IETE Technical Review

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/titr20

A Random Under-Sampling based Passive

Approach for Fast and Accurate Detection of
Islanding in Electrical Distribution System

Nimish Bhatt & Ashwani Kumar Chandel

To cite this article: Nimish Bhatt & Ashwani Kumar Chandel (2021): A Random Under-Sampling
based Passive Approach for Fast and Accurate Detection of Islanding in Electrical Distribution
System, IETE Technical Review, DOI: 10.1080/02564602.2021.1976291

To link to this article: https://doi.org/10.1080/02564602.2021.1976291

Published online: 19 Sep 2021.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=titr20
IETE TECHNICAL REVIEW
https://doi.org/10.1080/02564602.2021.1976291

A Random Under-Sampling based Passive Approach for Fast and Accurate

Detection of Islanding in Electrical Distribution System
Nimish Bhatt and Ashwani Kumar Chandel
Department of Electrical Engineering, National Institute of Technology, Hamirpur 177005, Himachal Pradesh, India

ABSTRACT KEYWORDS
In this article, a robust random under-sampling boosting (RUSB) islanding detection technique Distributed generation;
evolved from the machine learning (ML) approaches is proposed. Unlike the conventional passive Islanding detection; Modal
islanding techniques, the ML algorithms are superior in performance owing to their better dynamic current transformation; RUS
behaviour. However, a major challenge arises when non-islanding events are more as compared boosting; Skewed dataset;
Wavelet transform
to islanding events. This leads to skewness in dataset resulting in improper classification, and poor
accuracy are often observable. To address these issues, the proposed algorithm which relies on the
under-sampling approach can easily identify the dominating cases in the available dataset such that
overall detection accuracy improves with a better dynamic response. Also, the modal transforma-
tion employed at the point to measurement is a rescuer for reducing the dataset by decomposing
the three-phase current signal into single-phase current signal (also known as modal current (MC)
component). Therefore, the feature extraction is carried out from MC’s information by employing
wavelet transformation to detect islanding conditions quickly. The extensive numerical simulations
are carried out for a standard IEEE 15 bus distribution network to assess the improvement achieved
in the accuracy of classification and the ability to accurately detect the islanding condition in the
event of large number of non-islanding test cases with a fast dynamic response.

1. INTRODUCTION
(2) Reconnection of feeder in out of phase state due to
An ever increase in electrical power demand over the auto recloser operation of circuit breakers.
whole world has over-burdened the existing power sys- (3) Potential threat to the field personnel during repair
tem network. Further, due to the limited supply of fos- works.
sil fuels and rising environmental concerns, renewable-
based distributed generation units (DGs) have become The issues mentioned above indicate that the utility
a unanimous choice to cater to the increased demand. operator must detect the islanding condition as fast
The DG is a small to medium-scale electrical power as possible and quickly isolate the DG units from the
source integrated into the distribution system (DS). distribution network. The IEEE 1547–2018 standard
Apart from providing power, the DG assimilation also states that the DG must disconnect within 2 s from
inducts numerous advantages like improved voltage pro- the islanded network [5]. The islanding approaches
file, reduced power loss, and operational costs [1,2]. classification is as follows: 1) Remote-based island-
However, along with their benefits, DGs introduce oper- ing method (RIDM) and 2) Local islanding method
ational issues such as a change in frequency, variation in (LIDM). The remote or communication method employs
voltage, and islanding. When a DS segment is discon- a communication channel to keep vigilance on the
nected from the substation grid but remains energized relays connected to the substation and the DG [6].
from the DG unit, it is known as an islanding condition. Note that the RIDM has zero non-detection zones
Subsequently, the frequency and voltage of a DS are no (NDZs), faster detection time, and high accuracy. Despite
longer effectively controllable by the substation grid [3,4]. these advantages, the complicated infrastructure and
Thus, this inadvertent island formation imparts probable cost involved in implementing the RIDMs make them
hazards like: less attractive [7,8]. On the other hand, the LIDMs
employ the various parameters associated with the
(1) Degradation of power quality due to mismatch of DG at its location. The classification of the LIDMs
power in an islanded network giving rise to voltage is, namely, the active and passive islanding detection
and frequency deviation. methods.

© 2021 IETE
2 N. BHATT AND A. K. CHANDEL: A RANDOM UNDER-SAMPLING BASED PASSIVE APPROACH

The active method employs an arbitrary signal injec- transform (FT), S-transform (ST), and wavelet trans-
tion into the DS, followed by monitoring the parameters form (WT). Among these methods, the FT-based IDMs
like voltage and frequency using a feedback technique [24,25] rely on changing the input time-domain signal
[9–12]. Later on, few researchers introduced modified into frequency-domain for the feature extraction used for
active IDMs, for instance, active ROCOF [13], active slip islanding and non-islanding detection. Generally speak-
frequency technique [14]. Besides small NDZ, the active ing, the major collective drawback is a poor estimation of
IDMs suffer from several issues such as unwanted tran- the spectrum and low frequency resolution.
sient behaviour, instability in the system, and deteriora-
tion of power quality (PQ). It is known that introducing Nevertheless, the WT tool may overcome these issues
two perturbations at a fixed interval, irrespective of the by providing a multi-resolution analysis of the signal.
system’s prevailing situation, allows forecasting an island- Furthermore, by acquiring time localization of a fault
ing situation. A significant challenge is if the islanding event and monitoring the DG output power, the island-
event occurs during this interval, the detection shall hap- ing event could be detected [26,27]. Another prefer-
pen only after the subsequent perturbation interjection. ence could be obtaining the singular entropy from WT
to identify the islanding and non-islanding conditions
On the other hand, the passive IDMs rely on the surveil- [28]. However, in [28], the essential critical non-islanding
lance of the system’s critical parameters such as voltage, cases require further investigation to evaluate the pro-
frequency, and current. The passive IDMs includes: the posal’s maximum potential. Interestingly, the threshold
rate of change of voltage/frequency (ROCOV/ROCOF), setting approach-based WT-IDM can also be adopted
under- frequency/over-frequency (UF/OF), under- [29,30]. However, additional efforts are required from
voltage/over-voltage (UV/OV) [15,16] and unbalance in the designer’s point of view to obtain an exact threshold
voltage magnitude [17]. Further, some new passive IDMs value.
were introduced in the research community. The authors
in [18] have utilized the ROCOV combined with the As discussed earlier, the passive islanding methods are
rate of change of reactive power [ROCORP]. The IDM more flexible and cost-effective than the well-known
has been employed only for synchronous-based DGs. techniques. The widely adopted digital signal processing
ROCOV is also employed in [19] to identify the island- techniques must acquire the abilities to assist and aid the
ing conditions. An index-based IDM that focuses on the passive islanding methods. The artificial intelligence (AI)
rate of change of resistance regarding angular frequency and machine learning (ML) based classifiers are, there-
is proposed [20]. The method requires an appropriate fore, gaining interest in accurately detecting the islanding
threshold setting to detect the islanding condition. The and non-islanding conditions and overcoming the prob-
simplicity associated with the methods above and less lems of threshold settings. The hybrid technique based
sensitivity towards the system parameters allows for more on the decision tree (DT) classifier and the WT helps
robustness and ease of implementation. Unfortunately, in detect the islanding conditions by observing the transient
a small power balance between the load and the DG, there current and voltage signals [31]. Note that DT, being a
is a significantly large NDZ. weak learner, is prone to high classification error for a
complex dataset; therefore, a probabilistic neural network
Therefore, hybrid IDMs were developed, which amalga- (PNN) could be employed [32]. To better understand
mates the passive and active IDMs. Conventionally, the the classification accuracy, the PNN with DT and radial
passive IDMs are used to detect the various conditions basis function (RBFN) is studied. Still, the classification
based on the threshold concept. If the passive IDMs are accuracy of islanding cases is less than the non-islanding
unable to identify any ambiguous situation, then active issues.
IDMs are employed. In this manner, the constant inter-
jection of perturbation in the system is avoided, and Nevertheless, the quest for higher accuracy with better
PQ deterioration is minimized [21,22]. Despite the gains detection time using ML approaches continues [33,34].
mentioned earlier, the practical application of these IDMs Often, the non-islanding events are more than the island-
is a less feasible solution [23]. ing events creating a bias for the classifier affecting its
overall performance [35,36]. To overcome this issue, the
To address the concerns mentioned above and improve Hilbert-Huang transform (HHT) and an extreme learn-
the accuracy of detection, the up-gradation of passive ing machine (ELM) are hybridized [37]. Nevertheless, an
methods with the signal processing techniques (SP) is adaptive boosting (ADAB) technique [38] with higher
found very attractive. Some of the signal processing accuracy may aid in cutting down the detection time.
tools which are widely utilized are as follows: Fourier However, the possibility of nuisance tripping may affect
 N. BHATT AND A. K. CHANDEL: A RANDOM UNDER-SAMPLING BASED PASSIVE APPROACH 3

the overall accuracy of ADAB. An optimized random for- The various sections are as follows: The proposed
est (RF) combined with down sampling empirical mode methodology emphasizes the importance of modal cur-
decomposition (DEMD-ORFA) IDM proves to be a bet- rent transformation and WT in section 2. In Section
ter approach in terms of accuracy and detection. How- 3, the discussion regarding the implementation of the
ever, the DEMD-ORFA method is not suitable for real- RUSB algorithm is presented. Various simulations com-
time purposes owing to the implementation complexity prising of islanding and non-islanding events are dis-
[39]. A micro-phasor measurement unit and RF-based cussed in section 4. In Section 5, the performance
IDM have been projected to detect the islanding condi- analysis of the proposed IDM with other existing
tions. The procedure is quite accurate but requires a dedi- techniques is carried out. The summary and conclu-
cated hardware setup [40]. In view of the shortcomings of sion drawn from the proposed work are discussed in
the ML schemes as mentioned above, hybridization of the section 6.
WT with random under-sampling boosting (RUSB) ML
technique is proposed in the current proposal. Unlike the
conventional schemes, the proposal focuses on improv- 2. PROPOSED METHODOLOGY
ing the overall accuracy. In addition, the problem of The proposed IDM is presented in a flow chart in this
accurately separating the islanding events from the non- section, as exemplified in Figure 1. The usage of mathe-
islanding events is carried out. Often the non-islanding matical tools such as modal current (i.e. three-phase to
events are considered to be part of the dominant class; single-phase current transformation) and wavelet trans-
thus, the RUSB technique helps to identify the islanding form (WT) is discussed. It is to be noted that feature
event without any loss of generality. Hence, the classi- extraction is possible when the modal current compo-
fication accuracy of the proposed islanding method is nents are decomposed by WT. These modal components
improved for both the non-islanding and the islanding are derived from the input current (three-phase) signal
events. The advantages of performing hybridization of obtained from the target DG.
the ML schemes help recognize the false islanding events
in an efficient manner. Thus, this enhances the protec-
tion of the specialized electronic equipment and lives
of the crew workers. To ease the problem of handling a
three-phase current signal, the modal current (MC) is a
rescuer for the purpose of processing only a single-phase
current signal. Furthermore, the WT-based decomposed
current signal helps focus on the following six features:
energy, entropy, kurtosis, mean, skewness, and standard
deviation for the training purposes of the RUSB tech-
nique. Nevertheless, various islanding events encompass
the deficit and excess misalliance of power, considering
an exhaustive IDM performance evaluation. For this pur-
pose, the impact of severe islanding events is included in
the data set of the proposed work with the following key
potential highlights:

(1) The islanding events are classified with an improved

degree of accuracy.
(2) The classification of dominating non-islanding
events is well handled with the proposed approach.
(3) Despite the preferred hybridization, the use of MC
and minimum features, ease in implementation is
found to be attractive.
(4) The data set skewness is well handled with the RUSB
algorithm that helps in improving the overall accu-
racy of the proposed islanding scheme.
(5) Finally, an improved islanding detection time is
achievable when compared to the conventional ML
techniques suitable for the DS. Figure 1: Flow diagram of the proposed IDM
4 N. BHATT AND A. K. CHANDEL: A RANDOM UNDER-SAMPLING BASED PASSIVE APPROACH

2.1 Current Acquisition and Modal Current the faults and protection of transmission lines [46,47]
as well as in islanding detection [48]. The conversion
The implementation of the proposed IDM begins with an
matrix applied to obtain the modal components of the
input parameter which is captured to estimate the cur-
three-phase signal is given as:
rent state of the DS. The basic parameters that are easily
⎡ ⎤ ⎡ 3 ⎤⎡ ⎤
measurable without any error are current and voltage. Q1 /2 0 −3/2 P1
In the present article, the instantaneous three-phase cur- 1
⎣ Q2 ⎦ = ⎣ 1/ −1 1/ ⎦ ⎣ P2 ⎦
2 2 (1)
rents are used as an input parameter. The three-phase 3
Q3 1 1 1 P3
currents are measured from the junction of common cou-
pling (JCC) of DG and DS. These junctions are the buses where, P1 , P2 , and P3 correspond to the matrix of pha-
at which the DGs are mounted (here, bus 6 and bus 4). sor values of the three-phase system and Q1 , Q2 , and
The obtained currents are then converted into per-unit Q3 denotes the modal components with 1, 2, and 3
values, making the projected IDM applicable for vari- as three different modes, namely, aerial mode-1, aerial
ous systems. Since these currents act as input for WT mode-2, and the ground mode. Additionally, the authors
therefore, sampling of the input current is done. Broadly in [49,50] implemented a simplified form of modal trans-
the sampling frequency of a signal is selected as per the formation. The three phasor values are measured inde-
Nyquist principle which states that the sampling of a pendently, and a different coefficient scales each phasor.
signal must be carried out twice the rate of the signal Since scaling operation on a signal does not affect the sig-
frequency. However, it is difficult to adjudge the exact nal’s phase and frequency information, no loss of infor-
frequency of the transient components occurring in the mation occurs. Further, these scaled phasor values are
signal. If the signal is sampled at a low-frequency rate, the bunched together, resulting in a single-phase quantity
latent information present in the signal may be lost. Fur- containing all the characteristics of the three phasor val-
ther, the low sampling rate of the signal may hamper the ues. This version has been implemented too, stator faults
reconstruction of the signal within the embedded con- in induction machines [50], faults in the distribution sys-
troller. Hence the signal is sampled in the range varying tem [51], and islanding detection [38,52] as well. The
from kHz to MHz for preserving the maximum possible Modal Current (MC) is obtained by the linear amalga-
information. Apart from the IDM scheme, the controller mation of the three-phase using the following equation:
present in the inverter also has subsections such as MPPT,
grid synchronization, and PWM blocks. These subsec-
MC = EIr + FIy + GIb (2)
tions also operate at a high frequency [41]. Hence, a
higher sampling frequency becomes a reasonable choice where, Ir , Iy and, Ib correspond to the respective phase
for sampling the input signal. In the present article, the currents, and E, F, and G imply the corresponding
authors have selected the sampling frequency of 5MHz modal coefficients. The values of E, F, and G are 1, 2,
as per the ref [42,43]. and −3, respectively [38,50–52]. Since each phase cur-
rent has been assigned with a different modal coeffi-
The task of the IDM is not only to identify islanding con- cient, the direct addition or subtraction of any of the
ditions but also to recognize every non-islanding event phases is debarred. This method also ensures that all
to avoid nuisance tripping. The non-islanding events also the possible transient conditions are reflected in the
include faults which comprise of diverge group of phases transformed MC [48]. In the present article, the authors
and ground. Therefore, obtaining superior performance have employed the aforementioned simplified version of
of the IDM, the inclusion of all the three-phase of the modal transformation.
current becomes a compulsive task. However, handling
extensive data and storage issues may cause real-time 2.2 Wavelet Transform (WT)
implementation complexities. This problem is addressed
using Modal current transformation. The implementation of SP techniques is done to ana-
lyze and extract the latent attributes of the input sig-
Modal transformation or modal current (MC) is a math- nal under any PQ issue. In general, PQ issues have a
ematical tool that is similar to Clarke’s transforma- non-stationary nature; therefore, an effective SP tech-
tion (CT) [44,45]. MC disintegrates the phasor values nique that can obtain information of frequency and time
into corresponding mode values [43]. The basic concept domains is required. Numerous researchers have recom-
behind this approach is to consider a three-phase sys- mended the usage of WT [53,54]. As an islanding event
tem as three independent circuits. Several researchers is also a PQ event, WT has been used to analyze it. The
have employed MC in the detection of accurately detect wavelet transforms act as a powerful mathematical tool
 N. BHATT AND A. K. CHANDEL: A RANDOM UNDER-SAMPLING BASED PASSIVE APPROACH 5

wavelet coefficients aids in an easy splitting of noise com-

ponents and signal utilizing the concept of thresholding
[55,56]. The selection of an appropriate mother wavelet
is affected by the shape of the signal captured for analy-
sis. It is mandatory to choose the mother wavelet, which
has maximum resemblance with the signal. This leads
to perfect decomposition with optimum decomposition
levels, which reduce the computational burden. There
must be orthonormality between the selected mother
Figure 2: Wavelet Transform Decomposition wavelet because:

(1) The orthonormality assures that the energy of the

for analysing non-stationary signals. WT yields a sound signal remains conserved, which implies that no
localization in frequency at low frequencies and good transient component is lost while transforming the
localization in time at high frequencies. For discrete-time signal.
signals, the discrete wavelet transform (DWT) is applied. (2) WT provides a multi-resolution analysis (MRA),
The DWT of a function is as follows: which offers different scales to extract out low and
 high-frequency details of the signal.
∗
DWTf (s, u) = f (n)φs,u (n) (3) (3) The inner product between the signal and orthonor-
n mal basis results in wavelet coefficients.
where, φs,u stands for the mother wavelet and is shown
A variety of mother wavelets like Morlet, Haar, Sym-
as:
metric Mexican Hat are available to use. However, to

1 n − ud0 c0u analyse fast and impulsive transients, researchers have
φs,u (n) =  s φ (4)
c0 c0s recommended Daubechies Haar, and Morler. Daubechies
wavelet family is considered to be one of the most appro-
which results in decomposition details and finite impulse priate wavelet families for the analysis of transient con-
response (FIR) low pass filter (LPF) J(n), which leads to ditions and detection of islanding conditions in power
decomposition approximation, as shown in Figure 2. The systems, as explored in [57–61]. Since this paper utilizes
downscaling process is applied to the signals j1 (n) and current signals to prepare input features, Daubechie’s 5
k1 (n)obtained from K(n) and J(n) to eliminate the infor- (db5) with the order of layer of decomposition has been
mation redundancy. Similarly, the obtained signal j1 (n) is chosen for analysis of the MC [54] shown in the Figure
subsequently disintegrated into j2 (n) and k2 (n)by passing 3. It is inferable from the figure that this mother wavelet
it through the digital filters K(n) and J(n). The coefficients has a sharp transition makes it suitable for decompos-
of the two filters are dependent on the mother wavelet. ing the latent attributes of the input signal. Six features,
This filtration trend continues depending on the level namely, energy, entropy, kurtosis, mean, skewness, and
of decomposition. It is inferred from Figure 2 that the standard deviation, are extracted from WT. These fea-
computational time is proportional to the decomposition tures are subjected to the RUS Boosting algorithm to
level. label the several events as islanding and non-islanding
event.
2.3 Selection of Mother Wavelet and Feature
Extraction
Feature extraction is done for the identification of the 1.5

classification of islanding and non-islanding conditions. 1

These features are derived from the output of the WT. 0.5

Since these features directly impact the performance of 0

the classifier, therefore, it becomes a mandatory task to

-0.5
select a suitable mother wavelet to obtain accurate results.
-1
The mother wavelet imparts a key role as it helps central-
ize the maximum portion of the power of an input signal -1.5
0 1 2 3 4 5 6 7 8 9

in a small quantity of wavelet coefficients. This centraliza-

tion of the power of the input signal in a small quantity of Figure 3: Daubechies 5 mother wavelet
6 N. BHATT AND A. K. CHANDEL: A RANDOM UNDER-SAMPLING BASED PASSIVE APPROACH

3. RUS BOOSTING bagging ensemble technique relies on parallel learning

of weak correlative learners independently and further
The problem of classification of a skewed dataset is a
assimilate these learners. Boosting technique also utilizes
chief issue for data mining engineers. A dataset is known
the weak learners, but the learning pattern is sequen-
to be a skewed dataset when the majority of the dataset
tial, i.e. the training of the learners is done adaptively.
belongs to one particular class. Conventional classifiers
This technique is highly suitable for skewed datasets. On
tend to overfit the class with more dominance in the
the other hand, the stacking technique employs diversi-
dataset showing obliviousness to the minority class [62].
fied learners trained in a parallel manner [63,64]. RUS
In the present study, out of 707 cases, 41 belong to island-
Boosting classifier, as the name signifies, falls under
ing cases (minority class) and 667 instances correspond
the category of boosting ensemble technique. The pro-
to non-islanding conditions (majority class) as shown in
posed RUS Boosting classifier relies on the assimilation
Tables 1 and 2, respectively (discussed in the subsequent
of numerous weak or poor classifiers leading to an effec-
section). Thus, the obtained dataset is skewed in nature.
tive classifier. Notably, an effective classifier possesses
enhanced categorization ability with reference to the
Therefore, scopes exist to develop a classifier that is
weak classifiers. The proposed classifier applies random
immune to the skewness present in the dataset and classi-
under-sampling (RUS), which randomly removes exam-
fies the dataset without any biases. A conventional classi-
ples from the majority class from the subjected dataset.
fier is termed a weak learner when its categorization has a
Figure 4 depicts the scatter plot of one of the statistical
marginally improved accuracy than the random classifi-
features (energy) from the dataset. The abscissa denotes
cation of the dataset. This poor performance by the weak
the number of samples, while the ordinate represents the
classifier may be due to the presence of a single learner
value of energy. The scatter plot comprises islanding and
model. As a result, they face statistical issues, represen-
non-islanding classes, where the yellow dots belong to the
tation issues (skewed dataset), and computational issues.
islanding class, and the green dots correspond to the non-
Instead of using the conventional classifiers, committee-
islanding type. As it is visible in the Figure 4 that the non-
based or multiple learners results in better classification
islanding class heavily dominates the islanding class. This
for the skewed dataset. These classifiers or learners are
results in overfitting of the dominant class and becoming
better known as ensemble classifiers. The ensemble clas-
insensitive towards the minority class. Since the minority
sifiers consist of numerous learners derived from the
class belongs to the islanding condition, the classification
training dataset which are aided by a base learner. Hence,
of the islanding condition becomes poor. This problem
ensemble learner has multiple learners, which handles
is tackled by using the concept of under-sampling. The
the three issues described above quite easily.
RUS Boosting algorithm applies the under-sampling on
the non-islanding cases and randomly removes the sam-
There are three types of techniques available in ensemble
ples from the dataset. This process establishes the balance
classifiers, namely bagging, boosting, and stacking. The

Table 1: Various islanding test cases

Issue Parameter Alterations in parameter Simulated cases
Islanding Mismatch of active power amid load and DG 105% to 150% with an incremental rate of 5%. 10
95% to 50% with a decremental value of 5% 10
Mismatch of reactive power amid load and DG 105% to 150% with an incremental rate of 5%. 10
95% to 50% with a decremental rate of 5% 10
Perfect match of load and DG power 1
Sum of islanding cases 41

Table 2: Various non-islanding test cases

Issue Parameter Alterations in parameter Simulated cases
Non-islanding Sudden increment of load Increment in load from 50% to 150% with an incremental value of 5%. Total = 21 21×4 = 84
Variation of load at bus no.4,5,14, and 15. Total = 4
Increment of power factor of load Increment of load power factor (lagging) from 0.6–1.00 with an incremental value 4×21 ×4 = 336
of 0.1. Total = 4.
Increment in load from 50% to 150% with an incremental value of 5%. Total = 21.
Variation of load at bus no.4,5,14, and 15. Total = 4.
Faults Variety of faults: LLLG LLG, LL, and SLG. Total = 7. 7×5 ×7 = 245
Location of faults: Bus 4, Bus 4-5, Bus 4-14, Bus 4–15 and Bus 3-4. Total = 5
Variation in fault resistance (in ) = 3,5,10,15,20,25,30. Total = 7.
Induction Motor Starting Nominal values: 200HP, 575V, 60Hz, 1785 R.PM, delta connected. Total = 1 1
Capacitor Switching 5 kVAr, 25kV, 60Hz, delta 1
Sum of non-islanding cases 667
 N. BHATT AND A. K. CHANDEL: A RANDOM UNDER-SAMPLING BASED PASSIVE APPROACH 7

Figure 4: Graphical depiction of the energy obtained from the

simulated dataset

in the dataset, which results proper classification of both

islanding and non-islanding cases.

The detailed mechanism of the proposed algorithm is

demonstrated in Figure 5. The algorithm begins with ini-
tializing the weights of the given instance by 1l , where
l is the number of instances present in the dataset. In
other words, the weight assigned to each instance present
in the dataset is derived by dividing the instance value
by the total number of instances. The iterative training
of Weak Learners is done in the following steps. The
next step performs the task of under-sampling of major-
ity class instances till P% of the newly obtained dataset,
Cx correspond to minority class. Mx belong to the new
weight of Cx . Further, Mx and Cx are subjected to a poor
learner that results in weak hypotheses dx . The origi-
nal training dataset, Cand the weight distribution Mx are
used to obtain the pseudo loss ∈t in step 2.4. The next step Figure 5: A detailed description of the proposed RUS Boosting
deals with the calculation of the weight update parameter algorithm
β. Further, the weight distribution for the next iteration,
Cx+1 is updated and normalized. After X iterations, the
final hypothesis, G(x), is obtained as a weighted vote of
weak hypotheses [65].

4. TEST SYSTEM AND SIMULATION

The test conditions required to prepare the dataset for
training the classifier are simulated on IEEE 15 bus
DS in MATLAB environment with few alterations [66].
The distribution substation comprises of 25 MVA, 60Hz,
120/25 kV transformer, and the single line diagram of
the test setup is portrayed in Figure 6. Furthermore, two
DGs (photo-voltaic based);1 MW (sited at bus no. 6)
and 500kW (sited at bus no. 4) are integrated in the DS
[67]. To have a comprehensive assessment of the pro-
jected IDM, the present paper consists of 707 events, out
of which 667 are non-islanding events, and the rest are
islanding events. Tables 1 and 2 correspond the test cases
Figure 6: Single line diagram depicting the IEEE 15 bus distribu-
of various islanding and non-islanding conditions Fig- tion system
ures 7–14 portray the various islanding the non-islanding
8 N. BHATT AND A. K. CHANDEL: A RANDOM UNDER-SAMPLING BASED PASSIVE APPROACH

Figure 7: Three-phase current along with the MC (p.u) wave-

Figure 10: Three-phase current along with the MC (p.u) wave-
forms against time (s) during 15% active power (positive) mis-
forms against time (s) during fault condition (three-phase at bus 3)
match under islanding condition

Figure 11: Three-phase current along with the MC (p.u) wave-

Figure 8: Three-phase current along with the MC (p.u) wave- forms against time (s) during 10% positive change in the load
forms against time (s) during 5% of reactive power (negative)
mismatch under islanding condition

Figure 12: Three-phase current along with the MC (p.u) wave-

forms against time (s) during the start-up of induction motor

Figure 9: Three-phase current along with the MC (p.u) wave-

forms against time (s) during 0% mismatch of power load and DG 4.1 Islanding Event
under islanding condition An exhaustive dataset comprising of 40 islanding events
has been considered for detecting the islanding condi-
tion. These cases contain the variation in active and reac-
conditions. The figures consist of two waveforms; the tive power, including a perfect match of the power amid
upper waveform depicts the three-phase current condi- the load and the DG. All the islanding conditions are
tions, and the lower waveform corresponds to the MC simulated by opening the circuit breaker (CB) of the sub-
equivalent to the three-phase current. The simulation has station grid at 0.5 s. The details of all the 41 cases are men-
been done in MATLAB/Simulink R2018a. tioned in Table 1. Amongst the aforementioned number
 N. BHATT AND A. K. CHANDEL: A RANDOM UNDER-SAMPLING BASED PASSIVE APPROACH 9

becomes difficult. Under such conditions, if the island-

ing is left undetected, it can also harm personnel and
equipment.

4.1.3 Perfect Match

The last case belongs to a perfect match of power of load
and DER. It is visible from the Figure 9 the three-phase
waveform and the equivalent MC signal; there is a very
minute variation in one of the phases at 0.5 sec (visible in
blue) due to the impedance of the distribution line con-
necting between the load and the DG. Apart from this,
Figure 13: Three-phase current along with the MC (p.u) wave- there is no variation in any of the three phases and their
forms against time (s) during switching of delta connected capac-
corresponding MC. Under such conditions, the forma-
itor
tion of NDZ takes place, and the accuracy of the IDM
is reduced. Hence, such cases are very crucial in demon-
strating the competence of the proposed algorithm.

4.2 Non-islanding Events

In addition to islanding cases, 667 non-islanding events
are also taken into consideration for assessing the reliabil-
ity and accuracy of the proposed IDM. Some typical non-
islanding scenarios like several types of faults, including
single line to ground fault, line to line fault, double line to
ground fault are simulated. Different fault locations and
different fault resistances have been considered for testing
Figure 14: Three-phase current along with the MC (p.u) wave- the algorithm against the nuisance trip. Table 2 shows the
forms against time (s) during variation of power factor of load at
details of all the non-islanding events that are simulated.
bus 5
All the non-islanding events are initiated at 0.5 s. Total
five events are shown in the article, which are discussed
of cases, three critical test conditions are presented in the in the following subsections.
following subsections.
4.2.1 Fault
4.1.1 Active Power Mismatch The most common non-islanding event is a fault which
The first case deals with a 15% positive mismatch of has been shown in Figure 10. The figure depicts a three-
power amid load and the DG, which is depicted in Figure phase fault with a fault resistance of 10 ohms occurring at
7. The islanding occurs due to the occurrence of fault bus no. 3. After the initialization of the fault, a large cur-
resulting in the isolation of the DS. Since the demand of rent flow is observed in the DS. The impulsive rise is due
the load is higher than the power generated by the DG, to the DG supplies power to this fault. A total of 245 cases
the voltage suddenly drops, and the associated current corresponding to fault conditions where the variation of
shall increase from its nominal value, which is also vis- fault resistances, types of faults, and location are included
ible in the figure. All three phases have a higher peak at in preparing an exhaustive dataset.
0.5 s, and the corresponding MC also depict the same.
This increment in current can result in severe damage to 4.2.2 Variation of Load
the DER and the other components of the DS. Another non-islanding event that has the ability to mal-
operate DG relay is a sudden change in the load. The
4.1.2 Reactive Power Mismatch change in the load is done from 50% to 150% in steps
The following case corresponds to the islanding condi- of 5%. A total of 84 cases have been simulated under this
tion with 5% negative reactive power mismatch amid the condition. The variation in load power factor is done to
load and the DG as portrayed by Figure 8. Since the mis- maximize the number of cases and minimize the chances
match is meagre, there are the smallest variations in the of nuisance tripping. Figure 11 shows a sudden change in
three-phase current and MC, the detection of islanding load with 10% positive change that is 110% of its nominal
10 N. BHATT AND A. K. CHANDEL: A RANDOM UNDER-SAMPLING BASED PASSIVE APPROACH

value. It is visible in the Figure 11 that at 0.5 s, there is sig- visible in both three-phase current and corresponding
nificant variation present in three-phase current, which is modal current waveform. These transients are capable of
depicted by the MC. mal-operating the relay connected to the DG.

4.2.3 Starting of Induction Motor

5. PERFORMANCE OF CLASSIFIER
The start-up of the three-phase induction motor at 0.5 sec
has been shown to Figure 12. When the motor is started, The performance and comparative analysis of the RUS
the setting to fault is also visible in both the three-phase Boosting algorithm have been done with five classi-
and MC waveform. This impulsive current can lead to the fiers, namely, Logistic Regression (LR), Support Vector
nuisance tripping of the up of the flux requires a high Machine (SVM), Decision Tree (DT), K-Nearest Neigh-
amount of current. This current is of a highly lagging bour (KNN), and Linear Discriminant Analysis (LDA).
nature which results in poor power factor. Hence, the The dataset has been bifurcated with 65% for training and
motor start-up causes a slight distortion in the current the rest for the testing classifier. The comparison is made
waveform. Apart from this, the current value increases based on four parameters, namely, confusion matrix, sen-
from its nominal value. These two events are also visible sitivity, specificity, and accuracy. The table that describes
in the three-phase current and corresponding MC. the confusion present in making predictions is termed
a confusion matrix. It encompasses a distinct grouping
4.2.4 Switching of Capacitor of incorrect and correct predictions. There are four ele-
Figure 13 shows the switching of the delta-connected ments in the confusion matrix, namely, true positive, false
capacitor. Since the capacitor injects the leading reac- positive, true negative, and false negative. All the terms
tive power, therefore, the harmonics are introduced in are defined with reference to the presented article’s clas-
the DS. The harmonics are present in the current wave- sification problem (islanding and non-islanding). Con-
forms, which can affect the operation of the DG relay. fusion matrices corresponding to all the six classifiers are
Both the current waveforms (three-phase and MC) in shown in Figure 15. True-positive (denoted by NDtrue ) is
Figure 13. replicate the situation described above when the number of non- islanding cases that are correctly cat-
the capacitor is connected to the DS. egorized as non-islanding by the classifier. False-positive

4.2.5 Variation of Power Factor/Quality Factor

Another non-islanding event that has been taken into
account is the variation of power factor of the load or
quality factor of the load. Since power factor and qual-
ity factor are related to each by the following expression:

XL
QL = (5)
Z × cos ϕ
where, QL is the quality factor, XL is the load reactance,
Z is the load impedance, and cos ϕ is the power factor of
the load.

It is observable from the above expression that the quality

factor and power factor are inversely related. Therefore,
the variation of the power factor of load leads in variation
of the quality factor of the load. In the present study, 336
cases of power factor variation (quality factor variation)
are considered (mentioned in Table 2). Figure 14 corre-
sponds to the power factor (quality factor) variation of
the load connected at bus 5. The reactive power demand
suddenly increases, which degrades the power factor. The
power factor suddenly drops from 0.81–0.6, and the qual-
ity factor changes from 0.71–1.04. This variation causes
a change in reactive power requirement, which causes Figure 15: Confusion matrix corresponding to (a) DT, (b) KNN, (c)
transients in the current waveform. The transients are RUS Boosting, (d) SVM, (e) LR, and (f) LDA
 N. BHATT AND A. K. CHANDEL: A RANDOM UNDER-SAMPLING BASED PASSIVE APPROACH 11

(marked by NDfalse ) is defined as the count of non- is also called the overall accuracy of the IDM, which is
islanding cases that are miscategorized as islanding by shown below:
the classifier. True-negative (denoted by IDtrue ) is the NDtrue + IDfalse
count of the islanding cases that are classified as island- η=
NDtrue + NDfalse + IDTrue + IDfalse
ing. False-negative (denoted by IDfalse ) is the count of
the islanding cases that are labelled as non-islanding. All three indices for the six classifiers are calculated and
The different ratios of these elements are used to calcu- mentioned in Table 3. It is inferable from Figure 15 that
late the remaining three performance indices. Sensitivity all the classifiers are able to classify the non-islanding
(denoted by σ ) can be termed as the accuracy of non- events with high accuracy. However, the most crucial
islanding condition detection, which is calculated in the performance index in the present problem is specificity.
following manner: Since specificity deals with classification accuracy under
islanding conditions, it should be as high as possible. If
NDTrue
σ = (6) any islanding event is not detected, it shall lead to haz-
NDTrue + NDFalse ardous conditions and even pose death threats to utility
Specificity (denoted as ρ) can be labelled as the accuracy members. As the dataset is skewed in nature, all five clas-
of islanding condition detection, which is computed as: sifiers (DT, LDA, SVM, KNN, and LR) show maximum
sensitivity and relatively poor specificity. On the con-
IDTrue trary, random removal of majority class examples com-
ρ= (7)
IDTrue + IDFalse bined with boosting the weak learner leads to superior
performance for the RUS Boosting classifier. The sensi-
Accuracy or accuracy score (denoted by η) is the ratio tivity, accuracy, and specificity for the RUS Boosting are
of the number of correct predictions (islanding and non- 99.57%, 99.59%, and 100%, respectively.
islanding cases) to the total number of input samples. It
To avoid any catastrophic condition, it is vital to discon-
Table 3: Comparative Analysis of the all the classifiers nect the DG after islanding. Therefore, the minimum
Classifier Accuracy Sensitivity Specificity detection time is crucial for the development of IDM.
DT 98.78% 100% 78.57% The response time or the detection time of the IDM
KNN 99.59% 100% 92.85% includes the time elapsed in feature extraction and the
LR 97.16% 99.57% 57.14%
LDA 94.73% 100% 7.14% time taken by the classifier. Illustratively, it consists of the
SVM 98.78% 99.14% 92.85% time required to decompose the MC with the help of WT
RUS Boost 99.59% 99.57% 100%
along with the extraction of features. Similarly, the time

Table 4: Detection time of the proposed IDM for various events

Detection
Issue Parameter Alterations in parameter time (m-sec)
Islanding Mismatch of active power 110% of load 20.88
amid load and DG
120% of load 21.79
Mismatch of reactive power 90% of load 21.37
amid load and DG
80% of load 21.52
Non-islanding Capacitor Switching – 21.89
Motor-start-up – 21.24
SLG fault at bus 4 Fault resistance = 15 21.62
Fault resistance = 20 21.80
LLG at bus 4 Fault resistance = 15 22.11
Fault resistance = 20 21.87
SLG fault at bus 5 Fault resistance = 15 23.24
Fault resistance = 20 23.37
LLG at bus 5 Fault resistance = 15 22.08
Fault resistance = 20 22.44
Variation of load at bus 5 65% of the nominal load with power factor of 0.7 23.15
75% of the nominal load with power factor of 0.7 23.71
65% of the nominal load with power factor of 0.8 22.37
75% of the nominal load with power factor of 0.8 22.91
Variation of load bus 15 65% of the nominal load with power factor of 0.7 24.390
75% of the nominal load with power factor of 0.7 23.49
65% of the nominal load with power factor of 0.8 22.34
75% of the nominal load with power factor of 0.8 21.74
Average detection time 22.33
12 N. BHATT AND A. K. CHANDEL: A RANDOM UNDER-SAMPLING BASED PASSIVE APPROACH

Table 5: Comparative assessment of the proposed method- ORCID

ology with other prevalent methodologies
Nimish Bhatt http://orcid.org/0000-0002-6727-903X
IDM Technique Accuracy (%) Detection time (milli-second)
PAVMD [4] 99.20 30
SVM technique [33] 97.71 40
HHT-ELM [37] 99.09 Not available
HHT- Adaboost [38] 97.80 Not available REFERENCES
S-Transform [31] 97.20 26
Hybrid ST[31] 98.40 22 1. D. Sarkar, A. Kumar, and P. K. Sadhu, “A survey on
WT [31] 97.00 25 development and recent trends of renewable energy
DEMD-ORFA [39] 98.54 25.33 generation from BIPV systems,” IETE Tech. Rev, Vol.
WT-RUS Boost (proposed) 99.59 22.33
37, no. 3, pp. 258–80, 2020. doi:10.1080/02564602.2019.
1598294.

2. S. Sharma, and Y. R. Sood, “Microgrids: A review of status,

taken by the classifier corresponds to the time taken by technologies, software tools, and issues in Indian power
the RUS Boosting. Since the detection time shall be dif- market,” IETE Tech Rev. 2020. doi:10.1080/02564602.2020.
ferent for every event, it is not feasible to calculate the 1850367.
time for all 707 events. Therefore, 22 events, including all
3. N. Gupta, and R. Garg, “Algorithm for islanding detec-
types of islanding and non-islanding events, have been
tion in photo-voltaic generator network connected to low-
considered to estimate the average detection time. The voltage grid,” IET Gener. Transm. Distrib, Vol. 12, no. 10,
details of all 22 events have been shown in Table 4. As pp. 2280–7, 2018. doi:10.1049/iet-gtd.2017.1735.
per the obtained results, the average detection time of the
projected IDM is 22.33 milliseconds. Lastly, the compar- 4. T. Chakravorti, L. Priyadarshini, P. K. Dash, and B. N.
ison of the projected IDM with the other passive IDMs is Sahu, “Islanding and non-islanding disturbance detec-
tion in microgrid using optimized modes decomposi-
made, shown in the Table 5. Various IDMs, including sig- tion based robust random vector functional link,” Eng.
nal processing-based and AI-based, have been included Appl. Artif. Intell., Vol. 85, pp. 122–36, 2019. doi:10.1016/
for a fair comparison. It is explicit from the table that the j.engappai.2019.06.004.
projected IDM surpasses the other IDMs in accuracy and
results in the formation of smaller NDZ. 5. IEEE Standard for Interconnection and Interoperability
of Distributed Energy Resources with Associated Elec-
tric Power Systems Interfaces,” in IEEE Std 1547-2018
(Revision of IEEE Std 1547-2003), pp. 1–138, (2018).
6. CONCLUSION doi:10.1109/IEEESTD.2018.8332112.

An enhanced passive islanding scheme leveraging the 6. M. Kim, R. Haider, G. Cho, C. Kim, and C. Won, “Com-
benefits of the known digital signal processing tool and prehensive review of islanding detection methods for dis-
tributed generation systems,” Energies, Vol. 12, no. 5, pp.
an under-sampling ML approach for fast detection of 1–21, 2019. doi:10.3390/en12050837.
islanding conditions in DS is proposed. The proposed
algorithm can handle the dataset skewness and identify 7. S. K. G. Manikonda, and D. N. Gaonkar, “Comprehensive
the non-islanding events, i.e. including capacitor switch- review of IDMs in DG systems,” Smart Grid, Vol. 2, pp.
ing and the starting of an induction motor. It is possi- 11–24, 2019. doi:10.1049/iet-stg.2018.0096.
ble to reduce the computational complexity by employ-
8. S. Dutta, P. K. Sadhu, M. Jaya Bharata Reddy, et al., “Shifting
ing modal currents without losing information regarding of research trends in islanding detection method - a com-
the faulty conditions. Note that numerical results explic- prehensive survey,” Prot Control Mod Power Syst, Vol. 3, no.
itly demonstrate that the proposed IDM can identify 1, 2018. doi:10.1186/s41601-017-0075-8.
the islanding conditions even for a small power imbal-
ance amid the DG and the load. The proposed RUSB 9. H. Vahedi, and M. Karrari, “Adaptive fuzzy sandia
frequency-shift method for islanding protection of
technique eliminates the problem of class imbalance inverter-based distributed generation,” IEEE Trans. Power
present in the dataset. Thus, the islanding conditions Deliv, Vol. 28, no. 1, pp. 84–92, 2013. 10.1109/TPWRD.
are expected to be identified with 100% accuracy along 2012.2219628.
with a dynamic response time of 22.33 ms. Additionally,
the technique classifies the non-islanding event with an 10. D. Reigosa, F. Briz, C. Blanco, P. García, et al. “Active island-
ing detection for multiple parallel- connected inverter-
accuracy of 99.57%, reducing the probability of nuisance
based distributed generators using high frequency signal
tripping. The proposed method is, therefore, a suitable injection”, IEEE Energy Conver. Cong. Exposit. (ECCE),
choice for fast detection of islanding events under adverse Raleigh, NC, pp. 2719-2726, 2012. doi:10.1109/ECCE.
grid condition. 2012.6342534.
 N. BHATT AND A. K. CHANDEL: A RANDOM UNDER-SAMPLING BASED PASSIVE APPROACH 13

11. P. Gupta, R. S. Bhatia, and D. K. Jain, “Average absolute 22. A. Rostami, A. Jalilian, S. Zabihi, J. Olamaei, and E.
frequency deviation value based active islanding detection Pouresmaeil, “Islanding detection of distributed genera-
technique,” IEEE Trans. Smart Grid, Vol. 6, no. 1, pp. tion based on parallel inductive impedance switching,”
26–35, 2015. doi:10.1109/TSG.2014.2337751. IEEE Syst. J., Vol. 14, no. 1, pp. 813–23, Mar. 2020.
doi:10.1109/JSYST.2019.2923289.
12. M. Hamzeh, N. Rashidirad, and K. Sheshyekani, “A new
islanding detection scheme for multiple inverter-based dg 23. R. Bakhshi-Jafarabadi, J. Sadeh, J. de J. Chavez, and
systems,” IEEE Trans. Energy Convers, Vol. 31, no. 3, pp. M. Popov, “Two-Level islanding detection method for
1002–11, 2016. doi:10.1109/TEC.2016.2558631. grid-connected photovoltaic system-based microgrid with
small Non-detection zone,” IEEE Trans. Smart Grid, Vol.
13. P. Gupta, R. S. Bhatia, and D. K. Jain, “Active ROCOF relay 12, no. 2, pp. 1063–72, 2021. doi:10.1109/TSG.2020.30
for islanding detection,” IEEE Trans. Power Deliv, Vol. 32, 35126.
no. 1, pp. 420–429, 2017. doi:10.1109/TPWRD.2016.254
0723. 24. J. Kim, J. Kim, Y. Ji, Y. Jung, and C. Won, “An island-
ing detection method for a grid-connected system based
14. P. K. Ganivada, and P. Jena, “Active slip frequency based on the goertzel algorithm,” IEEE Trans. Power Elec-
islanding detection technique for grid-tied inverters,” IEEE tron, Vol. 26, no. 4, pp. 1049–55, April 2011. doi 10.
Trans. Ind. Informatics, Vol. 16, no. 7, pp. 4615–26, 2020. 1109/TPEL.2011.2107751.
doi:10.1109/TII.2019.2949009.
25. I. Kim, “Islanding detection technique using grid-harmonic
15. W. Freitas, W. Xu, C. M. Affonso, and Z. Huang, “Compar- parameters in the photo-voltaic system,” Proc. Energy, Vol.
ative analysis between rocof and vector surge relays for dis- 14, pp. 137–41, 2012. doi:10.1016/j.egypro.2011.12.908.
tributed generation applications,” IEEE Trans. Power Deliv,
Vol. 20, no. 2, pp. 1315–24, 2005. doi:10.1109/TPWRD. 26. A. Pigazo, V. M. Moreno, M. Liserre, A. D. Aquila, et al.,
2004.834869. “Wavelet-based islanding detection algorithm for single-
phase photo-voltaic (pv) distributed generation systems,”
16. H. H. Zeineldin, and J. L. Kirtley, “Performance of the Proc IEEE Int Symp Ind Electron, Vol. 56, no. 11, pp.
OVP/UVP and OFP / UFP method with voltage and fre- 2409–13, 2007. 10.1109/ISIE.2007.4374984.
quency dependent loads,” IEEE Trans. Power Deliv, Vol. 24,
no. 2, pp. 772–8, 2009. doi:10.1109/TPWRD.2008.2002959. 27. A. Pigazo, M. Liserre, R. A. Mastromauro, et al., “A.
wavelet-based islanding detection in grid-connected PV
17. S. I. Jang, and K. H. Kim, “An islanding detection systems,” IEEE Trans. Ind. Electron., Vol. 56, no. 11, pp.
method for distributed generations using voltage unbal- 4445–55, Nov. 2009. doi:10.1109/TIE.2008.928097.
ance and total harmonic distortion of current,” IEEE
Trans. Power Deliv, Vol. 19, no. 2, pp. 745–52, 2004. 28. A. Samui, and S. R. Samantaray, “Wavelet singular entropy-
doi:10.1109/TPWRD.2003.822964. based islanding detection in distributed generation,” IEEE
Trans. Power Deliv, Vol. 28, no. 1, pp. 411–8, Jan. 2013.
18. A. Rostami, H. Abdi, M. Moradi, J. Olamaei, and E. doi:10.1109/TPWRD.2012.2220987.
Naderi, “Islanding detection based on ROCOV and
ROCORP parameters in the presence of synchronous 29. C. Hsieh, J. Lin, and S. Huang, “Enhancement of
DG applying the capacitor connection strategy,” Electr. islanding-detection of distributed generation systems
Power Components Syst, Vol. 45, no. 3, pp. 315–30, 2017. via wavelet transform-based approaches,” Int. J. Electr.
doi:10.1080/15325008.2016.1250842. Power Energy Syst., Vol. 30, no. 10, pp. 575–80, 2008.
doi:10.1016/j.ijepes.2008.08.006.
19. S. Nikolovski, H. R. Baghaee, and D. Mlakic, “Islanding
detection of synchronous generator-based DGs using rate 30. H. K. Karegar, and B. Sobhani, “Wavelet transform method
of change of reactive power,” IEEE Syst. J, Vol. 13, no. 4, pp. for islanding detection of wind turbines,” Renew. Energy,
4344–54, 2019. doi:10.1109/JSYST.2018.2889981. Vol. 38, no. 1, pp. 94–106, 2012. doi:10.1016/j.renene.2011.
07.002.
20. R. Bekhradian, M. Davarpanah, and M. Sanaye-Pasand,
“Novel approach for secure islanding detection in syn- 31. N. W. A. Lidula, and A. D. Rajapakse, “A pattern recog-
chronous generator based microgrids,” IEEE Trans. Power nition approach for detecting power islands using tran-
Deliv, Vol. 34, no. 2, pp. 457–66, 2019. doi:10.1109/TPW sient signals—part I: design and implementation,” IEEE
RD.2018.2869300. Trans. Power Deliv, Vol. 25, no. 4, pp. 3070–77, Oct. 2010.
doi:110.1109/TPWRD.2010.2053724.
21. X. Chen, Y. Li, and P. Crossley, “A novel hybrid island-
ing detection method for grid-connected microgrids with 32. S. R. Samantaray, B. C. Babu, and P. K. Dash, “Proba-
multiple inverter-based distributed generators based on bilistic neural network based islanding detection in dis-
adaptive reactive power disturbance and passive crite- tributed generation” Electr Power components Syst, Vol.
ria,” IEEE Trans. on Power Electronics, Vol. 34, no. 9, pp. 39, no. 3, pp. 191–203, 2011. doi:10.1080/15325008.2010.
9342–56, Sep. 2019. doi:10.1109/TPEL.2018.2886930. 526986.
14 N. BHATT AND A. K. CHANDEL: A RANDOM UNDER-SAMPLING BASED PASSIVE APPROACH

33. B. Matic-Cuka, and M. Kezunovic, “Islanding detection for 44. A. J. Prado, J. P. Filho, S. Kurokawa, and L. F. Bovolato,
inverter-based distributed generation using support vec- “Modal transformation obtained from Clarke’s matrix -
tor machine method,” IEEE Trans. Smart Grid, Vol. 5, asymmetrical three-phase case,” IEEE Power Eng. Soc. Gen.
no. 6, pp. 2676–86, Nov. 2014. doi:10.1109/TSG.2014.233 Meet. PES, no. 2, pp. 1–6, 2007. doi:10.1109/PES.2007.
8736. 385636.

34. M. R. Alam, K. M. Muttaqi, and A. Bouzerdoum, “A 45. P. Torrez Caballero, E. C. Marques Costa, and S. Kurokawa,
multifeature-based approach for islanding detection of “Modal decoupling of overhead transmission lines using
DG in the subcritical region of vector surge relay,” IEEE real and constant matrices: influence of the line length,”
Trans. Power Deliv, Vol. 29, no. 5, pp. 2349–58, Oct. 2014. Int. J. Electr. Power Energy Syst, Vol. 92, pp. 202–11, 2017.
doi:10.1109/TPWRD.2014.2315839. doi:10.1016/j.ijepes.2017.05.006.

35. R. Azim, F. Li, Y. Xue, et al., “An islanding detection 46. L. M. Wedepohl, “Application of matrix methods to the
methodology combining decision trees and Sandia fre- solution of travelling-wave phenomena in polyphase sys-
quency shift for inverter-based distributed generations,” tems,” Proc. Inst. Electr. Eng, Vol. 110, no. 12, pp. 2200,
IET Gener. Transm. Distrib., Vol. 11, no. 16, pp. 4104–13, 1963.
2017. doi:10.1049/iet-gtd.2016.1617.
47. R. K. Aggarwal, D. V. Coury, A. T. Johns, and A. Kalam, “A
36. N. W. A. Lidula, and A. D. Rajapakse, “A pattern- practical approach to accurate fault location on extra high
recognition approach for detecting power islands using voltage teed feeders,” IEEE Trans. Power Deliv, Vol. 8, no.
transient signals — Part II: Performance evaluation,” IEEE 3, pp. 874–83, Jul. 1993.
Trans. Power Deliv., Vol. 27, no. 3, pp. 1071–80, 2012.
doi:10.1109/TPWRD.2012. 48. N. Perera, A. D. Rajapakse, and T. E. Buchholzer, “Isolation
2187344. of faults in distribution networks with distributed genera-
tors,” IEEE Trans. Power Deliv, Vol. 23, no. 4, pp. 2347–55,
37. M. Mishra, M. Sahani, and P. K. Rout, “An islanding 2008. doi:10.1109/TPWRD.2008.2002867.
detection algorithm for distributed generation based on
Hilbert–Huang transform and extreme learning machine,” 49. Y. M. Makwana, and B. R. Bhalja, “Experimental perfor-
Sustain Energy, Grids-Networks., Vol. 9, pp. 13–26, 2017. mance of an islanding detection scheme based on modal
doi:10.1016/j.segan.2016.11.002. components,” IEEE Trans. Smart Grid, Vol. 10, no. 1, pp.
1025–35, 2019. doi:10.1109/TSG.2017.2757599.
38. N. Bhatt, and A. Kumar, “A passive islanding detection
algorithm based on modal current and adaptive boost- 50. T. Ghanbari, “Autocorrelation function-based technique
ing,” Arab. J. Sci. Eng., Vol. 45, pp. 6791–801, 2020. for stator turn-fault detection of induction motor,” IET
doi:10.1007/s13369-020-04709-x. Sci Mearsure Technol, Vol. 10, no. 2, pp. 100–10, 2016.
doi:10.1049/iet-smt.2015.0118.
39. S. Mishra, R. K. Mallick, D. A. Gadanayak, and P. Nayak. “A
novel hybrid downsampling and optimized random forest 51. T. Ghanbari, and E. Farjah, “A multiagent-based fault-
approach for islanding detection and non-islanding power current limiting scheme for the microgrids,” IEEE Trans.
quality events classification in distributed generation inte- Power Deliv., Vol. 29, no. 2, pp. 525–33, 2014. doi:10.1109/
grated system,” IET Renew. Power Gener., January, pp. TPWRD.2013.2282917.
1–16, 2021. doi:10.1049/rpg2.12137.
52. R. Haider, C. H. Kim, T. Ghanbari, et al., “Passive island-
40. S. Dutta, S. Olla, and P. K. Sadhu, “A secured, reliable and ing detection scheme based on autocorrelation function
accurate unplanned island detection method in a renew- of modal current envelope for photo-voltaic units,” IET
able energy based microgrid,” Eng. Sci. Technol. an Int. J. Gener. Trans. Distrib., Vol. 12, no. 3, pp. 726–36, 2018.
2021. doi:10.1016/j.jestch.2021.01.015. doi:10.1049/iet-gtd.2017.0823.

41. W. Hu, et al., “One-Step-Prediction discrete observer- 53. Y. Wang, J. Ravishankar, and T. Phung, “Wavelet transform-
based frequency-locked-loop technique for three-phase based feature extraction for detection and classification
system,” IEEE. Access., Vol. 9, pp. 95401–11, 2021. of disturbances in an islanded micro-grid,” IET Gener.
doi:10.1109/ACCESS.2021.3093351. Transm. Distrib, Vol. 13, no. 11, pp. 2077–87, 2019.
doi:10.1049/iet-gtd.2018.5131.
42. P. Giroux, G. Sybille (Hydro-Quebec, IREQ) C. Osorio,
S. Chandrachood (The MathWorks) https://in.mathworks. 54. Y. Yu, W. Zhao, S. Li, and S. Huang, “A Two-stage wavelet
com/Help/physmod//sps/ug/average-model-of-a-100- decomposition method for instantaneous power quality
kW-grid-connected-pv-array-.html indices estimation considering interharmonics and tran-
sient disturbances,” IEEE Trans. Instrum. Meas., Vol. 70,
43. M. A. G. de Brito, et al. “Comparative analysis of MPPT pp. 1–13, 2021. doi:10.1109/TIM.2021.3052554.
techniques for PV applications,” 2011 Inter. Conf. Clean
Elect. Power (ICCEP), 2011, pp. 99–104. doi:10.1109/ICC 55. M. Sharie, M. R. Mosavi, and N. Rahemi, “Determina-
EP.2011.6036361. tion of an appropriate mother wavelet for de-noising of
 N. BHATT AND A. K. CHANDEL: A RANDOM UNDER-SAMPLING BASED PASSIVE APPROACH 15

weak GPS correlation signals based on similarity measure- 61. Y. Wang, J. Ravishankar, and T. Phung, “Wavelet transform-
ments,” Eng. Sci. Technol. an Int. J, Vol. 23, no. 2, pp. 281–8, based feature extraction for detection and classification
2020. doi:10.1016/j.jestch.2019.05.006. of disturbances in an islanded micro-grid,” IET Gener.
Transm. Distrib, Vol. 13, no. 11, pp. 2077–87, 2019.
56. D. Strömburgssom, et al., “Mother wavelet selection in doi:10.1049/iet-gtd.2018.5131.
the discrete wavelet transform for condition monitoring of
wind turbine drivetrain bearings,” Wind Energy, Vol. 22, 62. S. R. Mounce, K. Ellis, and M. Edwards JM, “Ensem-
pp. 1581–92, 2019. doi:10.1002/we.2390. ble decision tree models using rusboost for estimating
risk of iron failure in drinking water distribution sys-
57. A. I. Megahed, A. M. Moussa, H. B. Elrefaie, et al. “Selec- tems,” Water Resour. Manage, Vol. 31, pp. 1575–89, 2017.
tion of a suitable mother wavelet for analyzing power doi:10.1007/s11269-017-1595-8.
system fault transients”, IEEE Power and Energy Society
General Meeting - Conversion and Delivery of Electri- 63. R. A. Berk, “An introduction to ensemble methods for data
cal Energy in the 21st Century, Pittsburgh, pp. 1–7, 2008. analysis,” Sociol. Methods. Res., Vol. 34, no. 3, pp. 263–95,
doi:10.1109/PES.2008.4596367. 2006.

58. M. Heidari, G. Seifossadat, and M. Razaz, “Application 64. Z.-H. Zhou. Ensemble methods: foundations and algo-
of decision tree and discrete wavelet transform for an rithms. Chapman and Hall/CRC, 2012.
optimized intelligent-based islanding detection method in
distributed systems with distributed generations,” Renew. 65. C. Seiffert. “RUSBoost: Improving classification perfor-
Sustain. Energy Rev., Vol. 27, no. –, pp. 525–32, 2013. mance when training data is skewed”, Inter. Confer. Pat-
doi:10.1016/j.rser.2013.06.047. tern Recogn., Tampa, FL., pp. 1–4, 2008. doi:10.1109/ICPR.
2008.4761297.
59. S. R. Mohanty, N. Kishor, P. K. Ray, et al. “Comparative
study of advanced signal processing techniques for island- 66. V. K. Naik, and A. Yadav. “High Impedance Fault Detection
ing detection in a hybrid distributed generation system”, and Classification on IEEE-15 Bus Radial Distribution Sys-
IEEE Power Energy Society General Meet., Denver, CO, pp. tem by Using Fuzzy Inference System,” 2018 2nd Intern.
2409–13, 2015. doi:10.1109/PESGM.2015.7285854. Conf. on Power, Energy and Envir.: Towards Smart Tech-
nology (ICEPE), 2018, pp. 1–6. doi:10.1109/EPETSG.2018.
60. H. T. Do, X. Zhang, and N. V. Nguyen, “Passive-islanding 8658778.
detection method using the wavelet packet transform
in grid-connected photo-voltaic systems,” IEEE Trans. 67. Sunpower,’ https://in.mathworks.com/help/physmod/sps/
Power Electron, Vol. 31, no. 10, pp. 6955–67, 2016. ug/250-kw-grid-connected-pv array.html.
10.1109/TPEL.2015.2506464.

AUTHORS Prof. Ashwani Kumar Chandel gradu-

ated in Electrical Engineering from Ker-
Nimish Bhatt received the bachelor’s ala University. He was awarded post-
degree in electrical and electronics engi- graduation degree from Punjab Engineer-
neering from Rajiv Gandhi Technical Uni- ing College, Chandigarh. He was awarded
versity (Madhya Pradesh), India, in 2012, Ph.D. degree from Indian Institute of
the master’s degree in condition moni- Technology, Roorkee, India in 2005. Dr.
toring control and protection of electrical Chandel joined the Department of Electri-
apparatus from National Institute of Tech- cal Engineering, National Institute of Technology, Hamirpur,
nology (NIT), Hamirpur, India, in 2016. HP, India, as Lecturer in 1991, where presently he is working
He is currently pursuing Ph.D. degree in electrical engineering as a Professor and Head of the Department. His research work
from the National Institute of Technology, Hamirpur, India. His has been published in various International Journals of repute
research interests include distribution systems, islanding detec- including IEEE, IEE, Elsevier Science, Taylor & Francis and
tion and machine learning. others. He has worked extensively in the area of harmonic esti-
mation & elimination, condition monitoring of transformers
Corresponding author. Email: [email protected] and currently his research interest continues in these fields. He
is a Fellow of IETE, Member IEEE, and Life Member of ISTE.

Email: [email protected]