
53
connected MGs, achieving superior accuracy. Siavash Ranjbar for PQDE detection is identified.
54
et al. [15] employed Short-Time Fourier Transform (STFT) ➢ A total of 16 features are obtained from 5-level
55
extraction with Decision Tree (DT) classification for accurate detailed and approximate component and fed to a
56
fault detection in both grid-connected and islanded MGs. proposed novel classifier called Subspace
57
58
59 VOLUME XX, 2017 For Review Only 1
60
 IEEE Access Page 4 of 13

1
2
Weighted KNN, that shows significant outcomes MICROGRID C OMPONENT S PECIFICATION
3 Microgrid components Rating/specification
4 while detecting the PQDEs.
5 ➢ The detection time of the proposed MODWT- Utility 120KV, 50Hz, 2500MVA

6 SWKNN classifier is obtained through OPAL-RT TFM-1 120/25 KV, 50 MVA

7 real time simulation by taking the mean of 500 pi_T/L-1 14 Km, 50 Hz

8 disturbance observations. pi_T/L-2 8 Km, 50 Hz
9 The paper is structured as follows: Section 2 provides an
LOAD1 30 MW, 2Mvar
10 overview of the PV integrated microgrid system, detailing
LOAD2 2 MW, 2Mvar
11 specifications and power quality disturbance classes. Section
3 delves into signal pre-processing techniques. Section 4 Solar Plant 100KW*4 Units, 500V DC
12
13 discusses the formation of the feature matrix. Section 5 Voltage source converter 500KVA, 260 V
(VSC)
14 explores the weighted subspace KNN classifier and its
L-C Filter 45 𝜇H, 40 KVAR
15 proposed optimization methods. Section 6 presents the results
LOAD3 100 kW
16 and analysis of the study. Finally, the findings are summarized
and concluded in Section 7. TFM-2 0.26/25 KV, 450 KVA
17
II. MICROGRID SYSTEM DESCRIPTION
18
The considered microgrid is connected to 25 kV utility TABLE II
19 PQD EVENTS AND C ATEGORIES
system through breaker-1(BR-1) as presented in Fig. 2. It is
20 POWER
capable of operating both in on-grid mode and off-grid mode. EVENT
21 QUALITY EVENT SOURCE
CATEGORY
The power requirement during on-grid mode of operation is CLASS
22 PQC0 NORMAL UNDISTURBED
fulfilled by the utility. Alternatively, off-grid mode is activated
23 SINGLE LINE
by opening the contacts of BR-2. During this off-grid period, PQC1 LG
24 SAG
the 400-kW solar plant will take the charge of power delivery.
25 PQC2
LL DOUBLE LINE
The solar plant consists of 4 number of 100 kW units LLG SAG
26 SYMMETRICAL
cumulatively providing an output voltage of 500V DC through
27 FAULT (LLLL & LLLG)
MPPT controlled DC/DC converter. This DC output voltage
28 PQC3
TRANSFORMER TRIPPLE LINE
in converted to three phase AC through a PWM controlled
29 ENERGIZATION SAG
voltage source converter and connected to grid followed by a INDUCTION MOTOR
30 STATING
LC-filter. The specification and ratings of the microgrid
31 LARGE LOAD VOLTAGE
elements and components are listed in Table-I. Moreover, the PQC4
32 SWITCHING OFF SWELL
PQDs are simulated individually at the 25 kV BUS by
33 PQC5 CAPACITOR OSCILLATORY
actuating their respective breakers. The list of PQDs and the SWITCHING TRANSIENT
34 IMPULSE
event specific categories, sources and indices are presented in PQC6 LIGHTENING IMPULSE
35 TRANSIENT
Table-II.
36 NON-LINEAR LOAD
VOLTAGE
PQC7 NOTCH
37 SWITCHING
(HARMONICS)
38 ELECTRIC ARC
PQC8 FLICKER
39 FURNACE
40
III. SIGNAL PRE-PROCESSING
41
The initialization of the three-stage process of disturbance
42
detection as shown in Figure 1. starts with signal pre-
43
processing. The proposed work simplifies this stage as three
44
substages as follows,
45
46
A. DATA COLLECTION
47
This is the first step of signal pre-processing where the
48
voltage signal is kept on collected for a period 6 cycles with a
49
gap of one cycle from the 25kV bus (Fig. 2). This continues
50
process is carried out with 5kHz sampling frequency. The data
51
vectors are stored in a database as rows in sequential manner.
52
It is to be noted that the voltage data is converted to per unit
53 FIGURE 2. Circuit Diagram of the solar enabled microgrid system
before storage.
54
55
56
TABLE I
57
58
59 VOLUME XX, 2017 For Review Only 1
60
Page 5 of 13 IEEE Access

1
2
B. DATA REDUCTION unsuitable for real-time applications due to its limited
3
Various kinds of disturbances can initiate in a power redundancy during signal reconstruction. In contrast, DWT
4
network, affecting a single phase, either of the two phases, or addresses the CWT limitations by preserving the information
5
all three phases. Consequently, it's crucial to account for all of the input signal with least number of wavelet coefficients
6
three phases during analysis. But such comprehensive analysis [29]. For any signa l 𝑢(𝑡) DWT can be expressed as,
7
demands thrice memory space. A solution to this can be the 1 𝑘−𝑛𝑎 𝑖0
8 𝑊𝑇𝐷 (𝑖, 𝑘) = ∑𝑛 𝑢(𝑛) ∗ 𝑚 ( ) (3)
conversion of three-phase component into a single-phase 𝑎 𝑖0
9 √𝑎 𝑖0
component, leading to substantial reduction of processing time
10 Here, 𝑎 𝑖0 : scaling factor, 𝑛𝑎 𝑖0 : Translational Factor, 𝑚(∙) :
and minimal memory needs. One of such implementations is
11 mother wavelet, 𝑖, 𝑗 : non-negative integer values, 𝑘: sample
modal signal conversion [27]. But it is highly sensitive to
12 number in input signal
unbalanced disturbances. Therefore, a novel approach is
13 DWT performs a successive decomposition-reconstruction
adapted in this work with the help of Clarke’s Transformation.
14 of the disturbance signal using multi-resolution analysis. In
The basic form of Clarke’s Transformation is as follows,
15 2 1 1 this process two types of components are obtained namely
16 − −
𝑠𝛼 3 3 3 𝑠𝑎 approximate component (APC) and detailed component
1 1
17 [𝑠𝛽 ] = 0 √3 − √3 × [𝑠𝑏 ] (1) (DTC) in each level. In the first level, the APC-1 is obtained
18 𝑠0 1 1 1 𝑠𝑐 from a low pass filter (𝐹𝐿𝑃) whereas the DTC-1 are obtained
19 [3 3 3 ] from a high pass filter (𝐹𝐻𝑃). Further, APC-1 is decomposed
20 The single signal is constructed by combining all three on a similar fashion to obtain APC-2 and DTC-2 respectively
21 stationary reference frame components. Since sinusoids have and it continues till reaching number of levels demanded. It is
22 an exceptional ability to generate other sinusoids with to be noted that, after every decomposition stage the sampling
23 arithmetic operations, the new signal is obtained as, frequency of the signal is made half. A glimpse of the process
24 𝑠𝐶𝑇𝑀 = (𝑠𝛼 × 𝑠𝛽 ) + 𝑠𝑜 (2) can be seen on Fig. 4 which can be mathematically represented
25 The signal can be named as Clarke’s Transformed Modal as,
26 (CTM) Signal. It is to be noted that the signals must be 𝐼𝑛𝑝𝑢𝑡 𝑆𝑖𝑔𝑛𝑎𝑙, 𝐼 (𝑘) = ∑𝑛𝑖=1 𝐷𝑇𝐶𝑖 + 𝐴𝑃𝐶𝑛 (4)
27 converted to per-unit before making this conversion. A generic
28 CTM of 6 cycle normal signal can be seen in Fig. 3.
29
30
31
32
33
FIGURE 4. n-level DWT
34
35 In this study a more effective version of DWT is
36 implemented called Maximum Overlap Discrete Wavelet
37 Transform (MODWT) that employs filters similar to DWT,
38 but with the overlap between adjacent sub-bands. It helps to
39 reduce the boundary effects caused by filtering and down-
40 sampling. It further reduces the edge effects and allows for a
41 more accurate representation of the signal at different scales.
42 Moreover, MODWT can act on both stationary and non-
43 stationary components. The overlapping sub-bands can better
44 FIGURE 3. Clarke’s Transformed Modal of a Standard 3-phase Sine capture the transient behaviour of signals [28]. Due to such
45 Wave
characteristics, it can be extremely useful for analyzing
46 nonlinearity, transient and noise component present in power
47 C. DATA TRANSFORMATION (MAXIMUM OVERLAP
signals.
DISCRETE WAVELET TRANSFORM)
48
Before getting in to MODWT, it is necessary to understand
49 IV. FEATURE EXTRACTION
the basics of Wavelet Transform (WT). It proves to be a potent
50 The second stage of the three-stage process of disturbance
and efficient technique for analyzing swift transient signals
51 detection as shown in Fig. 1 is feature extraction. When
that arise within power networks during faults and unusual
52 designing features for a machine learning classifier, it's
incidents. By decomposing signals into various sub-
53 important to choose features that are relevant, informative, and
components while localizing both the time and frequency
54 representative of the underlying patterns in the data. The
domains, WT offers substantial utility. Typically, WT comes
55 specific features one MLC require will depend on the nature
in two primary forms: Continuous Wavelet Transform (CWT)
56 of the problem and the type of data. Therefore, the choice of
and Discrete Wavelet Transform (DWT). However, CWT is
57
58
59 VOLUME XX, 2017 For Review Only 1
60
 IEEE Access Page 6 of 13

1
2
3 features (also known as feature engineering) is a critical step k is set to a small value, the algorithm becomes more sensitive
4 in building effective machine learning classifiers. to outliers. Conversely, if k is chosen to be too large, the
5 Since, the chosen signal processor MODWT has inherent neighbourhood might encompass an excessive number of
6 property of preserving energy [28], the first feature is chosen points from different classes. Another concern is the strategy
7 to be a unique component called detail to approximate for aggregating class labels. The most straightforward
8 component energy ratio (DTAER). This can be defined as, approach is to rely on the majority vote, but this can be
𝑒𝐷𝑇𝐶 𝑗 ∑𝐾 2
9 𝑘=1 𝐷𝑇𝐶𝑗 (𝑘) problematic when the nearest Neighbour exhibit significant
𝐷𝑇𝐴𝐸𝑅𝐷𝑇𝐶𝑗 = = ∑𝐾 2
(5)
10 𝑒𝐴𝑃𝐶 𝑘=1 𝐴𝑃𝐶 (𝑘) variations in distance, and the closest Neighbour more
11 Secondly, the high frequency detailed component generally consistently represent the object's class. To understand this an
12 has a greater number of zero crossings to that of the example is taken in Fig. 5. Here 4 different classes (C1, C2,
13 approximate component. A ratio of these two can provide C3, C4) with a single feature are presented. If the unknown
14 hidden insights of the presence of high frequency components data is required to be determined through KNN classifier, it
15 in the signal. Therefore, the next feature called detail to will predict different classes randomly in every trial. But to
16 approximate component zero crossing ratio (DTAZCR) can naked eyes it is clearly seems to be nearer to C3. To overcome
17 be, this drawback, weighted KNN (WKNN) can be utilized which
𝑍𝐶𝐷𝑇𝐶 𝑗
18 𝐷𝑇𝐴𝑍𝐶𝑅𝐷𝑇𝐶𝑗 = (6) is an extension of the traditional KNN algorithm. In the
𝑍𝐶𝐴𝑃𝐶 standard KNN algorithm, each of the k nearest Neighbour
19 In general, a pure sinusoid is ideally having zero mean. But
20 contributes equally to the classification decision. However, in
the data reduction technique adapted in this study are the WKNN classifier, the influence of each Neighbour on the
21 generating CTM signals not uniform around the axis in most
22 classification is weighted based on certain criteria often called
cases if any disturbance exists. The same phenomenon will a kennel function. Typically, the weight is inversely
23 also propagate to MODWT multilevel decomposition of any
24 proportional to the distance, meaning that closer Neighbour
signal. Therefore, the simplest feature named mean of a signal have a greater influence on the classification decision, while
25 can be taken for all detailed and one approximate component.
26 more distant Neighbour have a reduced influence.
∑𝐾
𝑘=1 𝑠(𝑘)
27 𝑀𝐸𝐴𝑁𝑀𝑂𝐷𝑊𝑇𝑗 = (7) Moreover, to tackle the challenges occurred due to high
𝐾
28 dimensionality of the feature space, the random subs-pacing
V. PROPOSED CLASSIFIER technique is also utilized along with WKNN. Sub-spacing
29
KNN (K-Nearest Neighbour) is a non-parametric instance- introduces a randomization factor by considering only a subset
30
based learning method that provides solutions for problems of the available features to create a sub-classifier. Further an
31
with unknown distributions, particularly in non-nominal cases ensemble of sub classifiers together makes the prediction with
32
[34]. In instance-based learning, a distance metric function is a voting approach. This improves the generalization ability,
33
employed to compare the characteristics of each new instance provides robustness against irrelevant features and the
34
with existing ones. The nearest instance is then used to assign ensemble classifier as a whole becomes noise tolerant.
35
a class to the new instance. If a greater number of nearest If there are s subspaces such that,
36
Neighbour are considered, the class assigned to the new 𝑆𝑆 = [𝑆1 ,𝑆2 ,− − −, 𝑆𝑠 ] (8)
37
instance is determined by the majority class among the nearest And there are 𝑓 features such that,
38
k Neighbour [34]. 𝐹 = [𝑓1 , 𝑓2 , − − −, 𝑓𝑓 ] (9)
39
40 Further, 𝐹̀ is any random subset of 𝐹 such that {𝐹́1 ,𝐹́2 , − −
41 −, 𝐹́𝑠 } and if the dataset is denoted by 𝐷 and 𝐷́ is a random
42 subset of D such that {𝐷́1 , 𝐷́2 , − − −, 𝐷́𝑠 } then each subspace
43 𝑆𝑗 will going to be trained with only 𝐹́𝑗 features and 𝐷́𝑗 sub-
44 dataset.
45 Furthermore, in each subspace 𝑆𝑗 WKNN is
46 performed as follows,
1
47 𝑃(𝑇 = 𝑡|𝐹 = 𝑋𝑜) = ∑(𝑋𝑖 ,𝑡𝑖)∈𝐷́𝑗 𝑤𝑖 × 𝐼(𝑡𝑖 = 𝑡) (10)
𝑘
48 Here, 𝑃 denotes the conditional probability of each class.
49 𝑋𝑜 is the observation feature column matrix. 𝑤𝑖 is the weight.
50 𝐼(∙) is a Boolean operator that turns 1 when the given
51 observation (𝑋𝑖 , 𝑡𝑖 ) belongs to the class 𝑡 else 0.
52
53 VI. RESULT & DISCUSSION
54 The performance evaluation of the proposed MODWT-
FIGURE 5. A single Feature Representation for 4 Classes
55 SWKNN based PQ Detection system for gid connected PV is
One of the numerous factors impacting the performance of
56 discussed in this section. Initially, the model shown in Fig. 2
the KNN algorithm is the selection of the hyperparameter k. If
57
58
59 VOLUME XX, 2017 For Review Only 1
60
Page 7 of 13 IEEE Access

1
2
3 is designed and simulated in MALTAB Simulink elliptical circles which are occurred just before the
4 environment. synchronization and control algorithms takes their control.
5 The fault has been initiated at 0.1 sec to ensure the intentional
1
6 A disturbance should not interfere with the initial disturbances.

 0 − Transforma tion
0 6-Cycle BUFFER

7 -1 The subsequent disturbance event plots (Fig. 8) are therefore

8 1 only shown for 6 cycles i.e., 3 before and other 3 after the
0 B 6-Cycle BUFFER
initiation of disturbance events.
9 -1
10 1 Initial Disturbance at simulation Starting
11 0 C 6-Cycle BUFFER

12 -1

13  − component
CMT SIGNAL
14 X
15 Maximum  − component
16 Overlap
17 DWT
FEATURE SPACE (16 FEATURES)
+

18 D1 DTAERD1 DTAZCR D1 MEAND1 0 − component

19
20
SUB
SPACING
D2 DTAERD 2 DTAZCRD 2 MEAND 2
21
WKNN1
22
DTAER D 3 DTAZCRD 3 MEAND3
23 D3
FIGURE 7. One Line Sag
WKNN2
24
25 D4 DTAER D 4 DTAZCRD 4 MEAND4
26
27 D5 DTAER D 5 DTAZCR D 5 MEAND5
28
29 A5 MEANA5 WKNNS

30
31 CLASSIFICATION
32
33
MAX VOTING

34 PQD 1-8

35 FIGURE 6. Flow Diagram of The Proposed Work

36
37
The PQ event classes (PQC-0 to PQC-8) are intentionally (a) Double Line SAG
38
simulated at 0.1 sec. A 6-cycle data is getting captured with a
39
sampling frequency of 5kHz (100 data samples/cycle) i.e., 600
40
data samples at 100 sample gaps between each sample
41
collection instant. At every sample collection instant, all the
42
three-phase data are arithmetically combined to form theCMT
43
signal. This resulting signal is further processed through
44
MODWT multistep decomposition process to obtain 5
45
detailed and one approximate component. Further 16 features
46
are obtained from the decomposed signals and fed to the
47
SWKNN classifier to obtain the class of disturbance. A flow
48
diagram describing the work flow mentioned above is shown
49
Fig. 6. The choice of mother wavelet, value of K in WKNN
50
and the number of subspace variables are discussed in the later
51 (b) Triple Line SAG
sections.
52
53
A. WAVE FORM ASSESSMENT
54
B. Fig. 7 displays a single line sag event, which is intentionally
55
created in the test system by simulating a LG-Fault. The initial
56
disturbances at simulation starting are presented with dotted
57
58
59 VOLUME XX, 2017 For Review Only 1
60
 IEEE Access Page 8 of 13

1
2
3 It can be clearly observed from Fig. 8 the CTM signals of
4 different PQ events are distinctive with respect to each other.
5 It can be seen that; the unbalanced disturbances display good
6 presence in zero sequence component where the balance
7 disturbances have null or negligible presence. This particular
8 fact is used as an advantage in this study and a good arithmetic
9 mix of Clark’s Transformed signals is obtained to analyses the
10 class of disturbance.
11
12 C. DATASET FORMATION
13 Dataset formation simply means to create a tabular data
14 containing the features in columns along with the respective
(c) Swell class of event. The rows of the dataset represent different
15
16 occurrences of the events. Similarly, 5400*16 feature set and
17 a 5400*1 target set collectively forms the PQC dataset of this
18 study. Here 600 observations from each of the PQ disturbance
19 class have been created by processing the three-phase voltage
20 data collected at PCC of the studied system. This is termed as
21 Ideal Dataset (IDS). Moreover, on a technical aspect, the fault
22 data are collected with different fault resistances. Swell signals
23 are generated by suddenly removing different magnitude of
24 larger loads. Oscillatory Transients are generated with varying
25 capacitor bank switching. Further impulse transients are
26 created by lightening block with varying impulse amplitude.
27 Similarly harmonic signals are generated by non-linear load
28 (d) Capacitor Switching switching with varying the passive components. Lastly flicker
29 is initiated by Electric Arc Furnace switching. Further to
30 realize the real-world scenario another dataset is formed by
31 adding white gaussian noise to the existing dataset such that
32 20dB, 30dB and 40dB noise is added to 200 feature set of each
33 event. This will create another feature matrix termed as Noisy
34 Dataset (NDS). It has exactly the same dimension that of IDS.
35
D. OPTIMUM SWKNN PARAMETER SELECTION
36
37 Once the dataset is prepared, the next step is straight
38 forward to perform classification task. Since the chosen
39 classifier is subspace weighted KNN, the three major
40 parameters required to be obtained are the number of
41 subspaces, the number of decision variables and the number
(e) Lightening Impulse
42 of neighbours in each subspace. First of all, the number of
43 classes taken in the study are 9, therefore one cannot go below
44 9 neighbours which may lead to undesired outcomes. But the
45 nominal rule for the selection of ‘k’ in plain KNN is square
46 root of the available dataset i.e., √ 4800 = 69.28 ≈ 70.
47 Therefore, the number of subspaces can be set at 70/9 =
48 7.77 ≈ 8 such that there will be 9 neighbours in each
49 subspace. Now the decision variables (features) in each
50 subspace should be within 2 to 15 as the total number of
51 available variables are 16. If it is set to be very low then that
52 will affect the detection ability where if set to high will
53 increase computational burden. Therefore, a mathematical
54 study is done between the detection accuracy and
(f) Flicker
55 computational time to evaluate the number of decision
FIGURE 8. Other PQ Disturbances variable. It is to be noted that, the mother wavelet is set to
56
57 ‘sym4’ which is default for MODWT. Moreover, each
58
59 VOLUME XX, 2017 For Review Only 1
60
Page 9 of 13 IEEE Access

1
2
observation is 10-fold cross validated and the average SYMLETS ‘SYM2’ TO ‘SYM7’ 88.1%
3 ‘SYM45’
4 accuracy and computation time are presented in Table-III. COIFLETS ‘COIF2’ TO ‘COIF3’ 95.31%
5 ‘COIF5’
TABLE III DMEYER ‘DMEY’ ‘DMEY’ 99.75%
6 ACCURACY AND C OMPUTATION TIME FOR DIFFERENT DECISION FEJER- ‘FK4’, ‘FK6’, ‘FK6’ 94.31%
7 VARIABLE C OUNT IN SUBSPACES KOROVKIN ‘FK8’, ‘FK14’,
8 DECISION
COMPUTATION ‘FK18’, ‘FK22’
VARIABLE ACCURACY
9 COUNT TIME (IN MS)
10 2 27.52% 15.4
11 3 36.04% 26.7 F. PERFORMANCE EVALUATION
12 4 43.56% 34.9 The optimum MLC parameters along with the best mother
5 51.75% 44.5
13 6 59.27% 52.7
wavelet for MODWT are determined in the previous
14 7 72.5% 68.3 subsection. On account of those the performance of MODWT-
15 8 87.1% 81.9 SWKNN power quality detection scheme is evaluated in this
16 9 87.33% 99.7 subsection. Here, three other performance indices (PIs) are
10 87.47% 123.5
17 evaluated along with detection accuracy as of (11), (12), (13).
11 87.53% 157.2
18 12 87.64% 168.2 But the perquisites for calculating the PIs is the confusing
19 13 87.83% 183.7 matrix. Therefore, the confusion matrices of both IDS and
20 14 87.91% 201.6 NDS datasets are presented in Figure 9. and Figure 10. It can
21 15 87.97% 219.4 be seen that the overall accuracy of ideal dataset is found to be
22 99.74% whereas for noisy dataset it is 98.14%. It can also be
It can be observed that, the detection accuracy remains
23 observed that the PQC6 i.e., the impulse transient data are
steady near 87% when 8 or more variables are taken in a
24 detected with 100% accuracy both in ideal and noisy
subspace. But the overall computation time drastically
25 scenarios.
increasing with each surplus variable without significant
26
improvement in detection accuracy. Therefore, the random
27
variable per subspace is set to 8 so as to maintain an optimum
28
computation time without compromising the accuracy. The
29
obtained parameters for SWKNN are displayed in Table-IV.
30
31 TABLE IV
32 SWKNN P ARAMETER VALUES
33 NO. OF NEIGHBOUR PER FEATURE PER
SUBSPACES SUBSPACE SUBSPACE
34 8 9 8
FIGURE 9. Confusion Matrix of Ideal Dataset (IDS)
35
36
37 E. MDWT MOTHER WAVELET SELECTION
38 Further, the selection of best mother wavelet for the data
39 preprocessing stage could be a crucial factor which will hugely
40 affect the detection accuracy. To identify the same, all kinds
41 of orthogonal wavelet are rigorously tested with SWKNN
42 classifier with all possible wavelet combinations as shown in
43 Table-6. It can be clearly observed from Table-6 that the FIGURE 10. Confusion Matrix of Noisy Dataset (NDS)
44 Discrete Meyer Wavelet (‘dmay’) shows the highest accuracy
45 of detection amongst all other orthogonal wavelet variants. The performance indices are as follows,
46 Hence ‘dmay’ is chosen as the mother wavelet in this study. It 𝑃𝑟𝑒𝑐𝑖𝑠𝑜𝑛, 𝑃𝑋 =
𝐶𝑋
(11)
47 is to be noted that the parameter obtained for SWKNN as of
̂ 𝐶𝑋 +𝐶𝑥
𝐶𝑋
48 Table-V are used in this testing.
𝑅𝑒𝑐𝑎𝑙𝑙, 𝑅𝑋 =
𝐶𝑋 +𝐶𝑋
(12)
49 2∗(𝑃𝑋 ∗𝑅𝑋)
𝐹1𝑆𝑐𝑜𝑟𝑒, 𝐹1𝑋 = (13)
50 TABLE V 𝑃𝑋 +𝑅𝑋
51 OPTIMUM MOTHER WAVELET S ELECTION Where, 𝐶𝑋 is number of events truly detected as class -X
52 MOTHER VARIANTS OPTIMUM ACCURACY 𝐶𝑋 is number of events truly belongs to class -X but
WAVELETS OF WAVELET detected as any other class
53 FAMILY WAVELET ̂
𝐶𝑥 is number of events do not belongs to class -X but
54 HAAR ‘DB1’ ‘DB1’ 90.4%
55 DAUBECHIES ‘DB2’ TO ‘DB15’ 93.06% detected as class -X
56 ‘DB45’
57
58
59 VOLUME XX, 2017 For Review Only 1
60
 IEEE Access Page 10 of 13

1
2
3 Further it is loaded to OPAL-RT 4510 setup and the real-time
4 simulation is executed and results are opted in YOKOGAWA
5 multi signal oscilloscope (MSO) as shown in Fig. 10(b). The
6 detection is identified by an Event Index (EI) varies from 0.5
7 to 4 as per Table-2. It can be observed that, there is a detection
8 delay from the point of occurrence to point of detraction. It is
9 obvious because, the proposed detection scheme itself taking
10 a 6-cycle snapshot with a gap of one cycle, thereby introducing
11 a fixed delay of one cycle or 0.02 secs.
12
13
14
15
16
17 (a)Bar Diagram IDS
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33 (b)Bar Diagram NDS
34 (a)Real-Time Setup with MSO and HOST PC
FIGURE 11. Bar Chart of Performance Indices
35
36 The bar chart representations have been made to showcase
37 Double Line Sag Impulse Transient
different PIs valuation for every PQC in Fig. 11 considering
38 both ideal and noisy datasets. Here Fig. 9(a) displays the
39 outcome for IDS whereas Fig. 9(b) for NDS. While all the PIs 3

40 Triple Line Sag Oscillatory Transient 2.5

of IDS are well above 98%, the PIs of NDS for PQC7 are as 1.5
41 low as 93.46%, 95.33% and 94.38% (precision, recall, F1- 0.5
1

42 score) respectively. But the overall detection accuracy under ZOOM1 ZOOM2

43 noisy data condition is still sitting at 98.14%. Hence, the

44 MODWT-SWKNN classifier can be a good choice to be
45 implemented in real world PV based microgrid scenarios. Detection Duration
46 Detection Point

47 G. REAL-TIME VALIDATION
48 The results presented in the previous subsections are found Occurrence Point

49 through the model designed in MATLAB Simulink as well as (b)Disturbance CTM Signal and detected event indices
50 MATLAB script executions. But the outcomes are not enough
51 to assure whether the proposed MODWT-SWKNN PQ
FIGURE 12. Real Time Validation
52 Detection scheme will perform in practical grid scenarios or
53 This phenomenon reduces the computational
not. To resolve this concern, the PQ detection scheme is tested
54 complexity significantly and also provide a CPU cool off
in OPAL-RT real time simulation setup as shown in Fig. 12(a).
55 period in between successive snapshots. However, there are
First, the Simulink model is designed in the Host PC and the
56 some variable delays observed along with fixed delay which
respective C-codes are built in RT-LAB software platform.
57 differs from event to event. To counter this discrepancy, a set
58
59 VOLUME XX, 2017 For Review Only 1
60
Page 11 of 13 IEEE Access

1
2
of 500 observation are saved during real time simulation and H. COMPARATIVE STUDY
3
the average of sample difference between point of occurrence A comparative study involving the performance of
4
are detection are evaluated to calculate the detection time. various recently published PQ detection scheme in PV
5
Such an obtaining can be seen in Fig. 13. Finally, the mean integrated system are presented in Table-VI. The performance
6
detection time is calculated from the average detection of these methods in terms of detection accuracy are observed
7
difference of 142.3881samples with 5Hz frequency as both under ideal and noisy conditions with 3 SNR levels of
8
0.0285sec (142.3881/5000). 20dB, 30dB and 40dB respectively. Although [34, 38, 39]
9
have claimed to possess 100% ability of detection in ideal
10
condition but do not provide any information regarding the
11
performance under noise. Since in practical grid scenarios, the
12
current carrying conductors often experience electromagnetic
13
interference which is highly important to be considered while
14
designing a detection system. Therefore, the study tested the
15
performance both under ideal and noisy conditions. Moreover,
16
limited studies have focused upon the detection time [16, 39].
17
However, the study performed real-time detection of the
18
proposed scheme through OPAL-RT 4510 setup and the
19
detection time is calculated as explained in previous sub-
20
subsection. Although studies made in [16] and [39] obtained
21
the detection time, but lacking precise observations to made
22
during the calculation of it. However, one should take in to
23
account the factors like cycles to capture, signal propagation
24
delay, detection algorithm process delays etc. These factors
25
are already included in the applied real-time setup, hence an
26
accurate detection time of 0.0285 secs is obtained in this study.
27 FIGURE 13.
Observations
Average Detection Time Calculation with 500

28
29 TABLE VI
30 C OMPARATIVE ANALYSIS WITH PREVIOUSLY REPORTED PQD DETECTION S CHEMES IN PV EQUIPPED MICROGRIDS
31 MODEL PERFORMANCE (IN ACCURACY %)
TIME OF DETECTION
32 METHODS PQC IDEAL NOISY ACCURACY (SNR IN DB)
(IN SECS)
33 ACCURACY 20 30 40
34 [16] 6 99.5 - - - 0.07
35 [32] 5 - 94.7 97.2 99.75
36 [33] 16 98 - - - -
[34] 10 100 - - - -
37 [35] 10 97 - - - -
38 [36] 9 99.59 99.5 99.29
39 [37] 14 99.82 - 99.29 -
40 [38] 7 100 - - - -
[39] 16 100 97.54 99.81 99.96 0.184
41
PROPOSED 9 99.74 97.28 97.94 99.22 0.0285
42
43
44 VII. CONCLUSION KNN classifier called Subspace Weighted KNN so as to
45 The study solely focuses on the detection of 9 ensure efficient handling of high dimensional data. It is
46 different types of power quality events in a grid tied PV observed that, 8 subspaces each with 9 Neighbour and 8
47 system. At first, a novel signal reduction technique called features per subspace is showing highest detection accuracy of
48 Clack Transformed Modal is proposed to convert the three 99.75% with the Discrete Meyer mother wavelet. In addition
49 phase per unit PCC voltage to a solitary signal. This to that, the mean detection time of the proposed MODWT-
50 conversion ensures lesser memory utilization and reduced SWKNN based classifier is obtained by taking the detection
51 computational complexity in the signal processing stage. time mean of 500 random events in OPAL RT 4510 real-time
52 Further, 16 inherent features (DTAER, DTAZCR, MEAN) of simulation environment and found to be 0.0285 sec. Lastly,
53 the 6-cycle windowed CTM signals are obtained through the work is compared with few of the pre-published works that
54 MODWT 5-level decomposition. A dataset is created with especially emphases on PQ detection in grid tied PV system
55 dimension 5400*16 containing both ideal and noisy data and it is found that the detection accuracy and time of the
56 features for all 9 PQ events. Furthermore, an improved version proposed work is highly appreciating and can be implemented
57 of KNN classifier is proposed by sub-spacing the weighted in real grid scenarios. A wide variants of simple and complex
58
59 VOLUME XX, 2017 For Review Only 1
60
 IEEE Access Page 12 of 13

1
2
PQ event detection can be engrossed as a future work of this disturbances,’’ IEEE Trans. Smart Grid, vol. 9, no. 4, pp.
3 3018–3028, Jul. 2018.
4 study. [14] Chaovalit, Pimwadee, et al. "Discrete wavelet transform -
5 based time series analysis and mining." ACM
ACKNOWLEDGMENT Computing Surveys (CSUR) 43.2 (2011): 1-37.
6 [15] Ranjbar, Siavash, Ali Reza Farsa, and Sadegh Jamali.
The Authors wish to extend their appreciation to the Microgrid
7 "Voltage‐based protection of microgrids using decision
Research Laboratory at ITER, Siksha 'O' Anusandhan
8 tree algorithms." International Transactions on
Deemed to be University, Bhubaneswar, for their invaluable Electrical Energy Systems 30.4 (2020): e12274.
9
assistance and provision of resources throughout the duration [16] Prakash K. Ray, Asit Mohanty, Tapas Panigrahi, Power
10 quality analysis in solar PV integrated microgrid using
of this research work. The authors also thankful to the
11 independent component analysis and support vector
Department of Electrical Power Engineering, Newcastle machine, Optik 180 (2019) 691–698.
12
University, Singapore for providing technical assistance. [17] Diego H.S. Nolasco, Flavio B. Costa, Eduardo S.
13 Palmeira, Denis K. Alves, Benjamín R.C Bedregal,
14 Thiago O.A Rocha, Ricardo L.A Ribeiro, Juliano C.L
REFERENCES
15 Silva, Wavelet-fuzzy power quality diagnosis system
with inference method based on overlap functions: Case
16 [1] Adefarati, Temitope, and Ramesh C. Bansal. study in an AC microgrid, Eng. Appl. Artif. Intell. 85
17 "Integration of renewable distributed generators into the (2019) 284–294
18 distribution system: a review." IET Renewable Power [18] Shouxiang Wang, Haiwen Chen, A novel deep learning
Generation 10.7 (2016): 873-884. method for the classification of power quality
19 [2] Sharma, Arvind, et al. "Solar photovoltaic-based disturbances using deep convolutional neural network,
20 microgrid hosting capacity evaluation in electrical Appl. Energy 235 (2019) 1126–1140.
21 energy distribution network with voltage quality [19] Alshahrani, Saeed, Maysam Abbod, and Basem Alamri.
analysis." SN Applied Sciences 3 (2021): 1-22.
22 [3] Tatiana Chakravorti, N.R. Nayak, Ranjeeta Bisoi, P.K.
"Detection and classification of power quality events
based on wavelet transform and artificial neural
23 Dash, Lokanath Tripathy, A new robust kernel ridge networks for smart grids." 2015 Saudi Arabia Smart
24 regression classifier for islanding and power quality Grid (SASG). IEEE, 2015.
25 disturbances in a multi distributed generation based
microgrid, Renew. Energy Focus 28 (2019) 78–99,
[20] Khokhar, Suhail, et al. "A new optimal feature selection
algorithm for classification of power quality
26 [4] "IEEE Recommended Practice for Monitoring Electric disturbances using discrete wavelet transform and
27 Power Quality," in IEEE Std 1159-2019 (Revision of probabilistic neural network." Measurement 95 (2017):
28 IEEE Std 1159-2009) , vol., no., pp.1-98, 13 Aug. 2019, 246-259.
doi: 10.1109/IEEESTD.2019.8796486. [21] Zhao, Feng, et al. "Recognition of Hybrid PQ
29 [5] Reddy, Maddikara Jaya Bharata, Kondaveti Sagar, and Disturbances Based on Multi-Resolution S-Transform
30 Dusmanta Kumar Mohanta. "A multifunctional real‐tim e and Decision Tree."
31 power quality monitoring system using Stockwell [22] Samanta, Indu Sekhar, et al. "Dual-tree complex wavelet
transform." IET Science, Measurement & packet transform and regularized extreme learning
32 Technology 8.4 (2014): 155-169. machine-based feature extraction and classification of
33 [6] A. F. Bastos, W. Freitas, G. Todeschini and S. Santoso, power quality disturbances." Energy Systems (2023): 1-
34 "Detection of inconspicuous power quality disturbances 26.
through step changes in rms voltage profile", IET Gener.
35 Transmiss. Distrib., vol. 13, no. 11, pp. 2226-2235, Jun.
[23] Bashawyah, Doaa A., and Abdulhamit Subasi. "Power
quality event detection using fawt and bagging ensemble
36 2019. classifier." 2019 IEEE International Conference on
37 [7] A. F. Bastos and S. Santoso, "Universal Waveshape - Environment and Electrical Engineering and 2019 IEEE
38 Based Disturbance Detection in Power Quality Data Industrial and Commercial Power Systems Europe
Using Similarity Metrics," in IEEE Transactions on (EEEIC/I&CPS Europe). IEEE, 2019.
39 Power Delivery, vol. 35, no. 4, pp. 1779-1787, Aug. [24] Radhakrishnan, Padmavathi, et al. "A stacking ensemble
40 2020, doi: 10.1109/TPWRD.2019.2954320. classification model for detection and classification of
41 [8] G. S. Chawda et al., "Comprehensive Review on power quality disturbances in PV integrated power
Detection and Classification of Power Quality network." Measurement 175 (2021): 109025.
42 Disturbances in Utility Grid With Renewable Energy [25] Mishra, Sairam, et al. "A novel hybrid downsampling
43 Penetration," in IEEE Access, vol. 8, pp. 146807- and optimized random forest approach for islanding
44 146830, 2020, doi: 10.1109/ACCESS.2020.3014732. detection and non‐islanding power quality events
[9] M. Bollen and I. Gu, Signal Processing of Power Quality
45 Disturbances, Hoboken, NJ, USA:Wiley, 2006.
classification in distributed generation integrated
system." IET Renewable Power Generation 15.8 (2021):
46 [10] Mishra, Manohar. "Power quality disturbance detection 1662-1677.
47 and classification using signal processing and soft [26] D. E. Olivares et al., "Trends in Microgrid Control," in
48 computing techniques: A comprehensive IEEE Transactions on Smart Grid, vol. 5, no. 4, pp.
review." International transactions on electrical energy 1905-1919, July 2014, doi: 10.1109/TSG.2013.2295514.
49 systems 29.8 (2019): e12008. [27] Gupta, Preeti, and R. N. Mahanty. "An approach for
50 [11] Roy, Subhamita, and Sudipta Debnath. "PSD based high detection and classification of transmission line faults by
51 impedance fault detection and classification in wavelet analysis." Energy Education Science and
distribution system." Measurement 169 (2021): 108366. Technology 34.1 (2016): 109-122.
52 [12] Okelola, Muniru Olajide. "Power Quality Events [28] Alarcon-Aquino, Vicente, and Javier A. Barria. "Change
53 Classification on Real-Time Voltage Waveform Using detection in time series using the maximal overlap
54 Short Time Fourier Transform and Bayes discrete wavelet transform." Latin American applied
Classifier." International Journal of Applied 8.2 (2018). research 39.2 (2009): 145-152.
55 [13] K. Thirumala, M. S. Prasad, T. Jain, and A. C. Umarikar, [29] Ghanbari, Teymoor, and Ebrahim Farjah. "A multiagent-
56 ‘‘Tunable-Q wavelet transform and dual multiclass based fault-current limiting scheme for the
57 SVM for online automatic detection of power quality
58
59 VOLUME XX, 2017 For Review Only 1
60
Page 13 of 13 IEEE Access

1
2
microgrids." IEEE Transactions on Power Delivery 29.2
3 (2013): 525-533.
4 [30] Zakaria, Nur Adrinna Shafiqa, et al. "Comparison Prof. RANJAN KUMAR
5 analysis of different classification methods of power MALLICK was born in India, in
6
quality disturbances." International Journal of Electrical 1972. He received the bachelor’s
and Computer Engineering (IJECE) 12.6 (2022): 5754-
7 5764.
degree in electrical engineering
8 [31] Amin, Adnan, et al., eds. Advances in Pattern from the Institution of Engineers,
9 Recognition: Joint IAPR International Workshops, India, in 1996, the M.E. degree in
SSPR'98 and SPR'98, Sydney, Australia, August 11-13, power system engineering from
10 1998, Proceedings. Springer Science & Business Media,
11 1998. VSSUT, Burla, Odisha, in 2001,
12 [32] Prakash, Satya, Shubhi Purwar, and Soumya R. and the Ph.D. degree from BPUT,
Mohanty. "Adaptive detection of islanding and power Odisha, in 2013. He is currently
13 quality disturbances in a grid-integrated photovoltaic
14 system." Arabian Journal for Science and
working as a Professor with the
15 Engineering 45.8 (2020): 6297-6310. Department of Electrical and Electronics Engineering, ITER,
16
[33] T. P. Tun and G. Pillai, "Power Quality Event SOA Deemed to be University, Odisha, India. He is having 22
Classification in Distribution Grids Using Machine
17 Learning," 2021 56th International Universities Power
years of experience in teaching and research. His research
18 Engineering Conference (UPEC), Middlesbrough, interests include the application of power electronics,
19 United Kingdom, 2021, pp. 1-6, doi: optimization techniques in power systems, economic load
10.1109/UPEC50034.2021.9548222.
20 dispatch, and the design and control of HVDC converters,
[34] Vinayagam, Arangarajan, et al. "An ensemble approach
21 of classification model for detection and classification of Load frequency controllers, and Optimal Harmonic estimation
22 power quality disturbances in PV integrated microgrid in power systems, Stability of power system, power quality
network." Applied Soft Computing 106 (2021): 107294. along with machine learning and deep learning applications in
23 [35] Vinayagam, Arangarajan, et al. "A random subspace
24 ensemble classification model for discrimination of Power engineering.
25 power quality events in solar PV microgrid power
network." PloS one 17.1 (2022): e0262570.
26 [36] Yılmaz, Alper, Ahmet Küçüker, and Gökay Bayrak.
Dr. PRAVATI NAYAK
27 "Automated classification of power quality disturbances was born in 1983, India. She
28 in a SOFC&PV-based distributed generator using a received her Bachler degree in
29
hybrid machine learning method with high noise electrical engineering .in 2004
immunity." International journal of hydrogen
30 energy 47.45 (2022): 19797-19809.
and MTech. Degrees in power
31 [37] Dash, P. K., et al. "Multiple power quality disturbances electronics and drives in 2010 and
analysis in photovoltaic integrated direct current she has completed Ph.D.in SOA
32 microgrid using adaptive morphological filter with deep Deemed to be university in2022.
33 learning algorithm." Applied Energy 309 (2022): She is currently working as an
34 118454.
Associate Professor in the
35 [38] Pattanaik, Debasish, et al. "Power Quality Disturbance
Electrical Engineering
Detection and Monitoring of Solar Integrated Micro-
36 Grid." WSEAS Transactions on Power Systems 17 Department, ITER, SOA Deemed to be University, Odisha,
37 (2022): 306-315. India. She has a teaching experience of more than 16 years.
38 [39] Eristi, Belkis, and Huseyin Eristi. "Classification of
Her broad area of research includes Microgrids, Power
Power Quality Disturbances in Solar PV Integrated
39 Power System Based on a Hybrid Deep Learning Systems Protection, Artificial Intelligence, Power Quality,
40 Approach." International Transactions on Electrical Power Electronics. Machine learning and Deep learning
41 Energy Systems 2022 (2022). applications in Power Systems.
42
43 SAIRAM MISHRA is a graduate R. T. NAAYAGI R. T.
44 student member IEEE. He is NAAYAGI (Senior Member,
45 continuing his Ph.D. in Siksha O’ IEEE) received the bachelor’sand
46 Anusandhan University, master’s degrees (Hons.) in
47 Bhubaneswar-751002, Odisha, electrical and electronics
48 India in the department of engineering in India, and the
49 Electrical Engineering. He has Ph.D. degree in electrical and
50 completed his M.Tech with Power electronic engineering from The
51 Electronics and Drives University of Manchester,
52 Specialization in 2018 from Indira Manchester, U.K. She is working
53 Gandhi Institute of Technology, for Newcastle University in
54 Sarang, Dhenkanal-759146, Odisha and B.tech in 2014 from Singapore (NUiS) where she is currently the Director of
55 Siksha O’ Anusandhan University, Bhubaneswar-751030, Education & Associate Professor in Electrical Power
56 Odisha, India in Electrical Engineering Stream. Engineering. She received several merit certificates for her
57 academic proficiency, including the Best Outgoing Female
58
59 VOLUME XX, 2017 For Review Only 1
60
 IEEE Access Page 14 of 13

1
2
3
Graduate Award during her bachelor’s and the Outstanding proceedings / journals and filed 2 patents. He has supervised 8
4
Master’s Student Award. She was a first recipient of the PhD students and currently supervising 4 PhD Scholars. He is
Dorothy Hodgkin Post-Graduate Award from the School of a recipient of the Power Medal, IE(India) for one of his
5
Electrical and Electronic Engineering, The University of research papers. He has led several research projects on the
6
Manchester, for her Ph.D. degree, jointly sponsored by Rolls- integration of renewable energy generation and power quality
7 Royce plc and the Engineering and Physical Sciences improvement. He is a Senior Member of IEEE, Fellow of IE,
8 Research Council, U.K. She received the Woman Engineer and Life Member of ISTE.
9 Award from the Young Professionals Section Chennai, Dr. Panda is Associate Editor for IEEE Access, ITES, Wiley
10 Institution of Engineering and Technology (IET), U.K., in and also serves on the editorial board for the IJEEPS, De
11 2012. She received the Newcastle University Teaching Gruyter and the technical/steering committees for several
12 Award, in 2016. She received the Excellent Young Subject conferences. He has organized several technical events like
13 Editor Award, in 2020. She serves as a reviewer for the IEEE, IEEE Conference, Symposium, National Workshops, etc. He
14 IET, and many other international journals and conferences. has delivered several invited Talks at various national and
15 She was awarded Senior Fellow status of the Higher international events.
16 Education Academy, U.K., in 2019. In 2021, she has received His current research interests include automatic generation
17 the IEEE PES Outstanding Engineer Award in recognition of control, stability improvements using flexible alternating
18 her contributions to the IEEE PES Singapore chapter. She current transmission system devices, power quality, power
19 served as the Chair of NUiS Women in science and electronic converters, and distributed power generation. His
20 engineering network and organized many events to promote work involves design, implementation, and operation of
21 young professionals especially young women in engineering AC/DC microgrid with interfacing converters that use digital
22 and technology. She served as the Chair of the IEEE Power signal processing, artificial intelligence techniques and other
23 and Energy Society, Singapore Chapter, from January 2019 to novel control methods.
24 December 2020. She is an Associate Editor of IEEE/CSEE
25 Journal of Power and Energy Systems. Her research interests
26 include renewable energy integration and applications in smart
27 grid, power electronics for aerospace, electric vehicle
applications, low carbon electrical energy systems, and power
28
electronic solutions to sustainability.
29
30 PROF. GAYADHAR PANDA
31 has received his Ph.D. degree from
32 the Utkal University in 2007 in
33 electrical engineering. He is
34 currently Professor in the
35 Electrical Engineering
36 Department, National Institute of
37 Technology, Meghalaya, India.
38 Dr. Panda has over 24 years of
39 teaching and research experience.
40 He has published more than 120
41 technical papers in national and international conferences
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59 VOLUME XX, 2017 For Review Only 1
60