IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 51, NO.

2, MARCH/APRIL 2015 1249

Recognition of Power-Quality Disturbances Using

S-Transform-Based ANN Classifier and
Rule-Based Decision Tree
Raj Kumar, Member, IEEE, Bhim Singh, Fellow, IEEE, D. T. Shahani,
Ambrish Chandra, Fellow, IEEE, and Kamal Al-Haddad, Fellow, IEEE

AbstractThis paper deals with a modified technique for the in energy intensive equipment which are mostly nonlinear in
recognition of single stage and multiple power quality (PQ) distur- nature. Voltage waveform pollution in the distribution system
bances. An algorithm based on Stockwells transform and artificial depends upon different patterns of current waveforms drawn
neural network-based classifier and a rule-based decision tree is
proposed in this paper. The analysis and classification of single by these nonlinear loads. These distorted load currents cause
stage PQ disturbances consisting of both events and variations voltage distortion at the point of common coupling (PCC) of
such as sag, swell, interruption, harmonics, transients, notch, the power distribution system. Continuous process, multistage
spike, and flicker are presented. Moreover, the proposed algorithm batch operations, and data processing are mostly affected by
is also applied on multiple PQ disturbances such as harmonics poor PQ. Any PQ disturbance halts the process in industries
with sag, swell, flicker, and interruption. A database of these
PQ disturbances based on IEEE-1159 standard is generated in such as the paper and semiconductor industries, which results
MATLAB for simulation studies. The proposed algorithm extracts in semiprocessed products and huge production loss. Moreover,
significant features of various PQ disturbances using S-trans- to reboot a large number of workstations and recover pending
form, which are used as input to this hybrid classifier for the transactions is a time-consuming process [1].
classification of PQ disturbances. Satisfactory results of effective PQ events are defined as the sudden disturbances which
recognition and classification of PQ disturbances are obtained
with the proposed algorithm. Finally, the proposed method is also occur in the power system having a beginning and an ending
implemented on real-time PQ events acquired in a laboratory to [2], and occurrences of more than one type of event at the same
confirm the validity of this algorithm in practical conditions. time are called multistage events. These are mostly caused by
Index TermsDisturbances, event, multiresolution analysis, ground faults, switching of heavy loads, capacitor switching,
power quality (PQ), S-transform, wavelet. lightening strikes, and nonlinear loads. Multiple PQ events also
occur in the power system due to the simultaneous operations of
I. I NTRODUCTION these events. Disturbances like harmonics and flicker are called
as PQ variations [2] as these continuously exit in the power

E LECTRIC power quality (PQ) covers a wide range of

electromagnetic phenomena concerning the interaction of
the power system network and end user equipment. PQ is
system. Many of PQ disturbances are generated by consumer
loads, and they propagate on the electrical network, affecting
the transmission equipment and causing the malfunction of the
further worsening due to complicated power distribution with electronic equipment of others. PQ limits that caused equip-
the increase in grid connection for the new energy resources ment to malfunction are set by certain international standards
such as wind and solar energy conversion systems. Therefore, such as IEEE-1159, IEC 61000, and EN 50160 to maintain the
the demand for the clean power is increasing simultaneously PQ to an acceptable benchmark [3], [4].
due to the increase in the use of sensitive solid state converters Automatic PQ disturbance recognition is necessary in saving
the time of power engineers for detecting, classifying, and
Manuscript received February 12, 2014; revised July 3, 2014; accepted solving the PQ issues between the utilities and the consumers.
August 25, 2014. Date of publication September 11, 2014; date of current These disturbances cause huge economic losses due to the
version March 17, 2015. Paper 2014-IACC-0076.R1, presented at the 2013 maloperation of electronic equipment. PQ disturbance recog-
IEEE Industry Applications Society Annual Meeting, Orlando, FL, USA,
October 611, and approved for publication in the IEEE T RANSACTIONS nition requires the deployment of intelligent PQ monitoring
ON I NDUSTRY A PPLICATIONS by the Industrial Automation and Control instruments in the power system. Intensive research is needed
Committee of the IEEE Industry Applications Society. for developing intelligent techniques for the PQ monitoring
R. Kumar is with the Sant Longowal Institute of Engineering and Technol-
ogy, Longowal 148 106, India (e-mail: [email protected]). instruments.
B. Singh and D. T. Shahani are with the Indian Institute of Technology Delhi, Digital signal processing techniques like the Fourier trans-
New Delhi 110 016, India (e-mail: [email protected]; dt.shahani@ form (FT), short-time Fourier transform (STFT), wavelet trans-
gmail.com).
A. Chandra and K. Al-Haddad are with the cole de Technologie Suprieure, form (WT), Kalman filter, and Stockwell transform (ST) are
Montral, QC H3C 1K3, Canada (e-mail: [email protected]; kamal. extensively used to extract the time and frequency domain
[email protected]). information from PQ disturbances for analysis. FT does not
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. perform well for monitoring PQ signals which are mostly
Digital Object Identifier 10.1109/TIA.2014.2356639 nonstationary signals due to the reliance of resolution on the

0093-9994 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1250 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 51, NO. 2, MARCH/APRIL 2015

window size. IEC 61000-4-30 and IEC 61000-4-7 recommend of the S-matrix is a complex matrix of size k n and is
and explain the use of FT for harmonic group and subgroup represented as
calculations [5], [6].
Fourier analysis [7] also loses time information while trans- S(, f) = A(, f)ei(,f) (1)
forming to the frequency domain, so it does not provide time
information simultaneously. STFT [8] decomposes the nonsta- where A(, f) is the amplitude and (, f) is the phase. The
tionary signal into the timefrequency domain by considering rows of the S-matrix are the frequencies, and the columns are
it as a concatenation of stationary signals within the sliding corresponding to the time. Each column presents the frequency
window. The timefrequency resolution depends upon the size components present in the signal at a particular time. Each
of the window. Therefore, due to the same window size for all row presents the magnitude of a particular frequency with time
frequencies, it cannot track the signal dynamics properly. varying from 0 to N 1 samples. A detailed description of the
Wavelet analysis [9][11] is a technique to vary the window S-transform is given in the Appendix.
size. It allows the use of long time intervals for precise low
frequency information and shorter regions for high frequency B. ANN
information. Wavelet exhibits some disadvantages such as its
complicated computation and sensitivity to the noise level and An ANN consists of an interconnected group of artificial
the dependence of its accuracy on the chosen wavelet. This neurons and is trained to perform a particular function by
leads to the development of ST proposed by Stockwell et al. adjusting the values of the connections (weights and biases)
[12]. ST is an invertible timefrequency spectral localization between the neurons of different layers. The neural network is
technique that combines the elements of WT and STFT. ST defined by the interconnection pattern between different layers
uses an analysis window whose width is decreasing with the of neurons, the learning process, and the activation function
frequency, providing a frequency-dependent resolution. The of the neurons. Input and target pairs are needed to train a
processed signals yield certain unique features for a given PQ neural network. The weights are adjusted, based on a com-
disturbance that are used for the automatic recognition of the parison of the output and the target, until the network output
PQ disturbance using pattern classifiers like artificial neural matches the target. The number of features extracted from
network (ANN), fuzzy logic, and support vector machine [13] transforms decides the number of neurons in the input layer
[17]. In this paper, S-transform is applied for the recognition of of the neural network. The number of neurons in the output
single and multiple PQ disturbances, and the disturbances are layer is made equal to the number of elements of the target
well recognized using a rule-based decision tree and ANN. vector.
A two-layer feedforward network with sigmoid hidden and
output neurons is used for the classification of the PQ events.
The network is trained with a scaled conjugate gradient back
II. P ROPOSED A LGORITHM OF ST-BASED ANN propagation algorithm with statistical features like the skewness
C LASSIFIER AND RULE -BASED D ECISION T REE and kurtosis of the maximum amplitude with the time vector of
the S-matrix (S(, f)) for the harmonic, oscillatory transient,
The proposed algorithm for the recognition of the power spike, and notch.
PQ disturbances consists of the S-transform matrix, rule-based
decision tree, and the training of the neural network with the
feature extracted from an S-matrix for PQ disturbance classifi- C. Rule-Based Decision Tree
cation. The following sections describe the main features of the Decision rules map the observations about an item to con-
proposed algorithm. clude about the items target value and are used to represent the
decision-making process. A rule-based decision tree represents
the decision based on the rules driven from the data.
The rows and columns of the S-matrix for a particular event
A. S-Transform are investigated to find the rules to take the decisions for the
S-transform is one of the best techniques for performing classification of the PQ events.
the signal processing of nonstationary signals. S-transform is
an extension of the continuous wavelet transform (CWT) as it
III. S INGLE S TAGE AND M ULTIPLE PQ
utilizes the information contained in the phase of the CWT by
D ISTURBANCE A NALYSIS
modifying the phase of the mother wavelet. It produces a mul-
tiresolution analysis like a bank of filters with constant relative This section deals with the generation and analysis of the
bandwidth. It uniquely combines a frequency-dependent reso- PQ disturbances using the proposed algorithm in a MATLAB
lution while simultaneously localizing the real and imaginary environment.
spectra. The timefrequency localizing property is provided
by this transform in computing both the amplitude and phase
A. Generation of PQ Disturbances
spectra of discrete data samples.
The S-matrix is computed for each PQ event by taking the A database of single and multiple PQ disturbances is required
advantage of the FFT and convolution theorem. The output for the recognition of these disturbances. Synthetic data of PQ
KUMAR et al.: RECOGNITION OF PQ DISTURBANCE USING ANN CLASSIFIER AND RULE-BASED DECISION TREE 1251

TABLE I
N UMERICAL M ODELING OF THE S IMULATED PQ D ISTURBANCES [18]

disturbances are advantageous in estimating the generalization resulting in a matrix of size 640 640 elements with a 50-Hz
ability of the classifier, so they are accepted and widely used power frequency. These disturbances are characterized by their
[11], [16]. The availability of the real PQ disturbances data magnitude and frequency during their duration. The frequency
is limited as it requires a long monitoring time simultaneously resolution of this matrix depends upon the number of samples
at the number of locations with uncertainty in the occurrence taken for the FFT computation and the sampling frequency. In
of the PQ events. Therefore, numerical models of the PQ this application, the frequency resolution is calculated as 5 Hz.
disturbances are used for the simulation in MATLAB. These However, this resolution can be increased at the cost of a higher
simulated PQ disturbances obtained from numerical models number of computations.
closely depict the real-time data and can be generated as per The complex form of the S-transform matrix is calculated
the international standards. Multiple PQ disturbances are gen- for each PQ disturbance, and then, the absolute S-matrix is
erated from summation of numerical model of single stage PQ obtained for extracting the timefrequency information. The
disturbances. The numerical modeling of single and multiple maximum value of each column gives the information regard-
PQ disturbances [18] is presented in Table I. ing the main frequency present in the signal at a particular time.
The maximum value of each row gives the information regard-
ing the maximum amplitude of that frequency in the whole
B. PQ Disturbance Analysis
sample time. The frequency of the disturbance can be found by
The ST-based classifier presented in this paper is designed multiplying the sampling frequency with the frequency scale
to recognize eight single stage and four multiple PQ distur- value corresponding to the maximum amplitude of the row.
bances like sag, swell, interruption, flicker, oscillatory tran- The scale value is the normalized frequency value which is
sients, notch, spike harmonics, sag and harmonic, swell and the ratio of the frequency amplitude to the sampling frequency.
harmonic, flicker with harmonic, and interruption with har- Figs. 1(a)7(a) show the three types of single stage and four
monic. PQ disturbances are simulated in MATLAB as per types of multistage PQ disturbance signals. Figs. 1(b)7(b)
standard IEEE-1159 [3] with a sampling frequency of 3.2 kHz. show the maximum amplitude versus sample (time) contour by
The disturbances are simulated in a data window of ten cycles, searching the columns of the S-matrix. Figs. 1(c)7(c) show the
1252 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 51, NO. 2, MARCH/APRIL 2015

Fig. 1. (a) Oscillatory transient. (b) Maximum amplitude versus time contour. Fig. 5. (a) Voltage swell with harmonics. (b) Maximum amplitude versus time
(c) Amplitude versus frequency (normalized) contour. contour. (c) Amplitude versus frequency (normalized) contour.

Fig. 2. (a) Voltage spikes. (b) Maximum amplitude versus time contour. Fig. 6. (a) Voltage flicker with harmonics. (b) Maximum amplitude versus
(c) Amplitude versus frequency (normalized) contour. time contour. (c) Amplitude versus frequency (normalized) contour.

Fig. 3. (a) Voltage notches. (b) Maximum amplitude versus time contour.
Fig. 7. (a) Voltage interruption with harmonics. (b) Maximum amplitude
(c) Amplitude versus frequency (normalized) contour.
versus time contour. (c) Amplitude versus frequency (normalized) contour.

IV. C LASSIFICATION OF PQ D ISTURBANCES

PQ disturbance classification starts from the analysis of the
maximum amplitude versus time and frequency amplitude ver-
sus frequency (normalized) contours of the S-matrix obtained
for each PQ disturbance. The complete block scheme of the
classification algorithm is illustrated in Fig. 8. Broadly, as seen
from visual inspection, PQ disturbances can be separated into
two groups by counting the number of peaks in the frequency-
amplitude vector of the S-matrix. Peaks confirm the presence
of more than one frequency present in the disturbance signal.
Fig. 4. (a) Voltage sag with harmonics. (b) Maximum amplitude versus time PQ disturbance sag, swell, flicker, interruption, and pure sine
contour. (c) Amplitude versus frequency (normalized) contour. have only a single peak in the frequency-amplitude vector, so
they fall in the same group G1 . While transient, flicker, notch,
contour for the amplitude of the various frequencies present in harmonics, harmonic sag, harmonic swell, harmonics flicker,
the full disturbance signal. and harmonics interruption have more than a single peak in the
KUMAR et al.: RECOGNITION OF PQ DISTURBANCE USING ANN CLASSIFIER AND RULE-BASED DECISION TREE 1253

Fig. 8. Block schemes for the classifications of single and multistage PQ disturbances.

frequency-amplitude matrix, so all these disturbances belong to from the maximum value of the maximum amplitude versus
the second group G2 . time contour. The flicker factor can be estimated as [19]
G1 disturbances can be classified further by finding the mini-
mum and the maximum value of the maximum amplitudetime 
vector. The count of the values deviating from the set limits of s=Sub.har.F req.Ind |S[jT, fs ]|2
the amplitudetime vector indicates the type of the disturbance. Ff = . (4)
|S[jT, f1 ]|
The multiplication of the sample time and the counting of the
number provide the duration of that PQ event. These events can
be classified as short term and long term based on this time G2 disturbances can be classified into the G21 group with
information. PQ indices which represent the severity of the PQ disturbances like harmonics, spikes, and notches having unit
disturbances can be estimated from the S-matrix. The magni- amplitude during the entire data window and the G22 group
tude and the duration of the sag, swell, and interruption events with variation in the amplitudetime vector. Transient has al-
can be estimated from the amplitude time contour. Threshold most near to unit amplitude, so it can be grouped with the G21
and hysteresis limits for the voltage sag, swell, and interruptions disturbances. Group G22 consists of harmonic sag, harmonic
are defined in IEC-61000-4-30 [6]. The instantaneous rms and swell, harmonics flicker, and harmonics interruption. The dis-
sag energy index can be estimated from the amplitude time turbance of the G22 group can be recognized again by finding
contour as [19] the count of values from the threshold settings of the maxi-
  
mum amplitude vector of the S-matrix. Instantaneous harmonic

1  h distortion is estimated from the S-matrix as [19]
ViRM S =  |S[jT, fn ]| 2
(2)
2 n=1 

h
|S(jT, fn )|2
where f1 is the fundamental frequency and h is the highest n=2
harmonic order iHD = (5)
|S(jT, f1 )|

2
VSag
EV SAG = 1 TED (3) where h is the highest order frequency component.
Vn
The disturbances of the G21 group are recognized by the
where VSag is the sag depth, TED is the event duration, and training of the ANN. Statistical features like the skewness and
Vn is the nominal voltage swell energy that can be estimated kurtosis of the maximum amplitude with the time vector of the
1254 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 51, NO. 2, MARCH/APRIL 2015

TABLE II
C ONFUSION M ATRIX FOR THE E VENTS OF G21 G ROUP

S-matrix for these disturbances are taken as the input to the

neural network. Skewness and kurtosis are defined as

1  (xi x)3
N
Skewness, S = (6)
N 1 i=1 3
1  (xi x)4
N
Kurtosis, K = (7)
N 1 i=1 4

where N is the number of samples.

The neural network input layer consists of two neurons and
four neurons in the output layer. Features extracted from the
S-matrix of each of two hundred disturbance signals of each
event are used to train the neural network with 70% data for
testing and 15% data each for the validation and testing. The
sigmoid function is used as the activation function of the hidden
and the output layer neurons. A scaled complex conjugate
algorithm is used to the train the ANN. This algorithm does
not require any user-dependent parameter and is much faster as
it avoids the time-consuming line search per learning iteration
[19]. The classification accuracy is checked with the different
number of neurons in the hidden layer. The neural network is
able to classify these events with the overall accuracy of 99.9%
as shown in Table II. Harmonic, spike, and notch events are Fig. 9. (a) Laboratory prototype of the developed hardware for real-time PQ
classified with 100% accuracy, while the oscillatory transient disturbance generation and (b) the actual setup.
event is classified with an accuracy of 99.5%. voltage and current sensing. A signal conditioning circuitry for
these sensors is also developed for the protection of NI-cDAQ
from any voltage spike. These generated PQ disturbances are
V. H ARDWARE I MPLEMENTATION OF P ROPOSED
logged in MATLAB with a sampling frequency of 3.2 kHz
A LGORITHM FOR PQ D ISTURBANCE
via the NI-9215 module. These disturbance matrices have been
D ETECTION AND C LASSIFICATION
reshaped in the size of ten power cycles for carrying out the
There can be a number of reasons for the generation of the timefrequency analysis with S-transform.
same PQ disturbance in real world. Real-time PQ disturbances
have been generated in the laboratory by switching a number
A. Generation of PQ Disturbances
of linear and nonlinear loads such as in a practical distribution
system. Fig. 9(a) shows the developed hardware prototype for A three-phase rectifier and capacitor switching are used
the generation and acquisition of real-time PQ disturbances, to generate the voltage harmonics, oscillatory transients, and
and Fig. 9(b) shows the actual setup. A National Instruments notches, respectively. Sag and swell have been recorded by
Compact Data Acquisition Chassis (NI-cDAQ) with 9215 DAQ sensing the voltage across the inductor by switching the high
cards is used for the high speed acquisition of the PQ distur- rating linear load. Figs. 10(a)13(a) show the measured PQ
bances. A three-phase supply followed by series inductors and disturbances generated in the laboratory, and the analysis of
linear/nonlinear loads is used for the generation of selected these waveforms is shown in Figs. 10(b)13(b). The voltage
PQ disturbances such as sag, swell, etc. The voltage sensor sag with a magnitude of 0.85 p.u. is obtained after switching a
(LEM-make LV-25P) and current sensor (LA-55P) are used for high rating linear load as shown in Fig. 10(a).
KUMAR et al.: RECOGNITION OF PQ DISTURBANCE USING ANN CLASSIFIER AND RULE-BASED DECISION TREE 1255

Fig. 10. (a) Generation of real-time voltage sag signal and its (b) S-transform Fig. 12. (a) Generation of real-time voltage notch signal and its
analysis. (b) S-transform analysis.

Fig. 11. (a) Generation of real-time voltage swell signal and its Fig. 13. (a) Generation of real-time voltage transient signal and its
(b) S-transform analysis. (b) S-transform analysis.

In a similar way, a voltage swell with a magnitude of notches are generated by the switching of a three-phase rectifier
1.15 p.u. is obtained as shown in Fig. 11(a). Moreover, voltage as shown in Fig. 12(a). Finally, the voltage transients are
1256 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 51, NO. 2, MARCH/APRIL 2015

TABLE III trix clearly depict the disturbance patterns of the power system.
C OMPARISON OF S-T RANSFORM P ERFORMANCE W ITH
D IFFERENT C LASSIFIERS Features extracted from contours of the S-matrix are effectively
used for the automatic classification of the PQ disturbances.
The neural network is able to classify the disturbances with
accuracy more than 99% for harmonic, oscillatory transient,
spike, and notch events. In conclusion, using features from
the S-transform, which facilitates the use of multiresolution
analysis along with the ANN and rule-based decision tree, is
an effective and intelligent technique for the PQ monitoring
instruments.

A PPENDIX
Continuous S-Transform: The CWT W (, d) of a function
h(t) is defined as [12]

w(, d) = h(t)w(t , d)dt. (8)

produced by capacitor switching as shown in Fig. 13(a). These The scale parameter d determines the width and thus con-
PQ disturbances are measured and recorded in MATLAB at a trols the resolution. gives the position of the wavelet. The
sampling frequency of 3.2 kHz. S-transform is defined as a CWT with a specific mother wavelet
multiplied by the phase factor as
B. Application of Proposed Algorithm S(, f ) = ei2f (t)w(t, d) (9)
The real-time PQ disturbances generated in the laboratory are where the mother wavelet w(, d) is defined as
recorded via NI-9215 cards. These disturbances are acquired, f t2 f 2
w(t, f ) = e 2 ei2f t . (10)
and a data bank is generated of these events, each for a couple 2
of cycles. These data banks are further divided into submatrices The frequency f is the inverse of scale parameter d. Therefore,
for handling the data efficiently. The proposed algorithm is the continuous S-transform is obtained as
applied on these disturbance patterns. 
|f | t2 f 2
S(, f ) = h(t) e 2 ei2f t dt (11)
C. Results and Discussion 2

Figs. 10(b)13(b) show the results obtained on the applica- and the width of the Gaussian window is (f ) = T = 1/|f |.
tion of the proposed algorithm for four different cases of voltage The relationship between the S-transform and FT can be
sag, voltage swell, voltage notches, and voltage oscillatory written as
transients. As shown in these figures, the maximum amplitude  2 2
2
versus time contour and amplitude versus frequency contour S(, f ) = H( + f )e f 2 ei2 dt f = 0. (12)
of these events clearly depict the magnitude and duration of
the disturbances. The features extracted from these contours
By taking advantage of the fast FT and the convolution theorem,
are effectively classifying the single-stage and multistage PQ
the discrete analog of the aforementioned equation can be used
disturbances. The classification accuracy of the S-transform
to compute the discrete S-transform.
with different classifiers is compared and presented in Table III.
Discrete S-Transform: The PQ disturbance signal h(t) can
The proposed classifier is able to classify 13 PQ disturbances
be sampled as h(kT ), where T is the sampling interval time,
with an accuracy of 99.9% which is higher than others. These
and the discrete Fourier transform of the sampled signal h(kT ),
obtained results show a satisfactory performance of this algo-
where k = 0 to N 1 is as
rithm for classifications of various single-stage and multistage
PQ disturbances.  n  N 1
1 
h(kT )e N
i2nk
H = (13)
NT N
VI. C ONCLUSION k=0

Single-stage and multistage disturbances have been effec- where n = 0, 1, . . . N 1. The discrete version of the
tively classified with the S-transform-based ANN classifier and S-transform is obtained as jT and is f n/(N T ) as
rule-based decision tree. PQ disturbances have easily been   
n  
N 1
detected and localized as amplitude, frequency, and time in- m+n i2mk
S jT, = H G(m, n)e N n = 0
formation of the PQ signals which is easily obtained with NT m=0
NT
S-transform. S-matrix contours like maximum amplitude versus (14)
where G(m, n) = e(2 m /n ) .
2 2 2
time and amplitude versus frequency from the S-transform ma-
KUMAR et al.: RECOGNITION OF PQ DISTURBANCE USING ANN CLASSIFIER AND RULE-BASED DECISION TREE 1257

For n = 0, (14) is modified as [20] I. W. C. Lee and P. K. Dash, S-transform-based intelligent system for
classification of power quality disturbance signals, IEEE Trans. Ind.
1   m 
N 1 Electron., vol. 50, no. 4, pp. 800805, Aug. 2003.
[21] M. V. Chilukuri and P. K. Dash, Multiresolution S-transform-based fuzzy
S[jT, 0] = h (15) recognition system for power quality events, IEEE Trans. Power Del.,
N m=0 NT
vol. 19, no. 1, pp. 323330, Jan. 2004.
[22] M. Biswal and P. K. Dash, Detection and characterization of multiple
where j, m, and n = 0, 1, . . . N 1. power quality disturbances with a fast S-transform and decision tree
based classifier, J. Digit. Signal Process., vol. 23, no. 4, pp. 10711083,
Equation (11) makes the constant average of the time series Jul. 2013.
into the zero frequency voice, thus assuring that the inverse is
exact.
The inverse of the discrete S-transform is given as
 
1    n  i2nk
N 1 N 1 Raj Kumar (M13) received the Bachelor of En-
h[jT ] = S jT, e N . (16) gineering (Electronics and Instrumentation) degree
N n=0 NT from Punjabi University, Patiala, India, in 1994 and
k=0
the M.Tech. degree in instrumentation and con-
trol engineering from Punjab Technical University,
Jalandhar, India, in 2005. He is currently working
toward the Ph.D. degree at the Indian Institute of
R EFERENCES Technology Delhi, New Delhi, India.
[1] A. Chapman, Costs of poor power quality, in Power Quality Appli- He is currently an Assistant Professor with the
cation Guide. Hertfordshire, U.K.: Copper Development Association, Electrical and Instrumentation Engineering Depart-
Mar. 2001. ment of Sant Longowal Institute of Engineering and
[2] M. H. J. Bollen and I. Y.-H. Gu, Signal Processing of Power Quality Technology, Longowal, India. His areas of research include power quality,
Disturbances. New York, NY, USA: Wiley-IEEE Press, 2006. digital signal processing, and process control systems.
[3] Recommended Practice for Monitoring Electric Power Quality, IEEE
1159:1995.
[4] Voltage Characteristics of Electricity Supplied by Public Distribution
Systems, EN Std. 50160, 1999.
[5] Testing and Measurement TechniquesPower Quality Measurement
Bhim Singh (SM99F10) received the Bachelor of
Methods, IEC Std. 61000-4-30, 2003.
Engineering (Electrical) degree from the University
[6] Testing and Measurement TechniquesGeneral Guide on Harmonics and
of Roorkee, Roorkee, India, in 1977 and the M.Tech.
Interharmonics Measurements and Instrumentation for Power Supply,
and Ph.D. degrees from the Indian Institute of Tech-
IEC Std. 6100-4-7, 2002.
nology Delhi (IIT Delhi), New Delhi, India, in 1979
[7] A. Rodriguez et al., Time-frequency transforms comparison for power
and 1983, respectively.
quality analysis, in Proc. 5th Int. Conf. EEM, 2008, pp. 16. In 1983, he joined the Department of Electrical
[8] Y. H. Gu and M. H. J. Bollen, Time-frequency and time scale domain
Engineering, University of Roorkee, as a Lecturer.
analysis of voltage disturbance, IEEE Trans. Power Del., vol. 15, no. 4,
He became a Reader there in 1988. In December
pp. 12791284, Oct. 2000.
1990, he joined the Department of Electrical Engi-
[9] S. Santoso, E. J. Powers, W. M. Grady, and P. Hofmann, Power quality neering, IIT Delhi, as an Assistant Professor, where
assessment via wavelet transform analysis, IEEE Trans. Power Del.,
he became an Associate Professor in 1994 and a Professor in 1997. He has
vol. 11, no. 2, pp. 924930, Apr. 1996.
guided 50 Ph.D. dissertations and 150 M.E./M.Tech. theses. He has been
[10] F. Jurado and J. R. Saenz, Comparison between discrete STFT and
granted one U.S. patent and filed ten Indian patents. He has executed more than
wavelets for the analysis of power quality events, Elect. Power Syst. Res., 70 sponsored and consultancy projects. His fields of research interest include
vol. 62, no. 3, pp. 183190, Jul. 2002.
power electronics, electrical machines, electric drives, power quality, flexible
[11] M. Gouda, M. M. A. Salama, M. R. Sultan, and A. Y. Chikhani, Power
ac transmission systems, high-voltage direct-current transmission systems, and
quality detection and classification using wavelet multiresolution signal
renewable energy generation.
decomposition, IEEE Trans. Power Del., vol. 14, no. 4, pp. 14691476, Prof. Singh is a Fellow of the Indian National Academy of Engineering, the
Oct. 1999.
National Science Academy (NSc), the Indian Academy of Science (IASc), the
[12] R. G. Stockwell, L. Mansinha, and R. P. Lowe, Localization of the
Institution of Engineering and Technology, U.K., the Institution of Engineers
complex spectrum: The S-transform, IEEE Trans. Signal Process., vol. 4,
(India), the World Academy of Sciences (FTWAS), the Indian National Science
no. 4, pp. 9981001, Apr. 1996.
Academy, and the Institution of Electronics and Telecommunication Engineers.
[13] P. K. Dash, B. K. Panigrahi, and G. Panda, Power quality analysis us-
ing S-transform, IEEE Trans. Power Del., vol. 18, no. 2, pp. 406411,
Apr. 2003.
[14] P. Janik and T. Lobos, Automated classification of power quality distur-
bances using SVM and RBF networks, IEEE Trans. Power Del., vol. 21,
no. 3, pp. 16631669, Jul. 2006. D. T. Shahani received the Bachelor of Technol-
[15] F. Z. Zhao and R. G. Yang, Power-quality disturbance recognition us- ogy (Electronics) degree from the Indian Institute of
ing S-transform, IEEE Trans. Power Del., vol. 22, no. 2, pp. 944950, Technology, Kharagpur, India, in 1973 (silver medal-
Apr. 2007. ist) and the Ph.D. degrees from the Indian Institute of
[16] S. Mishra, C. N. Bhende, and B. K. Panigrahi, Detection and classifi- Technology Delhi (IIT Delhi), New Delhi, India, in
cation of power quality disturbances using S-transform and probabilistic 1979.
neural network, IEEE Trans. Power Del., vol. 23, no. 1, pp. 280287, He has been associated with IIT Delhi since 1973,
Jan. 2008. and currently, he is a Professor in the Instrument
[17] M. Uyar, S. Yildirim, and M. T. Gencoglu, An expert system based on S- Design and Development Centre. He has a number
transform and neural network for automatic classification of power quality of research, consultancies, and numerous transfers of
disturbances, Expert Syst. Appl., vol. 36, pt. 2, no. 3, pp. 59625975, technologies to his credit. His areas of specialization
Apr. 2009. are electronics, instrumentation, electromagnetic transducers, and EMI/EMC.
[18] R. H. G. Tan and V. K. Ramachandramurthy, Numerical model frame- Prof. Shahani was the recipient of the prestigious award of Excellence in
work of power quality events, Eur. J. Sci. Res., vol. 43, no. 1, pp. 3047, Teaching (IIT Delhi, Golden Jubilee Year) 2011. He was also the recipient
Jun. 2010. of the Outstanding Development and Technology award in Electro Magnetic
[19] M. F. Moller, A scaled conjugate gradient algorithm for fast supervised Systems and of the Outstanding Instructional Element Award (teaching at IIT
learning, Neural Netw., vol. 6, no. 4, pp. 525533, 1993. Delhi) in 1991.
1258 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 51, NO. 2, MARCH/APRIL 2015

Ambrish Chandra (SM99F14) received the B.E. Kamal Al-Haddad (S82M88SM92F07)

degree from the University of Roorkee, Roorkee, received the B.Sc.A. and M.Sc.A. degrees from
India, in 1977, the M.Tech. degree from the Indian the University of Quebec at Trois-Rivieres,
Institute of Technology Delhi, New Delhi, India, in Trois-Rivieres, QC, Canada, in 1982 and 1984,
1980, and the Ph.D. degree from the University of respectively, and the Ph.D. degree from the Institut
Calgary, Calgary, AB, Canada, in 1987. National Polytechnique, Toulouse, France, in 1988.
He was a Lecturer and later a Reader at the Univer- Since June 1990, he has been a Professor with the
sity of Roorkee. Since 1994, he has been a Professor Electrical Engineering Department, Ecole de Tech-
in the Electrical Engineering Department at the cole nologie Superieure (ETS), Montreal, QC, Canada,
de Technologie Suprieure, Universit du Qubec, where he has been the holder of the Canada Research
Montral, QC, Canada. His main research interests Chair in Electric Energy Conversion and Power Elec-
are power quality, active filters, static reactive power compensation, flexible tronics since 2002. He is the Chief of the ETS Bombardier Transportation North
ac transmission systems, and the control and integration of renewable energy America division, a joint industrial research laboratory on electric traction
resources. system and power electronics. His research interests are highly efficient static
Dr. Chandra is a Fellow of the Institution of Engineering and Technology, power electronic converter design, harmonics, and reactive power control using
U.K., the Institution of Engineers (India), and the Institution of Electronics power filters, active rectifiers, and resonant converters.
and Telecommunication Engineers. He is also a Life Member of the Indian Prof. Al-Haddad is a Fellow of the Canadian Academy of Engineering and a
Society for Technical Education. He is a Professional Engineer in the Province Life Member of the Circle of Excellence of the University of Quebec and was
of Quebec, Canada. the recipient of the outstanding researcher award from ETS in 2000.