Applied Soft Computing 59 (2017) 243257

Contents lists available at ScienceDirect

Applied Soft Computing

journal homepage: www.elsevier.com/locate/asoc

Recognition of power quality disturbances using S-transform based

ruled decision tree and fuzzy C-means clustering classiers
Om Prakash Mahela , Abdul Gafoor Shaik
Department of Electrical Engineering, Indian Institute of Technology, Jodhpur 342011, India

a r t i c l e i n f o a b s t r a c t

Article history: A method based on Stockwells transform and Fuzzy C-means (FCM) clustering initialized by decision
Received 25 July 2016 tree has been proposed in this paper for detection and classication of power quality (PQ) disturbances.
Received in revised form 2 May 2017 Performance of this method is compared with S-transform based ruled decision tree. PQ disturbances are
Accepted 31 May 2017
simulated in conformity with standard IEEE-1159 using MATLAB software. Different statistical features
Available online 6 June 2017
of PQ disturbance signals are obtained using Stockwells transform based multi-resolution analysis of
signals. These features are given as input to the proposed techniques such as rule-based decision tree
Keywords:
and FCM clustering initialized by ruled decision tree for classication of various PQ disturbances. The PQ
Fuzzy C-means clustering initialized by
decision tree
disturbances investigated in this study include voltage swell, voltage sag, interruption, notch, harmonics,
Pattern recognition spike, icker, impulsive transient and oscillatory transient. It has been observed that the efciency of
Power quality classication based on ruled decision tree deteriorates in the presence of noise. However, the classication
Ruled decision tree based on Fuzzy C-means clustering initialized by decision tree gives results with high accuracy even in
Statistical feature the noisy environment. Validity of simulation results has been veried through comparisons with results
Stockwells transform in real time obtained using the Real Time Digital Simulator (RTDS) in hardware synchronization mode.
The proposed algorithm is established effectively by results of high accuracy to detect and classify various
electrical power quality disturbances.
2017 Elsevier B.V. All rights reserved.

1. Introduction quality of power may cause motor failures, mal-operation of protec-

tion devices and equipments, metering with low accuracy, reduced
Nowadays, electrical power quality (PQ) is becoming a relevant efciency of appliances, premature ageing of appliances and equip-
concern for both providers and consumers of electrical power to ments and overheating of lines. This may also cause interference
identify the cause of disturbances for protection of equipments [1]. with communication systems [4]. The causes and sources of PQ dis-
PQ disturbances are introduced in the network mainly because of turbances are required to initiate a PQ improvement action. This can
extensive use of solid state switching devices in the electronics and be achieved by effective detection, classication and localization of
electrical circuits, digital computers, power electronically switched PQ events available with voltage and current signals.
loads, variable speed drives, non-linear loads, rectiers, data pro- Advanced mathematical, signal processing and articial intelli-
cessing equipments and inverters. Use of the microprocessor based gent (AI) techniques are proposed to effectively detect and classify
controllers and devices in large number of complicated industrial the power quality disturbances whose features may overlap with
processes also affect quality of power [2]. Problem of power quality each other [57]. The Fourier Transform (FT) and its version short
is related to the deviation of current and voltage from their stan- time FT (STFT) are utilized for power quality detection which gives
dard pure sinusoidal waveforms. PQ problem covers a variety of rise to poor efciency of recognition. These transforms cannot ade-
electromagnetic phenomenon which may be either transient or quately analyse transient phenomenon due to xed window size.
steady-state in nature. These include voltage swell, interruption, A Wavelet transform (WT) which has a variable size window has
voltage sag, oscillatory transient (OT), harmonics, impulsive tran- been used extensively overcoming the limitations of STFT which
sient (IT), spikes, notches, voltage unbalance and icker [3]. Poor in turn gives better efciency of classication. Use of the contin-
uous wavelet transform (CWT) and its discrete version known as
discrete wavelet transform (DWT) proved to be more efcient for
Corresponding author.
analysis of the non-linear and transient phenomenon. However,
E-mail addresses: [email protected], [email protected] (O.P. Mahela),
electrical noise present in signal has signicantly affected the per-
[email protected] (A.G. Shaik). formance of these transforms for detection of PQ disturbances [8,9].

http://dx.doi.org/10.1016/j.asoc.2017.05.061
1568-4946/ 2017 Elsevier B.V. All rights reserved.
244 O.P. Mahela, A.G. Shaik / Applied Soft Computing 59 (2017) 243257

Stockwell transform (S-transform) is a technique based on spectral

localization of time-frequency which has advantageous features of
both WT and STFT. This transform gives resolution of signal, hav-
ing PQ disturbance, dependent on frequency using a window of
frequency dependent width. This transform also provides phase
spectrum and amplitude of analysed PQ disturbance signals. High
accuracy is achieved using this transform even in the presence of
high level of electrical noise [10,11]. An approach using S-transform
for online monitoring of PQ disturbances in distribution network is Fig. 1. Ruled decision tree based scheme for recognition of PQ disturbances.
reported in [12].
The various pattern recognition techniques which are capable
decision tree. Firstly, PQ disturbances are generated in MATLAB
to identify extremely non-linear functional mapping are deployed
software as per parameters specied in the standard IEEE-1159.
for PQ disturbances classication. Articial intelligent methods are
Then S-transform is utilized for feature extraction of 10 simulated
commonly used to determine various patterns of PQ disturbances.
PQ disturbances. Afterwards, a set of features (F1F8) are given as
In recent years, various approaches such as articial neural network
input to decision tree based algorithm to classify investigated PQ
(ANN) [13,14], modied euclidean ARTMAP neural network [15],
disturbances. However, the classication based on FCM clustering
sparse signal decomposition (SSD) with a matrix based on com-
algorithm is initialized using the feature F2. After this, the feature
plete hybrid dictionary (OHD) [16], a binary feature matrix based
vectors F9F13 are used as input to this proposed decision tree ini-
decision making system [17], bacterial foraging optimization (BFO)
tialized FCM clustering algorithm for classication purpose. Finally,
algorithm based on Fuzzy clustering (FC) [3] and the support vector
to evaluate performance, the two algorithms are tested using hun-
machine (SVM) [18]. An algorithm based on ruled decision tree pro-
dred data sets of each power quality disturbance. This data set is
vides the decision of classication using rules driven from data with
obtained by varying the various parameters of PQ disturbances. Per-
advantages of low computational burden and fare accuracy [19].
formance of algorithms is also tested by adding electrical noise of
Fuzzy C-means (FCM) clustering approach is used for PQ analysis
20 dB SNR with the signal. The results of decision tree and deci-
with minimum computational complexity [20].
sion tree initialized FCM clustering algorithms are compared with
A symmetrical components (in time domain) based approach
each other. Moreover, to establish the feasibility and effective-
to detect and classify the PQ events is reported in [21]. This
ness of algorithms, a comparative study is carried out in terms
task has been achieved using features extracted from sum of
of percentage accuracy between proposed approaches and sev-
positive-sequence components. This approach has advantages of
eral existing methods. The validity of simulation results has been
low computational burden and easy segmentation for recognition
veried through comparisons with results in real time obtained
of PQ disturbances. Samantaray et al. [22], proposed a decision tree
using Real Time Digital Simulator (RTDS) in hardware synchro-
and variational mode decomposition based algorithm for recogni-
nization mode. Hence, main contribution of the presented work
tion of both single stage as well as complex power quality events
is the design of an approach using S-transform based decision tree
in an utility grid interfaced with distributed generation. The pro-
initialized FCM clustering for the recognition of PQ disturbances.
posed method has advantages of accurate estimation of amplitudes,
This algorithm has the advantages of minimum effect of noise on
frequencies and phases of fundamental component, harmonic,
the performance and high efciency even in noisy environment.
inter-harmonic and icker components even in the noisy envi-
The sum of absolute values curve introduced in this paper is an
ronment. A technique for online recognition of PQ disturbances
additional contribution which helps to achieve high efciency. The
using tunable-Q WT (TQWT) for detection and dual multi-class
proposed approach can be used for PQ monitoring devices.
SVM (MSVM) for classication of these disturbances is reported
The contents in this paper are organized into eight sec-
in [23]. This technique is capable to adjust redundancy of wavelet
tions. The proposed methodologies of PQ events recognition are
transform and Q-factor for feature extraction of PQ disturbances
detailed in Section 2. Section 3 describes the S-transform based
to achieve high efciency of recognition. A method for segmenta-
multi-resolution analysis (MRA) of PQ disturbance signals. The S-
tion of single stage and complex PQ disturbances based on adaptive
transform based features are dened in Section 4. Section 5 covers
thresholds determined using DWT has been reported in [24]. This
classication results of PQ disturbances obtained using ruled deci-
technique uses the calculated energy and its entropy resulted from
sion tree and decision tree initialized FCM clustering classiers.
WT based decomposition of signal. Abdelsalam et al. [25], described
Real time validation of results is detailed in Section 6. Section 7
an effective approach for detection and classication of PQ distur-
presents the comparative study of various PQ recognition tech-
bances. This approach uses the Kalman lter with linear nature
niques. Finally, conclusions drawn based on the presented work
and DWT for feature extraction from voltage or current waveforms.
are detailed in Section 8.
These extracted features are fed to fuzzy-expert system for the pur-
pose of classication of disturbances. The use of DWT with Kalman
lter improves the performance of recognition of power quality 2. Proposed PQ events recognition methodology
disturbances. Mahela et al. [26], proposed an algorithm for recog-
nition of PQ disturbances associated with solar energy penetration The proposed algorithms based on ruled decision tree and
in distribution network during the events of grid synchronization decision tree initialized FCM clustering for recognition of PQ dis-
and outage of solar PV system. A technique using fast variant of turbances are presented in this section. The PQ disturbances are
S-transform and fuzzy decision tree based classier for recognition detected, classied and localized with the help of these algorithms
of PQ disturbances is proposed by the authors in [27]. Proposed based on features extracted from signals using S-transform. Block
method has the advantage of less computational burden and accu- diagram schemes of the proposed methodologies are illustrated in
rately classies the disturbances with overlapping features. Figs. 1 and 2respectively.
This paper presents a method based on Stockwells transform Mathematical models in MATLAB software are used to generate
and Fuzzy C-means clustering initialized by decision tree for recog- signals (of 50 Hz) with power quality disturbances in conformity
nition of power quality disturbances which is the extension of work with standard IEEE-1159. These signals depict real time data and
reported in the articles [28] and [29]. Results of this algorithm have establish generalized capability of classier. Mathematical models
been compared with an approach based on S-transform and ruled of signals, their control parameters and simulated values of control
 O.P. Mahela, A.G. Shaik / Applied Soft Computing 59 (2017) 243257 245

Fig. 2. Decision tree initialized FCM clustering based scheme for recognition of PQ disturbances.

Table 1
Mathematical modeling of simulated PQ disturbances [30].

PQ disturbance Class symbol Mathematical model Technical parameters

Standard Simulated

Pure sine wave C1 V(t) =Asin(t) = 2f A = 1 pu, f = 50 Hz

Voltage sag C2 V(t) = (1 (u(t t1 ) u(t t2 )))sin(t) 0.1 0.9, T t2 t1 9T = 0.3, t1 = 0.06, t2 = 0.14
Voltage swell C3 V(t) = (1 + (u(t t1 ) u(t t2 )))sin(t) 0.1 0.8, T t2 t1 9T = 0.3, t1 = 0.06, t2 = 0.14
Interruption C4 V(t) = (1 (u(t t1 ) u(t t2 )))sin(t) 0.9 1.0, T t2 t1 9T = 0.95, t1 = 0.06, t2 = 0.14
Flicker C5 V(t) = (1 + f sin(t))sin(t)
(tt1 )
0.1 f 0.2, 5 20 Hz f = 0.15, = 15
Oscillatory transient C6 V (t) = sin(t) + e  sin n (t t1 ){u(t2 u(t1 ))} 0.1 0.8, = 0.8, t1 = 0.08,  = 0.02,
0.05T t2 t1 3T, t2 = 0.10
8 ms  40 ms,
(tt1 ) (tt1 )
300 fn 900 Hz
Impulsive transient C7 V (t) = sin(t) + e  e  {u(t2 u(t1 ))} 1 10, = 10, t1 = 0.085,  = 0.02,
0.05T t2 t1 3T, t2 = 0.088
8 ms  40 ms
Harmonics C8 V(t) = 1 sin(t) + 3 sin(3t) + 5 sin(5t) + 7 sin(7t) 0.1 3 , 5 , 7 0.15 3 = 0.05, 5 = 0.10,
7 = 0.15
Notch C9
9 
V (t) = sin(t) sign(sin(t))
 0.1 K 0.4, 0 t1 , K = 0.4, t1 = 0.006,
K u (t (t1 + 0.02n)) u (t (t2 + 0.02n)) t2 0.5T, t2 = 0.0065
n=0
0.01T t2 t1 0.05T
Spike C10
9 
V (t) = sin(t) + sign(sin(t))
 0.1 K 0.4, 0 t1 , K = 0.4, t1 = 0.002,
K u (t (t1 + 0.02n)) u (t (t2 + 0.02n)) t2 0.5T, t2 = 0.0023
n=0
0.01T t2 t1 0.05T

A: amplitude; f: frequency; V: voltage; T: time period; : time constant; : angular frequency; u(t): unit step function.

parameters are provided in Table 1 [30]. Class symbol C1 is used retained [32]. Hence, the S-transform can be used effectively to
to designate the pure sine wave whereas the symbols C2C10 are extract features of non-stationary PQ disturbance signals. The S-
assigned to investigated PQ disturbances as illustrated in Table 1. transform based decomposition of signal gives a matrix of size
Generated PQ disturbance signals are sampled with 64 samples per n m as output which is complex in nature. This matrix is known as
cycle (3.2 kHz sampling frequency) over a period of 10 cycles. Then S-matrix and represented by the following mathematical relation
S-transform based multi-resolution analysis is used to decompose [33].
these signals and S-matrix is obtained. Time frequency represen-
S(, f ) = A(, f )e(i(,f )) (1)
tation of this matrix is utilized to extract various features F1F14.
These features are grouped into two separate classes. First class of where (, f): phase; A(, f): amplitude. Rows and columns of S-
features comprising of F1F8 is used to design decision rules uti- matrix corresponds respectively to frequency and time. Magnitude
lized by decision tree based classication. Second class of features corresponding to a particular frequency with respect to time is
F9F13 is used for classication by the algorithm based on FCM represented by each row in terms of samples from 0 to N 1.
clustering initialized by decision tree. The PQ disturbances such as Frequency components associated with the signal at a particular
voltage swell, voltage sag and OT can be localized with the help of instant of time are represented by each column. Important infor-
feature F14. The performance of proposed classication techniques mation related to the frequency, phase and magnitude is extracted
has been validated using a data set comprising of hundred signals from S-matrix. At a particular instant of time, locus of the maximum
which are generated by varying various parameters of each PQ dis- values of S-matrix represent magnitude contour. The phase corre-
turbance. Performance of both rule-based decision tree and FCM sponding to regions of amplitude with maximum values represents
clustering algorithms has been investigated in noisy environment the phase of S-matrix. Frequency contour represents frequency
with a noise of 20 dB signal to noise ratio (SNR). This noise level contents associated with the signal which are derived from S-
is added to signals using MATLAB command awgn. Mathematical matrix [18].
tools such as Stockwells transform, rule-based decision tree and
FCM clustering used in proposed algorithms are described in the 2.2. Rule based decision tree
following subsections.
The decision tree is a classication tool supported by decision
2.1. Stockwell transform rules which is invented in 1980 by Breiman. Wehenkel was the
rst scientist to implement decision tree in the year 1989 for clas-
R. G. Stockwell proposed a signal processing transform in 1996 sication purpose in the power system eld. A binary tree graph is
known as Stockwell transform (S-transform). It conceptually con- used by this technique to nd unknown relationship between input
tains the elements of both wavelet analysis and short-time Fourier and output parameters [19]. It represents decision with the help of
analysis but it falls in a different category [31]. The S-transform rules driven from data. Decisions are presented in the form of tree
decomposes a time varying signal (PQ disturbance signals) using like graph starting from root node and nal decision contained in
MRA where absolute phase of each constituent frequency is the leaf node. Response is contained by the leaf node [34]. A set
246 O.P. Mahela, A.G. Shaik / Applied Soft Computing 59 (2017) 243257

of features obtained from the S-matrix is utilized for the decision

rules.

2.3. Fuzzy C-means clustering

Features extracted from the PQ disturbance signals using S-

transform based multi-resolution analysis are given as input to FCM
clustering for classication purpose. Dunn was the rst to introduce
the FCM clustering technique which was generalized by Bezdek in
the year 1981 [20]. It is basically a multiple-clustering technique
which is implemented by optimizing the basic C-means objective
function or its modication [20]. FCM clustering has the capability
to determine and iteratively update the membership values corre-
sponding to a data point in accordance with pre-dened number
of clusters. Hence, the data point has corresponding membership
values for all the clusters [35,36]. Objective function is given by the
following relation


N 
C

Jm = um
ij
||xi cj ||2 (2)
i=1 j=1

where m: cluster numbers; xi : ith element of n-dimensional mea-

sured data; uij : degree of membership of xi in cluster j; cj : the
n-dimensional centre of cluster. FCM clustering technique has fol-
lowing advantages over other clustering techniques:
Fig. 3. Stockwell transform based plots (a) pure sine wave, (b) frequency con-
tour, (c) amplitude plot, (d) sum of absolute values plot, (e) phase plot and (f)
FCM clustering has a kind of uniqueness that all samples of sig- amplitudefrequency plot.
nals belong to every cluster with a respective membership value
against various clustering methods such as articial neural net-
works, Wards method and co-clustering where a sample belongs
to single cluster. 3.1. Pure sine wave
Concept of the break points and similarities used in this algo-
rithm decide more precisely and with minimum time interval Fig. 3 shows various S-transform based plots of pure sine wave.
that whether a sample should be outside or inside of a cluster. These plots are used as reference for detection of PQ disturbances.
This quantied concept of similarity makes clustering of variables The peak observed in amplitudefrequency curve at 50 Hz indi-
more effective and easier as compared to other techniques. cates pure sinusoidal nature of the wave. The deviations from these
FCM clustering algorithm may accounts for uncertainties. This curves will be an indication of presence of power quality distur-
helps to produce more appropriate results while determining the bance in the signal.
class boundaries.
FCM clustering has minimum computational complexity. The
computational complexity may be reduced further by the use of 3.2. Voltage sag
interval-valued fuzzy set.
FCM clustering is vulnerable to outliers. The S-transform based plots related to voltage sag (30%) are
It is relatively insensitive to initializations. presented in Fig. 4. Sag in the voltage is effectively detected by
Minimum number of fuzzy rules is used by FCM clustering. decreased amplitude in plots like amplitudetime curve, S-contour
(frequency contour) and sum of absolute values plot shown respec-
tively in Fig. 4(c), (b) and (d). However, magnitude in other plots
3. Analysis of PQ disturbances using S-transform remains same as in case of pure sine wave. Incidence of two peaks
at the time instants 0.06 s and 0.14 s in the second derivative plot
The S-transform based analysis of investigated PQ disturbances related to sum absolute values plot described in Fig. 4(f) helps to
is detailed in this section. The S-matrix is used to obtain various effectively localize sag in the voltage.
plots using S-transform to detect and classify PQ events. These
include phase curve, time-frequency contour (also known as S-
contour), amplitudefrequency curve and amplitude curve. A curve 3.3. Voltage swell
known as sum of absolute values curve has also been introduced
in this work which is intended to add a new feature for classi- A swell in voltage and respective plots based on S-transform
cation of PQ disturbances with improved efciency. This curve is are depicted in Fig. 5. Voltage swell (30%) is effectively detected by
obtained by summing absolute values over a column of S-matrix. increased magnitudes of amplitudetime curve, S-contour and sum
The sharp peaks are detected in the second derivative of this curve absolute values plot as illustrated respectively in Fig. 5(c), (b) and
at the instant of start and end of PQ events like voltage swell, voltage (d). However, magnitude of phase and frequencyamplitude curves
sag, momentary interruption and OT. These peaks can be utilized will remain same as observed in case of pure sine wave. Incidence of
for localizing such events. The S-transform based plots of pure sine two peaks at time instants 0.06 s and 0.14 s in the second derivative
wave are taken as reference which will help in detecting PQ distur- curve related to the sum absolute values as observed in Fig. 5(f)
bances. helps to effectively localize the voltage swell.
 O.P. Mahela, A.G. Shaik / Applied Soft Computing 59 (2017) 243257 247

Fig. 4. Stockwell transform based plots (a) voltage sag, (b) frequency contour, (c) Fig. 6. Stockwell transform based plots (a) voltage interruption, (b) frequency con-
amplitudetime plot, (d) sum of absolute values plot, (e) phase plot, (f) second tour, (c) amplitude plot, (d) sum of absolute values plot, (e) phase plot, (f) second
derivative plot and (g) amplitudefrequency plot. derivative plot and (g) amplitudefrequency plot.

respective plot related to the pure sine wave. Duration of inter-

ruption in voltage can be obtained from second derivative plot
by calculating the interval between peaks available at the instants
0.06 s and 0.14 s as shown in Fig. 6(f).

3.5. Harmonics

Harmonics and respective plots based on S-transform are pre-

sented in Fig. 7. Continuous ripples on the surface of S-contour
as well as sum absolute values plot as illustrated respectively in
Fig. 7(b) and (d) can effectively be utilized for detection of harmon-
ics in the voltage signal. The nite values of frequency amplitude
curve between the normalized frequencies of 0.05 and 0.15 also
detect presence of the harmonics in voltage signal as illustrated in
Fig. 7(f).

3.6. Flicker

A icker and respective plots based on S-transform are pre-

sented in Fig. 8. Continuous circles in the form of a series observed
in frequency contour (S-contour) of Fig. 8(b) indicate the presence
of icker. Continuously occurring ripples in sum absolute values
plot illustrated in Fig. 8(d) effectively demonstrates the presence
Fig. 5. Stockwell transform based plots (a) voltage swell, (b) frequency contour, (c) of icker in the signal. Presence of this icker leads to a nite value
amplitude plot, (d) sum of absolute values plot, (e) phase plot, (f) second derivative in the frequencyamplitude plot between 0.15 and 0.35 normalized
plot and (g) amplitudefrequency plot. frequencies as depicted in Fig. 8(f).

3.4. Momentary interruption 3.7. Oscillatory transient

A momentary interruption in voltage signal and its related plots Fig. 9 depicts an oscillatory transient (0.080.10 s) and its related
based on MRA using S-transform are described in Fig. 6. A discon- S-transform based plots. This oscillatory transient can effectively
tinuity in S-contour as well as decreased magnitude below 10% of be sensed by a single isolated contour present in the S-contour as
amplitude and sum of absolute value plots as illustrated respec- shown in Fig. 9(b). The presence of OT in voltage signal can also
tively in Fig. 6(b), (c) and (d) indicate presence of momentary be sensed by large magnitude changes present between 0.08 s and
interruption in voltage. It can be observed from Fig. 6(e) that magni- 0.10 s in all the plots of Fig. 9. These simultaneous changes in all
tude of phase curve is also decreased with the event of interruption the curves are not detected in other types of PQ disturbances. This
in voltage. However, frequencyamplitude curve is identical with helps in discrimination of OT from other PQ events. Localization of
248 O.P. Mahela, A.G. Shaik / Applied Soft Computing 59 (2017) 243257

Fig. 9. Stockwell transform based plots (a) oscillatory transient, (b) frequency con-
tour, (c) amplitude plot, (d) sum of absolute values plot, (e) phase plot, (f) second
derivative plot and (g) amplitudefrequency plot.
Fig. 7. Stockwell transform based plots (a) harmonics, (b) frequency contour,
(c) amplitude plot, (d) sum of absolute values plot, (e) phase plot and (f)
amplitudefrequency plot.

Fig. 10. Stockwell transform based plots (a) impulsive transient, (b) frequency con-
tour, (c) amplitude plot, (d) sum of absolute valuestime plot, (e) phase plot and (f)
amplitudefrequency plot.

Fig. 8. Stockwell transform based plots (a) icker, (b) frequency contour, (c)
amplitudetime plot, (d) sum of absolute values plot, (e) phase plot and (f)
amplitudefrequency plot.
3.8. Impulsive transient

Fig. 10 depicts an impulsive transient and its S-transform based

plots. The impulsive magnitude in both the sum absolute values
plot and S-contour described respectively in Fig. 10(d) and (b) helps
the oscillatory transient (0.080.10 s) is validated by sharp peaks to detect the impulsive transient present in voltage signal. These
available in the second derivative curve as detailed in Fig. 9(f). impulsive magnitudes also help in the discrimination of IT from
 O.P. Mahela, A.G. Shaik / Applied Soft Computing 59 (2017) 243257 249

Fig. 11. Stockwell transform based plots (a) multiple notches, (b) frequency con- Fig. 12. Stockwell transform based plots (a) multiple spikes, (b) frequency con-
tour, (c) amplitude plot, (d) sum of absolute values plot, (e) phase plot and (f) tour, (c) amplitude plot, (d) sum of absolute values plot, (e) phase plot and (f)
amplitudefrequency plot. amplitudefrequency plot.

other PQ events. Since impulsive transient contains all the frequen-

cies, nite values have been observed over the entire frequency extracted to classify these PQ disturbances are detailed in the next
range in frequencyamplitude plot illustrated in Fig. 10(f) which section.
also helps in discriminating IT from other PQ events.

3.9. Multiple notches 4. S-transform based feature extraction

Voltage signal with multiple notches and respective plots based This section details the different features of PQ disturbances
on S-transform are illustrated in Fig. 11. A series of regularly spaced extracted from S-transform based plots discussed in Section 3 using
contours observed in S-contour of Fig. 11(b) effectively detects the various statistical techniques. These features are designated by the
multiple notches. The equally spaced ripples observed continu- symbols F1F14. All the features are dened as described below.
ously in the sum absolute values plot of Fig. 11(d) also indicate the F1: Proposed sum factor. It is dened as
presence of multiple notches. Gradual and continuous increasing Sf = max(S) + min(S)max(R)min(R), where R: a data array of
magnitude of frequencyamplitude plot followed by a peak cor- sum absolute values related to reference voltage signal (pure sine
responding to 50 Hz frequency also indicates presence of multiple wave); S: a data array of sum absolute values of voltage signal
notches. having PQ disturbance. Zero value of this sum factor indicates no
PQ disturbance in the signal. Finite value of sum factor will indicate
that a PQ disturbance is present in voltage signal.
3.10. Multiple spikes
F2: This feature indicates the number of peaks available in S-
transform based frequencyamplitude plot.
Multiple spikes in voltage signal and respective plots based on
F3: This is the skewness of S-transform based phase plot. Fol-
S-transform are shown in Fig. 12. Equally spaced contours observed
lowing relation can be utilized to calculate skewness of a signal.
in S-contour as detailed in Fig. 12(b) detects the spikes. These spikes
are also validated by the ripples continuously present in sum abso-
lute values curve as detailed in Fig. 12(d). However, the number of E(x )3
s= (3)
contours are different from that of multiple notches. Hence, number 3
of these contours and ripples will help in discrimination of spikes
from the notches. Gradual and continuous increasing magnitude of where x represents an array of signal data;  represents the stan-
amplitudefrequency curve after the single peak corresponding to dard deviation of x;  indicates the mean of x; E is the expected
50 Hz is also observed due to the presence of multiple spikes in a value of quantity.
way similar to the multiple notches. F4: Proposed amplitude factor. It is dened as
Thus, it has been observed that several similarities and dissimi- Af = (1 + (C A) + (D B)) where C: maximum value of ampli-
larities exist in the S-transform based plots related to various power tude plot of test signal with PQ disturbance; D: minimum value
quality disturbances. Hence, these disturbances can be classied of amplitude curve of test signal; A: maximum value of amplitude
into various groups of similar features. These groups can further be plot of reference signal; B: minimum value of amplitude plot of
classied from sub-groups to single entities. The statistical features pure sine wave (reference).
250 O.P. Mahela, A.G. Shaik / Applied Soft Computing 59 (2017) 243257

F5: This indicates the Kurtosis (k) of S-transform based ampli- peak in frequency amplitude plot. PQ disturbance signals included
tude plot. Following relation is used to dene kurtosis of a signal. in the group G1 are sub-classied further in sub-groups G11 and
G12 using value of sum factor (F1). The signals with feature F1 < 5
E(x )4
k= (4) are included in subgroup G11 whereas the subgroup G12 contains
4 the signals with F1 > 5. The value of amplitude factor (F4) is fur-
where x represents an array of signal data;  represents the stan- ther utilized to identify the two signals contained in the subgroup
dard deviation of x;  indicates the mean of x; E is the expected G12. PQ disturbance signal with F4 > 1 is identied as spike (C10)
value of quantity. whereas the signal with F4 < 1 is notch (C9). The features F4 and
F6: This is Kurtosis of S-transform based phase plot. F1 are used for further classication of four PQ disturbance signals
F7: This is Kurtosis of S-transform based sum absolute values included in group G11. The data with F4 > 1 and F1 > 0 will indicate
plot. presence of voltage swell (C3). The voltage signal corresponding to
F8: This is Kurtosis of S-transform based amplitudefrequency pure sine wave (C1) is observed to have feature values such as F4 = 1
plot. as well as F1 = 0. PQ disturbances such as voltage sag (C2) and MI
F9: This is mean derived from S-matrix. (C4) having low magnitude compared to pure sine wave will have
the values of F1 < 0. The value of feature F4 can be utilized to dis-
F9 = mean(abs(S(j, n))) (5)
criminate these disturbances from each other. The values of feature
F10: This is standard deviation derived from S-matrix. F4 > 0.3 indicate the presence of voltage sag whereas F4 < 0.3 will
indicate the presence of interruption.
F10 = std(abs(S(j, n))) (6)
PQ disturbance signals contained by the group G2 are further
F11: Variance derived from the S-matrix (amplitude). sub-classied in two groups G21 and G22 using values of features
F3 and F8. Subgroup G21 contains signals having values of fea-
F11 = var(abs(S(j, n))) (7)
tures F3 > 0.01 and F8 < 30 whereas all other signals of group G2 are
F12: This indicates the energy content of S-transform based S- classied as subgroup G22. Signal corresponding to oscillatory tran-
contour. sient (C6) and impulsive transient (C7) have been included in the
2 subgroup G21. The features F5 and F6 are used to classify these dis-
F12 = (abs(S(j, n))) (8) turbances into individuals. Oscillatory transient will have F6 > 5 and
F13: This represents the maximum deviation calculated from F5 > 5. However, impulsive transient has the feature values F6 < 5
S-matrix (amplitude). and F5 < 5. Flicker (C5) and harmonics (C8) are grouped together
into the subgroup G22. The disturbance with a value of F7 < 10 is
F13 = max(Smatrix) F9 (9) icker whereas the disturbance having a value of F7 > 10 is har-
F14: Second order derivative calculated from the S-transform monics. Fig. 13 illustrates the ruled decision tree algorithm based
based sum absolute values plot of voltage signal with PQ distur- classication with the help of a owchart. Threshold values used
bance. in the proposed algorithm are obtained using S-transform based
MRA of one hundred data of PQ disturbance signals of each category
2 f
F14 = = f (n + 1) + f (n 1) 2f (n) (10) (C1C10) generated by varying various parameters (refer Table 1).
n2 Numbers of correctly classied and misclassied PQ disturbance
where n: represents sample number; f(n): represents sum absolute signals have been used to evaluate the efciency of proposed ruled
values. Sharp peaks available in the curve of F14-time will help decision tree algorithm. Hundred sets of data of each PQ distur-
to localize the PQ disturbances like voltage swell, voltage sag, OT, bance are obtained by varying various parameters of the PQ events
momentary interruption, etc. for testing the performance of proposed algorithm. In noisy envi-
ronment, the performance is evaluated by adding Gaussian noise
5. Classication of power quality disturbances of 20 dB SNR to the signals. Table 3, presents performance of the
algorithm for classication of various PQ disturbance signals with
This section presents the classication results of PQ distur- and without the presence of noise.
bances based on proposed algorithms like rule based decision tree
and FCM clustering initialized using decision tree. The features F1,
F2,. . .,F13 as dened in the Section 4 are being utilized by these 5.2. Classication using decision tree initialized fuzzy C-means
algorithms for classication purpose. Features F1, F2,. . .,F8 are used clustering
by the decision tree to design rules. Features F9, F10,. . .,F13 are used
by the FCM clustering algorithm to obtain cluster plots required In this method, the classication is initialized by rule based deci-
for classication purpose. Finite value of feature F1 indicates a PQ sion tree and data are grouped into groups G1 and G2 based on
disturbance available in the signal. Results related to PQ events clas- the feature F2 as explained in Section 5.1. The data included in
sication using proposed approaches are presented in the following these groups are further classied using FCM clustering based on
subsections. features F9F13. The FCM clustering based scatter plots between
two features taken at a time are used for this classication. The six
5.1. Ruled decision tree based classication combinations of scatter plots F13-Time, F10F9, F11F9, F11F10,
F13F10 and F11F12 are found effective for FCM clustering based
Table 2 presents the numerical values of extracted features (F1, classication. These plots give distinctive regions to separate and
F2,. . .,F8) utilized to frame decision rules for classication purpose distinguish the PQ disturbance from each other with high accuracy.
[28]. Once a PQ disturbance is detected based on the value of feature The scatter plots related to signals included in group G1 are illus-
F1, the PQ disturbance signals are included in two major groups G1 trated in Fig. 14. It has been observed that data related to voltage
and G2. The classication of PQ disturbances into these two groups sag (C2), pure sine wave (C1), MI (C4) and voltage swell (C3) occupy
is based on number of peaks detected in frequencyamplitude plot separate regions in the scatter plots. However, signal data of notch
(feature F2). PQ disturbances data with one peak (F2 = 1) in fre- (C9) and spike (C10) occupy regions very close to each other as
quency amplitude plot are included in the group G1. However, the well as around sine wave. Therefore, the subgroup G11 includes
group G2 includes the PQ disturbances data with more than one the signals related to pure sine wave, voltage swell, voltage sag and
 O.P. Mahela, A.G. Shaik / Applied Soft Computing 59 (2017) 243257 251

Table 2
Numerical values of statistical features used for ruled decision tree based classication of PQ disturbances.

Power quality disturbance S-transform based features related to PQ disturbances

F1 F2 F3 F4 F5 F6 F7 F8

C1 0 1 1.7454e13 1.0000 2.8204 1.6011 59.6536 74.0081

C2 00.5782 1 1.3260e13 0.7125 1.4927 1.4602 8.3947 71.6985
C3 00.0757 1 1.2321e13 1.2875 1.4883 1.9833 4.0969 72.1032
C4 01.7409 1 0.0438 0.1066 1.4819 1.8108 2.5148 59.7960
C5 04.3990 2 1.1975e06 1.0000 2.8204 1.6011 1.7092 54.6984
C6 48.8003 2 1.6563 1.6713 20.4860 11.8703 11.3693 5.9078
C7 65.7188 2 0.2879 1.1267 2.6787 1.8527 53.0595 16.2903
C8 07.8434 2 4.6024e14 1.0008 2.8443 1.6005 24.1222 60.3197
C9 10.9126 1 4.8087e05 0.9540 2.8215 1.6010 3.0229 49.8862
C10 10.6793 1 2.1734e05 1.0160 2.8199 1.6009 6.5734 51.3621

Fig. 13. Flow chart of ruled decision tree based classication of PQ disturbances.

Table 3
Performance of ruled decision tree algorithm for classication of PQ disturbances in noisy and noise free environments.

Power quality event Signals correctly classied Signals mis-classied Percentage efciency (%)

Without noise 20 dB SNR noise Without noise 20 dB SNR noise Without noise 20 dB SNR noise

C1 100 99 0 1 100 99
C2 99 99 1 1 99 99
C3 100 100 0 0 100 100
C4 100 99 0 1 100 99
C5 98 98 2 2 98 98
C6 100 99 0 1 100 99
C7 99 99 1 1 99 99
C8 99 97 1 3 99 97
C9 99 98 1 2 99 98
C10 98 97 2 3 98 97
Overall efciency of classication: 99.2% (in noise free environment), 98.5% (in noisy environment)

MI. The subgroup G12 contains the data related to notch and spike turbances included in the group G11 are classied effectively using
signals. FCM clustering.
Fig. 15 illustrates the FCM clustering based scatter plots of signal Fig. 16 depicts the FCM clustering based scatter plots of signals
data included in the subgroup G11. It is observed that signal data included in the group G12. It can be observed that upper position
of this group form non-overlapping clusters corresponding to each in the scatter plots of Fig. 16(a)(e) is occupied by data of spike.
PQ disturbance. It is also evident from these plots that the volt- However, in these plots the data of notch are clustered in lower
age swell, pure sine wave, voltage sag and MI occupy the isolated position. Fig. 16(f) shows a clear discrimination between notch and
regions in a respective regular order from top to bottom. Signal spike due to widely separated clusters of data.
data of pure sine wave are shown by a straight line in the F13-time Fig. 17 illustrates the scatter plots related to signals of group
scatter plot and by a dot in other scatter plots. Thus, all the PQ dis- G2. Scatter plots of data related to oscillatory and impulsive tran-
252 O.P. Mahela, A.G. Shaik / Applied Soft Computing 59 (2017) 243257

Fig. 14. FCM clustering based scatter plots of group G1 data.

Fig. 15. FCM clustering based scatter plots of group G11 data.

sients are clustered separately from clusters related to that of icker The scatter plots of harmonics and icker included in the sub-
and harmonics as described in Fig. 17(a). Hence, the oscillatory and group G22 are shown in Fig. 19 which demonstrates the clear
impulsive transients are classied under subgroup G21. However, discrimination of these disturbances. It is observed that these plots
the harmonics and icker are classied under the subgroup G22. are widely separated in all the plots of Fig. 19 which clearly illus-
Scatter plots of subgroup data G21 are illustrated in Fig. 18. It trates the successful classication of signals in subgroup G22. In
is observed that scatter plots related to OT and IT occupy non- all scatter plots, the Flicker and harmonics data are clustered sep-
overlapping regions as shown in Fig. 18(a) and others. The plots arately with icker data clustered above the harmonics data.
help to discriminate these PQ disturbances from each other. The A hundred data sets of each PQ disturbance is obtained by vary-
data related to OT are clustered on left hand side in Fig. 18(a) scatter ing parameters of these events to test the performance of proposed
plot and above the IT data in all other plots. decision tree initialized FCM clustering algorithm. Performance of
 O.P. Mahela, A.G. Shaik / Applied Soft Computing 59 (2017) 243257 253

Fig. 16. FCM clustering based scatter plots of group G12 data.

Fig. 17. FCM clustering based scatter plots of data group G2.

Table 4
Performance of FCM clustering initialized by decision tree algorithm for classication of PQ disturbances.

Power quality event Number of correctly classied signals Number of mis-classied signals Percentage efciency (%)

Without noise 20 dB SNR noise Without noise 20 dB SNR noise Without noise 20 dB SNR noise level

C1 100 100 0 0 100 100

C2 100 99 0 1 100 99
C3 100 100 0 0 100 100
C4 100 100 0 0 100 100
C5 99 98 1 2 99 98
C6 100 99 0 1 100 99
C7 99 98 1 2 99 98
C8 99 99 1 1 99 99
C9 100 100 0 0 100 100
C10 100 99 0 1 100 99
Overall efciency of classication: 99.6% (in noise free environment), 99.3% (in noisy environment)
254 O.P. Mahela, A.G. Shaik / Applied Soft Computing 59 (2017) 243257

Fig. 18. FCM clustering based scatter plots of group G21 data.

Fig. 19. FCM clustering based scatter plots of group G22 data.

this algorithm under noisy conditions has been established with interacts with the RTDS. Fig. 20 illustrates the complete experi-
addition of noise level of 20 dB SNR to each disturbance. Table 4 mental set-up used for the study. Signals are modelled in simulink
demonstrates the performance of proposed algorithm using num- using standard mathematical models in the MATLAB-2011b on HID.
bers of correctly classied and misclassied signals of disturbances. These models are loaded on RTDS target ML605 for processing. The
results obtained in real time using the RTDS in hardware synchro-
6. Validation of results in real time nization mode are as good as experimental results. The ether-net
communication system is used for communication between lap-
The simulated results are validated using Real Time Digital Sim- top and RTDS. OpWrite block of RT-Lab is used to take out the data
ulator (RTDS) of OPAL-RT. A laptop of 64-bit operating system, 4 from RTDS and analysed using S-transform. Performance of results
GB RAM, Intel(I) Core(TM)i5-3230M [email protected] GHz processor is is validated in terms of numbers of correctly classied and misclas-
utilized as human interface device (HID) in this study. This HID sied PQ disturbance signals using a set of hundred data of each PQ
 O.P. Mahela, A.G. Shaik / Applied Soft Computing 59 (2017) 243257 255

Table 5
Comparison of simulated and real time classication results of algorithm using FCM clustering initialized by decision tree in terms of accuracy (%).

PQ event Simulated results Real time results Deviation (%)

Without noise 20 dB SNR noise Without noise 20 dB SNR noise Without noise 20 dB SNR noise

C1 100 100 100 100 0 0

C2 100 99 100 100 0 1.01
C3 100 100 100 100 0 0
C4 100 100 100 100 0 0
C5 99 98 98 97 1.01 1.02
C6 100 99 100 99 0 0
C7 99 98 100 97 1.01 1.02
C8 99 99 98 97 1.01 2.02
C9 100 100 100 99 0 1.00
C10 100 99 98 97 2.00 2.02
Overall classication accuracy (RT): 98.33% (without noise), 96.85% (20 dB SNR noise)

methods in the noise free and noisy environment is illustrated in

Table 6. Noise level of 20 dB SNR is considered in this work which
is higher than noise level considered in the references [4] and [16]
and equal to that considered in all other articles. It is observed that
noise present in signal deteriorates classication efciency of the
algorithms in general. It is established that the proposed algorithm
of FCM clustering initialized by decision tree provides better accu-
racy than the rule-based decision tree algorithm. From Table 6, it
is also observed that efciency of proposed algorithm of FCM clus-
tering initialized by decision tree is higher than the efciency of
algorithms reported in references [3,4,12,16,18].
The Fuzzy C-means clustering based algorithm could achieve
100% efciency in classifying the events investigated in the ref-
erence [10] i.e. harmonics, oscillatory transient, spike and notch.
However, the efciency of classication reduces to 99.25% in the
presence of noise for these events mentioned above. The reference
[10] does not present the results in noisy environment.
Thus, the algorithm based on Fuzzy C-means clustering has clas-
sication efciency better than various algorithms including the
algorithm proposed in this paper and making use of rule-based
decision tree.

8. Conclusions

This paper presents a Stockwells transform based algorithm of

Fuzzy C-means clustering initialized by decision tree for recogni-
Fig. 20. Experimental set up for real time results.
tion of power quality disturbances. The results of this algorithm
has been compared with the ruled decision tree algorithm based
disturbance. Performance under noisy environment is evaluated on S-transform. The various plots obtained with the help of S-
with Gaussian noise of 20 dB SNR. Table 5, presents performance of transform can effectively be utilized for detection of power quality
algorithm in terms of real time results and comparison with sim- disturbances. Performance of detection of PQ disturbances has been
ulated results. The percentage deviation of real time results from improved using the proposed sum absolute values curve. Various
simulation results is described using a proposed relation as detailed features extracted from these curves are used effectively for detec-
below. tion and classication of power quality disturbances. Out of the
 SR RT proposed algorithms, the FCM clustering initialized by decision tree
D= 100% (11) algorithm has been found to be more effective even in noisy envi-
SR
ronment. The S-transform based features are divided into two sets
where SR: simulation result; RT: real time result. From Table 5, it is i.e. F1F8 and F9F13. First set of features (F1F8) are used by algo-
clearly observed that simulated results are very close to real time rithm based on ruled decision tree. However, the second set of
results with an error below 3%. features (F9F13) is used by algorithm of FCM clustering initial-
ized by decision tree. PQ disturbances are localized with the help
7. Performance comparison of proposed feature F14.
A set of hundred data related to each PQ disturbance obtained by
The comparison of efciency of methods proposed in this varying the parameters is used to effectively test the performance
paper with the accuracy of algorithms reported in the references of proposed algorithms. Performances of these algorithms are com-
[3,4,10,12,16,18] in terms of classication of PQ disturbances is pre- pared with algorithms already reported in literature. Efciency of
sented in this section. These articles are selected because almost proposed algorithm of FCM clustering initialized by decision tree is
all the disturbances considered in the present study had also been found to be above 99% even in noisy environment. Validity of sim-
investigated in these articles. Performance comparison of various ulation results has been veried through comparisons with real
256 O.P. Mahela, A.G. Shaik / Applied Soft Computing 59 (2017) 243257

Table 6
Comparative study of performance of PQ classication algorithms.

Reference Type of technique No. of PQ disturbances Overall percentage efciency (%)

Without noise Noise with 20 dB SNR

[3] Fast ST + FC + BFO 9.00 99.80 99.20

[4] DWT + wavelet networks 16.00 98.18 97.87
[10] ST + DT + ANN 13.00 99.90
[12] ST + Extreme learning machine 5.00 99.50 99.09
[16] SSD + HD 18.00 97.11 95.75
[18] ST + SVM 14.00 92.30
Proposed ST + DT 10.00 99.20 98.50
Proposed ST + DT + FCM 10.00 99.60 99.30

time results obtained from the RTDS. These results in real time are [14] M. Valtierra-Rodriguez, R. de Jesus Romero-Troncoso, R.A. Osornio-Rios, A.
very close to respective simulated results with an error less than 3% Garcia-Perez, Detection and classication of single and combined power
quality disturbances using neural networks, IEEE Trans. Ind. Electron. 61 (5)
which shows the effectiveness of the algorithm. Thus, it is estab- (2014) 24732482, http://dx.doi.org/10.1109/TIE.2013.2272276.
lished that decision tree initialized FCM clustering based algorithm [15] A.C. Barros, M.S. Tonelli-Neto, J.G.M.S. Decanini, C.R. Minussi, Detection and
using S-transform based features can effectively be utilized for clas- classication of voltage disturbances in electrical power systems using a
modied euclidean artmap neural network with continuous training, Electr.
sication of various investigated PQ disturbances. This algorithm Power Componen. Syst. 43 (19) (2015) 21782188, http://dx.doi.org/10.1080/
has the advantage of high classication efciency even in noisy 15325008.2015.1073814.
environment with minimum effect of noise. The high accuracy and [16] M.S. Manikandan, S.R. Samantaray, I. Kamwa, Detection and classication of
power quality disturbances using sparse signal decomposition on hybrid
low computational burden of this method make it feasible, effective
dictionaries, Instrum. Meas. IEEE Trans. 64 (1) (2015) 2738, http://dx.doi.
and an intelligent technique for PQ monitoring devices. The future org/10.1109/TIM.2014.2330493.
work includes the application of proposed algorithm on the signals [17] T. Nguyen, Y. Liao, Power quality disturbance classication utilizing
S-transform and binary feature matrix method, Electr. Power Syst. Res. 79 (4)
in hybrid power system including renewable energy sources.
(2009) 569575, http://dx.doi.org/10.1016/j.epsr.2008.08.007.
[18] B.K. Panigrahi, P.K. Dash, J.B.V. Reddy, Hybrid signal processing and machine
intelligence techniques for detection, quantication and classication of
References power quality disturbances, Eng. Appl. Artif. Intell. 22 (3) (2009) 442454,
http://dx.doi.org/10.1016/j.engappai.2008.10.003.
[1] S. Khokhar, A.A.B.M. Zin, A.S.B. Mokhtar, M. Pesaran, A comprehensive [19] Y.-L. Chen, H.-W. Hu, K. Tang, Constructing a decision tree from data with
overview on signal processing and articial intelligence techniques hierarchical class labels, Expert Syst. Appl. 36 (3, Part 1) (2009) 48384847,
applications in classication of power quality disturbances, Renew. Sustain. http://dx.doi.org/10.1016/j.eswa.2008.05.044.
Energy Rev. 51 (2015) 16501663, http://dx.doi.org/10.1016/j.rser.2015.07. [20] T.C.-K. Huang, W.-H. Hsu, Y.-L. Chen, Conjecturable knowledge discovery: a
068. fuzzy clustering approach, Fuzzy Sets Syst. 221 (2013) 123, http://dx.doi.org/
[2] M. Uyar, S. Yildirim, M.T. Gencoglu, An expert system based on S-transform 10.1016/j.fss.2012.12.006, Theme: clustering.
and neural network for automatic classication of power quality [21] R. Kumar, B. Singh, D.T. Shahani, Symmetrical components-based modied
disturbances, Expert Syst. Appl. 36 (3, Part 2) (2009) 59625975, http://dx. technique for power-quality disturbances detection and classication, IEEE
doi.org/10.1016/j.eswa.2008.07.030. Trans. Ind. Appl. 52 (4) (2016) 34433450, http://dx.doi.org/10.1109/TIA.
[3] B. Biswal, H.S. Behera, R. Bisoi, P.K. Dash, Classication of power quality data 2016.2536665.
using decision tree and chemotactic differential evolution based fuzzy [22] S. Samantaray, P. Achlerkar, M.S. Manikandan, Variational mode
clustering, Swarm Evol. Comput. 4 (2012) 1224, http://dx.doi.org/10.1016/j. decomposition and decision tree based detection and classication of power
swevo.2011.12.003. quality disturbances in grid-connected distributed generation system, IEEE
[4] M.A.S. Masoum, S. Jamali, N. Ghaffarzadeh, Detection and classication of Trans. Smart Grid PP (99) (2016) 1, http://dx.doi.org/10.1109/TSG.2016.
power quality disturbances using discrete wavelet transform and wavelet 2626469.
networks, Sci. Meas. Technol. IET 4 (4) (2010) 193205, http://dx.doi.org/10. [23] K. Thirumala, S.P. Maganuru, T. Jain, A. Umarikar, Tunable-q wavelet
1049/iet-smt.2009.0006. transform and dual multiclass SVM for online automatic detection of power
[5] O.P. Mahela, A.G. Shaik, N. Gupta, A critical review of detection and quality disturbances, IEEE Trans. Smart Grid PP (99) (2016) 1, http://dx.doi.
classication of power quality events, Renew. Sustain. Energy Rev. 41 (2015) org/10.1109/TSG.2016.2624313.
495505, http://dx.doi.org/10.1016/j.rser.2014.08.070. [24] L.C.M. Andrade, M. Oleskovicz, R.A.S. Fernandes, Adaptive threshold based on
[6] D. Granados-Lieberman, R.J. Romero-Troncoso, R.A. Osornio-Rios, A. wavelet transform applied to the segmentation of single and combined power
Garcia-Perez, E. Cabal-Yepez, Techniques and methodologies for power quality disturbances, Appl. Soft Comput. 38 (2016) 967977, http://dx.doi.
quality analysis and disturbances classication in power systems: a review, org/10.1016/j.asoc.2015.10.061.
IET Generat. Trans. Distrib. 5 (4) (2011) 519529, http://dx.doi.org/10.1049/ [25] A.A. Abdelsalam, A.A. Eldesouky, A.A. Sallam, Classication of power system
iet-gtd.2010.0466. disturbances using linear Kalman lter and fuzzy-expert system, Int. J. Electr.
[7] P. Thakur, A.K. Singh, Signal processing and AI based diagnosis of power Power Energy Syst. 43 (1) (2012) 688695, http://dx.doi.org/10.1016/j.ijepes.
quality disturbances: a review, Energy Economics and Environment (ICEEE), 2012.05.052.
2015 International Conference on (2015) 16, http://dx.doi.org/10.1109/ [26] O.P. Mahela, A.G. Shaik, Power quality recognition in distribution system with
EnergyEconomics.2015.7235071. solar energy penetration using S-transform and fuzzy C-means clustering,
[8] M.B. Latran, A. Teke, A novel wavelet transform based voltage sag/swell Renew. Energy 106 (2017) 3751, http://dx.doi.org/10.1016/j.renene.2016.12.
detection algorithm, Int. J. Electr. Power Energy Syst. 71 (2015) 131139, 098.
http://dx.doi.org/10.1016/j.ijepes.2015.02.040. [27] M. Biswal, P.K. Dash, Measurement and classication of simultaneous power
[9] S. Santoso, E.J. Powers, W.M. Grady, P. Hofmann, Power quality assessment via signal patterns with an S-transform variant and fuzzy decision tree, IEEE
wavelet transform analysis, Power Deliv. IEEE Trans. 11 (2) (1996) 924930, Trans. Ind. Inf. 9 (4) (2013) 18191827, http://dx.doi.org/10.1109/TII.2012.
http://dx.doi.org/10.1109/61.489353. 2210230.
[10] R. Kumar, B. Singh, D.T. Shahani, A. Chandra, K. Al-Haddad, Recognition of [28] O.P. Mahela, A.G. Shaik, Recognition of power quality disturbances using
power-quality disturbances using S-transform-based ann classier and S-transform and rule-based decision tree, 2016 IEEE 1st International
rule-based decision tree, Ind. Appl. IEEE Trans. 51 (2) (2015) 12491258, Conference on Power Electronics, Intelligent Control and Energy Systems
http://dx.doi.org/10.1109/TIA.2014.2356639. (ICPEICES) (2016) 16, http://dx.doi.org/10.1109/ICPEICES.2016.7853093.
[11] P. Kanirajan, V.S. Kumar, Power quality disturbance detection and [29] O.P. Mahela, A.G. Shaik, Recognition of power quality disturbances using
classication using wavelet and {RBFNN}, Appl. Soft Comput. 35 (2015) S-transform and fuzzy C-means clustering, 2016 International Conference on
470481, http://dx.doi.org/10.1016/j.asoc.2015.05.048. Cogeneration, Small Power Plants and District Energy (ICUE) (2016) 16,
[12] H. Eristi, O. Yildirim, B. Eristi, Y. Demir, Automatic recognition system of http://dx.doi.org/10.1109/COGEN.2016.7728955.
underlying causes of power quality disturbances based on S-transform and [30] R.H.G. Tan, V.K. Ramachandaramurthy, Numerical model framework of power
extreme learning machine, Int. J. Electr. Power Energy Syst. 61 (2014) quality events, Eur. J. Sci. Res. 43 (1) (2010) 3047.
553562, doi:http://dx.doi.org/10.1016/j.ijepes.2014.04.010. [31] R.G. Stockwell, L. Mansinha, R.P. Lowe, Localization of the complex spectrum:
[13] A.K. Ghosh, D.L. Lubkeman, The classication of power system disturbance the s transform, Signal Process. IEEE Trans. 44 (4) (1996) 9981001, http://dx.
waveforms using a neural network approach, Power Deliv. IEEE Trans. 10 (1) doi.org/10.1109/78.492555.
(1995) 109115, http://dx.doi.org/10.1109/61.368408.
 O.P. Mahela, A.G. Shaik / Applied Soft Computing 59 (2017) 243257 257

[32] P.K. Ray, N. Kishor, S.R. Mohanty, Islanding and power quality disturbance [35] S.R. Samantaray, Decision tree-initialised fuzzy rule-based approach for
detection in grid-connected hybrid power system using wavelet and s power quality events classication, Generat. Trans. Distrib. IET 4 (4) (2010)
-transform, Smart Grid IEEE Trans. 3 (3) (2012) 10821094, http://dx.doi.org/ 530537, http://dx.doi.org/10.1049/iet-gtd.2009.0508.
10.1109/TSG.2012.2197642. [36] B. Biswal, P.K. Dash, K.B. Panigrahi, Power quality disturbance classication
[33] S. Ventosa, C. Simon, M. Schimmel, J.J. Danobeitia, A. Manuel, The S-transform using fuzzy C-means algorithm and adaptive particle swarm optimization,
from a wavelet point of view, Signal Process. IEEE Trans. 56 (7) (2008) Ind. Electr. IEEE Trans. 56 (1) (2009) 212220, http://dx.doi.org/10.1109/TIE.
27712780, http://dx.doi.org/10.1109/TSP.2008.917029. 2008.928111.
[34] A. Rodriguez, J.A. Aguado, F. Martin, J.J. Lopez, F. Munoz, J.E. Ruiz, Rule-based
classication of power quality disturbances using S-transform, Electr. Power
Syst. Res. 86 (2012) 113121, http://dx.doi.org/10.1016/j.epsr.2011.12.009.