This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2626469, IEEE
Transactions on Smart Grid
1

Variational Mode Decomposition and Decision Tree

Based Detection and Classification of Power
Quality Disturbances in Grid-Connected Distributed
Generation System
Pankaj D. Achlerkar, S. R. Samantaray, Senior Member, IEEE, and M. Sabarimalai Manikandan, Member, IEEE

Abstract—This paper presents a variational mode decomposi- monitoring is essential for online applications like compensa-
tion (VMD) and decision tree based detection and classification tion and protection. Also, the near to real time monitoring is
method of single and mixed power quality (PQ) disturbances in essential for offline processes like post disturbance analysis,
grid-connected distributed generation (DG) system. Applicability
of VMD technique is investigated for discrimination of stationary planning for installation of compensating devices, planning
PQ disturbances (such as harmonics, interharmonics and flicker), for improvements in existing system, performance checking
non-stationary events (e.g. transients) and noise. Studies indicate of existing compensation schemes, etc. Monitoring includes
usefulness of VMD for accurate estimation of phasor quanti- deriving proper indices to quantify systems performance in
ties such as amplitude, phase angle and frequency and other presence of disturbance and this quantification requires exact
describing features. Features namely mode central frequencies
(MCFs), relative mode energy ratios (RMERs), zero crossings and determination of magnitude, phase, frequency contents of
instantaneous amplitude (IA) are extracted for classification of disturbance signal and time instant at which disturbance is
single and mixed PQ disturbances using a decision tree algorithm. occurring. [14]
A set of synthetic test signals, disturbance signals obtained from Various methods are studied for detection and classification
real events as well as signals generated from Real Time Digital of PQ disturbances due to the variations in linear and non-
Simulator (RTDS) platform are used to test effectiveness of
proposed method, under various system operating scenarios and linear loads, and the environmental factors [2]–[9]. Curve-
noise levels. Proposed method is found to be capable of accurate fitting, and frequency estimation approaches, such as Kalman
detection, estimation, localization and classification of all kinds filtering, Prony analysis, matrix pencil, recursive least square
of PQ disturbances in both noisy and noise-free cases. (RLS), recursive and non-recursive Newton algorithms are
Index Terms—Power quality (PQ) analysis, distributed gener- studied to analyse PQ disturbances. Most PQ disturbance
ation (DG), variational mode decomposition, harmonic analysis detection and classification methods are developed based on
the signal processing techniques such as Fourier transform
(FT), chirp Z-transform (CZT), short-time Fourier transform
(STFT), discrete wavelet transform (DWT) and S-transform
I. I NTRODUCTION
(ST) [2], [5], hyperbolic S-transform (HST) [3], time-time

R ECENTLY, renewable energy resources such as photo-

voltaic (PV) and wind energy are leading green resources
that are integrated with conventional power systems to meet
transform (TTT) [6], mathematical morphology (MM) [2],
empirical mode decomposition (EMD) [7], [9], sparse repre-
sentation [8] and the pattern recognition techniques including
power demand. High penetration of distributed generation artificial neural network (ANN), probabilistic neural network
(DG) systems causes serious issues related to power quality (PNN), fuzzy, and neuro-fuzzy, decision tree (DT), and support
(PQ) and stability problems [1]–[13]. The behavior of the wind vector machines (SVMs) [2]–[9], [16]–[27].
energy and PV systems is affected by many environmental In [2], classification of islanding and different PQ distur-
factors such as solar radiation intensity, shading of panel, bances is well studied using the ST, HST, TTT, and MM
the temperature of the cell, wind speed and tower shadow techniques, with SVM classifier. It is observed that the wavelet
effect [2]–[13]. The variations in load, switching phenomena, transform (WT) completely fails to detect the PQ events and
power electronic devices, charging of transformers, use of the ST method exhibits deterioration in localization to some
nonlinear loads along with the environmental factors cause extent in the presence of noise whereas the other techniques
PQ problems such as harmonics, interharmonics, sag, swell, such as TTT, HST and MM are capable of effectively detecting
interruptions, transients, notch and their combinations [10]– and localizing the islanding and other events. The standard
[13]. In order to ensure reliable, secure and quality supply of deviation, kurtosis, skewness and energy features are extracted
power, it has become utmost important for distribution system to study the effectiveness in classifying islanding and different
operator to continuously monitor these disturbances. Real time PQ disturbances.
In [3], authors studied detection and classification of PQ
Pankaj D. Achlerkar, S. R. Samantaray, and M. Sabarimalai Manikan- disturbances obtained with the modified Nordic 32-bus test
dan are with the School of Electrical Sciences, Indian Institute of Tech-
nology Bhubaneswar, Odisha-751013, India. E-mail: [email protected], systems with wind/PV penetration using WT/ST/HS-transform
sbh [email protected], [email protected] and DT/SVM classifier. In this study, ten statistical features

1949-3053 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2626469, IEEE
Transactions on Smart Grid
2

such as energy, standard deviation, mean, skewness, kurtosis are extracted from each mode for discriminating PQ distur-
of magnitude/phase contour are obtained from the HST matrix bances under noise-free and noisy environments. Finally, a
and optimal feature selection is done by genetic algorithm. decision-tree algorithm is implemented to classify the PQ
Results show that the HST with DT significantly provides signals into different PQ classes based on the predefined
better classification accuracy. thresholds as per the typical parameters described in IEEE
The FT is not effective tool for the analysis of non- PQ Standards 1159-2009 [34].
stationary PQ disturbances including notches, impulsive and The remainder of this paper is organized as follows: Section
oscillatory transients. The STFT has limited localization abil- II presents VMD approach with a suitable decomposition
ity for transient signals due to the fixed size of analysis parameters and a few test cases to demonstrate the effective-
window. The DWT has poor detection rates for the non- ness of the VMD technique. Section III presents the VMD
stationary PQ disturbances with harmonics and flickers due based method with four features and a decision tree algorithm.
to spectral leakage and noise sensitivity [2], [21]. DWT sets Section IV presents classification results using different types
hard-band limits for signal resolution in frequency domain of PQ disturbances in real-time generated and RTDS simulated
and its output depends on selection of mother wavelet and scenarios. Finally, conclusions are drawn in Section V.
number of levels of decomposition. It has been shown that
the DWT completely fails to detect disturbances under noisy II. VARIATIONAL M ODE D ECOMPOSITION
conditions [2], [3]. The study further shows that the ST The variational mode decomposition (VMD) is a recently
fails to detect and localize some transient disturbances under introduced adaptive and quasi-orthogonal signal decompo-
noisy conditions [3]. Further, it is not suitable for harmonic sition technique [33], which non-recursively decomposes a
analysis, particularly when high-order harmonics are present multicomponent signal x(t) into a finite number of compactly
[23]. The sparse representation based method achieved better band-limited IMFs (i.e. sub-signals or modes), xk (t) that have
classification rates with much more computational burden specific sparsity properties while reproducing the input signal.
[8]. There are few time domain methods that make use of A mode is modelled as a sinusoid with time-varying amplitude
three-phase data for disturbance detection at a bus, including and phase,
synchronous fundamental dq-frame method, synchronous har-
K K
monic dq-frame method, enhanced PLL method [28], method X X
x(t) = xk (t) = Ak (t) cos φk (t) (1)
based on instantaneous power theory and method based on
k=1 k=1
instantaneous symmetrical components [29]. Though these
are suitable for real-time implementation, they do not reveal Each mode is almost considered compact around a cor-
frequency information, they are highly noise sensitive and they responding center frequency. Major building blocks of non-
do not indicate phase in which disturbance has occurred. Other recursive VMD model are the Wiener filtering, the Hilbert
time domain methods, mainly used for detection of events (sag, transform and analytic signal and the frequency shifting
swell, interruption) are missing voltage technique, RMS/peak through harmonic mixing. VMD is based on constrained
value evaluation method [30], Teager energy operator [31] etc., variational optimization problem, solved by alternate direction
possess same drawbacks. method of multipliers (ADMM). The bandwidth of a mode is
N. E. Huang et.al. in [32] proposed a method called Hilbert- determined along with the decomposition in spectral domain,
Huang Transform that recursively breaks down a signal into with following steps [33]: (i) The real signal is converted to the
several unknown components called Intrinsic Mode Functions analytic signal using Hilbert transform [9] such that one-sided
(IMFs) in time domain, by empirical mode decomposition frequency spectrum of the signal is obtained; (ii) the mode
(EMD) and then estimates instantaneous amplitude, phase frequency spectrum is shifted to baseband regions by mixing
and frequency of individual modes by Hilbert transform. with an exponential with the estimated center frequency ωk ;
This method though free from any need of signal modelling and (iii) the bandwidth is computed by the H 1 Gaussian
and can catch non-stationary features, it has no supporting smoothness of the demodulated signal.
mathematical theory. Also, it is noise sensitive, it suffers from
problem of end effect and mode mixing. Several attempts have A. Variational Mode Decomposition Algorithm:
been made to mathematically develop the algorithm for mode Step 1: Initialize [xˆk 1 ], [ωk1 ], λ̂1 ; iteration count n = 1.
decomposition [15]. A more robust and completely variational Here, [xˆk 1 ] and [ωk1 ] is spectrum and center frequency of
approach is proposed by K. Dragomiretskiy and D. Zosso, k th mode respectively; λ̂1 is the Lagrangian multiplier for
named as variational mode decomposition (VMD) [33]. iteration 1.
This paper proposes application of VMD for detection,
localization and classification of single and mixed PQ dis- Step 2: For iteration n + 1, for all modes k = 1 to
turbances in grid-connected DG system. VMD is used for K, update the spectrum of each mode [xˆk ], ∀ ω ≥ 0
decomposing PQ signals into the fundamental, harmonics and n
x̂(ω) − i<k x̂n+1 (ω) − i>k x̂ni (ω) + λ̂ 2(ω)
P P
interharmonics, the non-stationary disturbances including the i
x̂n+1
k (ω) = 2
impulse and oscillatory transients, notches, spikes, and the 1 + 2α (ω − ωkn )
background noises. Then, four features such as mode central (2)
frequencies (MCFs), relative mode energy ratios (RMERs), Here, α is a bandwidth parameter, that determines range of
instantaneous amplitude (IA), number of zero crossings (ZCs) frequencies that are captured in various modes. This equation

1949-3053 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2626469, IEEE
Transactions on Smart Grid
3

is similar to that used for extraction of original signal from large, then most of the modes are discarded (if Lagrangian
noise using Wiener filtering, classically used for denoising. multiplier is set to zero). Underbinning can be identified
The inverse Fourier transform of spectrum gives analytic signal during post-processing, by checking the spectral overlap or
of modes, the modes in time domain, xk (t), are then obtained orthogonality between modes. However, underbinning may
by taking real part of time domain analytic signal. not produce perfect decomposition or it may loose important
Update the center frequency of each mode, as follows: mode informations, it is not recommended to occur. In case
R∞ of overbinning where specified number of modes are more
n+1 ω|x̂n+1
k (ω)|2 dω
ωk = R0 ∞ n+1 (3) than actual, if α is small, one or several extra modes will
0
|x̂k (ω)|2 dω largely consist of noise and thus have very low structure and
This equation puts the updated center frequency at the center broad spectral density. If α is large in this case, important part
of gravity of the power spectrum of corresponding mode. of signal/spectrum will be shared by two or more different
Update the Lagrangian multiplier as: modes and their center frequencies shall coincide (mode
! duplication). A component of particular frequency can be
n+1 n
X
n+1 obtained by adding duplicated modes. In case of detection of
λ̂ (ω) = λ̂ (ω) + τ x̂(ω) − x̂k (ω) (4)
any disturbance from signal, number of modes shall be kept
k
more than the number of harmonic order we want to extract
Here, τ is Lagrangian multiplier update parameter. to accommodate for extra noise, transient events and other
modelling errors.
Step 3: Increase iteration count and go back to step 2,
until convergence. The convergence criterion is,
2 C. Effectiveness of VMD for Analysis of PQ Disturbances
X x̂n+1

k − x̂nk 2 To demonstrate effectiveness of VMD approach in esti-
2 < (5)
k
kx̂nk k2 mating phasors and disturbance components of PQ signal,
synthetic test signals are generated in MATLAB.
where,  is an acceptable tolerance value.
1) Estimation of Harmonics and Interharmonics: In this
study, the synthetic PQ signal is computed as,
B. Selection of Decomposition Parameters 9
X
1) Lagrangian Multiplier (λ): The role of the Lagrangian x(t) = An cos (2πfn t + φn ) (6)
multiplier is to enforce the equality constraint of reconstruc- n=1
tion. If exact reconstruction is not required, particularly in with seven harmonics and two interharmonics as given in Table
presence of strong noise, Lagrangian multiplier can be set to I. The waveform is uniformly sampled at rate of 6.4 kHz. A
zero. It is done by simply choosing its update parameter τ = 0. time window of 200 ms is considered. Mirror extension is
2) Data Fidelity Constraint Parameter (α): It is also called used to remove end effect. Since it is known that waveform
bandwidth parameter. It represents range of center frequencies consists of nine components of very wide range of frequencies,
of modes in which the given signal is going to decompose. the values α = 250 (low value) and number of modes K = 9
In order to capture signals containing vary wide range of are chosen. Table I gives result of decomposition by VMD.
frequencies (e.g. harmonically distorted signals), value of α
can be kept smaller (of the range of few hundreds); while for TABLE I: Estimation of power quantities of PQ signal with
detection of frequency contents in a signal over smaller range harmonics and interharmonics
(e.g. detection of flicker, detection of fundamental frequency
Synthesized Signal [23] Result by VMD
oscillations), it can kept very high (of order of few tens of Freq. Amp. Phase Freq. Amp. Phase
thousands). fn (Hz) An (pu) φn (deg.) fn (Hz) An (pu) φn (deg.)
3) Number of Modes (K): Number of modes into which 50 1 -85.23 49.7 0.9876 -85.1357
150 0.77 89.41 149.6 0.7485 89.5131
signal has to be decomposed needs to be predefined in VMD. 250 0.56 -91.02 249.4 0.5408 -90.8553
Since, VMD concurrently updates all the modes, it avoids 350 0.32 88.52 349 0.3137 88.7335
mode mixing problem to great extent. 473 0.12 -92.03 472.7 0.0983 -94.7986
If in specific cases, the number of disturbance components 650 0.06 87.25 649.5 0.059 87.5469
775 0.06 -93.48 774.6 0.0515 -92.8434
present in a signal is known, that much number of modes can 850 0.03 85.63 848.1 0.021 86.5749
be specified to get correct component in each mode. Extra 1150 0.03 92.59 1150 0.0298 92.924
modes can be reserved in presence of noise so that noise
component will occur in those modes. However, in practice, The magnitude spectrum of decomposed modes (Fig. 2)
particularly in the application of detection of stochastic PQ dis- shows that each mode (represented by different color) is
turbances/harmonics, spectral content of voltage/current signal compact around its center frequency that matches with actual
is unknown. If number of modes specified does not match with frequency content of the signal. The amplitude and phase of
actual, resulting decomposition depends upon parameter α. modes, obtained from spectrum have reasonable accuracy.
Consider a case where specified number of modes is less It can be noted that, 200 ms time frame corresponds to
than actual; this is called as case of underbinning. If α is small, frequency resolution of 5 Hz. In this case, existing methods of
some of the modes are shared by neighbouring modes. If α is harmonic estimation such as discrete Fourier transform (DFT)

1949-3053 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2626469, IEEE
Transactions on Smart Grid
4

Mode 1
2 1
Signal

Flicker
Signal
0 0
−2 −1 0 (a)
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2
−1
1 0.5
Mode 2

Mode 3
0 0.05 0.1 0.15 0.2
0 0 Time (sec.)
−1 −0.5
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2
0.4

Magnitude
1 0.2
Mode 4

Mode 5
0 0 (b)
−1 −0.2 0.2
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2
0.1 0.1
Mode 6

0 Mode 7 0
0
20 30 40 50 60 70 80
−0.1
0 0.05 0.1 0.15 0.2 −0.1 Freq. (Hz)
0 0.05 0.1 0.15 0.2
0.05 0.05 Fig. 2: Illustrating effectiveness of VMD approach in estimat-
Mode 8

Mode 9

0 0
ing amplitudes, frequencies and phases of 50 Hz fundamental
−0.05 −0.05
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2 and flicker components
Time (sec.) Time (sec.)

1 TABLE II: Estimation of power quantities for flicker signal

Magnitude

vo fo φo , φf1 vf1 vf2 ff1 ff2

Parameter
0.5 (pu) (Hz) φf2 (deg.) (pu) (pu) (Hz) (Hz)
Synthesized
1 50 0 0.3 0.2 10 18
Signal
0 Result by
0 200 400 600 800 1000 1200
0.999 50 2.4 0.3 0.19 10 17.65
VMD
Freq. (Hz)

Fig. 1: Performance of VMD based method in estimating

Expanding Eqn. (7) and simplifying using trigonometric iden-
power quantities of harmonic and interharmonic contained PQ
tities, we get:
signal
1X
v(t) =vo cos(ωo t + φo ) − vf {cos[(ωo + ωf )t
2 (8)
f
give accurate estimation of components that are integer mul-
tiple of 5, whereas DFT fails to detect components with non- + (φo + φf )] + cos[(ωo − ωf )t + (φo − φf )]}
integer multiple frequencies of 5 Hz due to spectral leakage. Thus, for each flicker component of frequency ωf (< ωo ),
However, here, interharmonics of 473 Hz is estimated with there are two components in voltage signal (ωf ±ωo Hz). Also,
great accuracy by VMD (Table I). Since, center frequencies are frequencies of these components are closer to each other; thus
computed from Eqn. (3), it gives a numerical value irrespective modes from VMD can detect these components by keeping
of frequency resolution. Though the Eqn. (3) uses area under α very high. The values of amplitue, phase & frequency of
spectral curve, whose accuracy varies while implementation in flicker envelope (vf , ωf and φf ) as well as of fundamental
discrete domain, however it has little impact on computation component can be detected from VMD modes by comparison
of center frequencies. with Eqn. (8).
It is further emphasized that, VMD based method does It has been observed that flicker frequencies within smaller
not require a spectral parameter estimation method since the range of fundamental frequency can be measured by selecting
signal filtering is directly done in frequency-domain. However, the large value of data fidelity constraint (α = 10000).
while representing spectra by Fast Fourier Transform (FFT) The effectiveness of the VMD technique in measuring the
(Fig. 1), spectral leakage can be observed around 473 Hz. amplitudes, frequencies, phases of the flicker is shown in Fig.
Thus, though center frequency estimates it correctly, leakage 2. With number of modes K = 5, the estimated parameters
phenomenon in FFT do not allow to reflect it on spectrum. In are summarized in Table II. From the results, it is found that
such case, magnitude and phase computed from spectrum may the VMD technique can capture the flicker components and
be incorrect and it is better to compute them from actual time- accurately estimate the parameters of the fundamental and
domain mode signal. Though FFT is faster for estimation of flicker components. Since flicker frequencies are 10 and 18
stationary features, VMD has added advantage of estimating Hz, spectrum shows components at 50 ± 10 and 50 ± 18
time-varying spectra and detection of interharmonics. Hz. One of the component of flicker (18 Hz) is not integer
2) Estimation of Flicker Components: Flicker waveform multiple of 5 Hz, thus, it indicates leakage around 30 and 70
has amplitude envelope which has one or more low frequency Hz in spectrum. However, its frequency is correctly estimated
components. A general multicomponent flicker is mathemati- by VMD. However, if flicker component has lesser frequency
cally modelled as amplitude modulated sinusoid, such that it cannot complete at least one one cycle during
considered time window, the method fails to detect that
 
X
v(t) = vo − vf cos(ωf t + φf ) cos(ωo t + φo ) (7) component. Thus, flicker frequencies less than 5 Hz cannot
f be detected for time window of 200 ms.

1949-3053 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2626469, IEEE
Transactions on Smart Grid
5

2 Signal
Signal 0 (a) Variational Mode
Variational Mode Decomposition with
Decomposition with α = 250, K=10
−2
0 0.02 0.04 0.06 0.08 0.1 α = 10000, K=10

For each mode k=1 to K,

2 No
Yes
Extracted
LF signal

err (k)=No. of zero crossings calculated from

C.Freq.=60±2 Hz?
(b) Center freq. – actual no. of zero crossings
0
Fundamental
−2 Yes Yes
err (k) <=1 ?
No
0 0.02 0.04 0.06 0.08 0.1 35≤C.Freq.≤85 Hz?
xHF=xHF+kth mode
RMS envelope

1.5 No
of LF signal

Flicker
1 X: 0.02283 Noise/
X: 0.0727 (c) Residual k>K ?
Y: 1.246 Yes Center Freq.=60 ± 5 Hz ?
0.5 Y: 0.1076 Yes
IA envelope
0 thresholding No Max. of IA
0 0.02 0.04 0.06 0.08 0.1 Sag Yes
envelope of
No
0.1<IA<0.9 pu Center Freq.
0.5 Duration xHF<0.02 ?
Harmonics

Of =60×n ± 5 Hz ?
Event n is integer Noise Mean of
(d) 1.1<IA<1.8 pu Yes
0 Swell > No xHF<1e-5 ?
0.5 cycle No Impulse/ +ve
Of fundamental? IA<0.1 pu Yes
Interharmonics Spike Polarity of
−0.5
0 0.02 0.04 0.06 0.08 0.1 max|xHF| Oscillatory
Interruption Harmonics Notch Transient
-ve
0.5
Extracted
HF signal

0 (e)
−0.5 Fig. 4: Flow chart of proposed method for detection and
−1
0 0.02 0.04 0.06 0.08 0.1 classification of single and mixed PQ disturbances using the
0.8
two-stage VMD and decision tree-based algorithm
Magnitude

0.6 X: 60.01
Y: 0.6977 X:300
0.4 (f)
Y:0.05178

0.2
0
0 100 200 300 400 500 600 700 800
bances in grid-connected DG systems. Firstly, the proposed
Freq. (Hz) method implements two-stage VMD with different data fi-
Fig. 3: Effectiveness of VMD approach in detection of com- delity constraint (α) for decomposing the stationary and non-
bined power quality disturbances generated in real-time (signal stationary PQ disturbance signals into the fundamental com-
taken from IEEE-1159 PQD database [35]) ponent, the harmonics and interharmonics, the non-stationary
disturbances including the impulse and oscillatory transients,
notches, spikes, and the background noises. In the second
3) Estimation of Combined PQ Disturbances: The effec- step, the time-domain and frequency-domain features such as
tiveness of the VMD approach is evaluated for real-time mode central frequencies (MCFs), relative mode energy ratios
fault event generated voltage signal taken from the IEEE- (RMERs), mode instanta-neous amplitude (IA), number of
1159 Monitoring Electric Power Quality website [35]. The zero-crossings (ZC) are extracted from the each of the modes
decomposition results in Fig. 3 (b), (c), (d) and (e) demonstrate for discriminating the fundamental and PQ disturbance signals
the ability of VMD to discriminate time-varying fundamental from noisy (or residual) modes under noise-free and noise
(here, 60 Hz) (named as LF signal), harmonics (300 Hz) and conditions. A decision-tree algorithm is implemented to detect
oscillatory transient (named as HF signal) components. It also and classify the input PQ signals stationary and non-stationary
indicates applicability of VMD for non-stationary signals (e.g. disturbance classes based on the four features with predefined
transients), having time-varying amplitudes (also the cases of feature thresholds as per the typical parameters defined in
sag, swell and interruption), which can be captured in a mode IEEE PQ Standards 1159-2009 [34].
by its root mean square (RMS) envelope. The spectra of the
fundamental and harmonics can be seen in Fig. 3 (f).
A. Feature Extraction by VMD
III. VMD AND D ECISION -T REE BASED D ISTURBANCE This subsection presents four features that are extracted
D ETECTION AND C LASSIFICATION A LGORITHM from the mode spectra and the time-domain modes for detec-
The preliminary decomposition results emphasize that the tion, localization and classification of various PQ disturbances.
number of modes must be more than the number of harmonic Acquired disturbance signal is decomposed into K = 10
terms to adequately capture the non-stationary PQ disturbances modes. Decomposed modes represent separated stationary
including impulsive and oscillatory transients, spikes, notches (DC, fundamental, harmonics, interharmonics etc.), quasi-
and noises. It is further noted that the flicker and time- stationary and non-stationary (sag, swell, interruption, tran-
varying envelopes can be adequately captured by choosing sient events) and noise parts. Parameter α is chosen as 250
larger value of α. Based on studies on spectral ranges of both for detection of all power quality (PQ) disturbances except
stationary and non-stationary PQ disturbances, a two-stage α = 10000 is chosen for flicker detection. Since flicker
VMD technique with two sets of decomposition parameters, signal is viewed as amplitude modulated sinusoid that has side
α and K, for robust detection and measurement of power lobes around fundamental frequency in magnitude spectrum
quantities (amplitude, frequency and phase) and for classi- (within ±25 Hz), high value of α gives decomposition within
fying sinusoidal disturbances and PQ indices (instantaneous narrow spectral range. The values of parameters α and K
amplitude and duration) for non-stationary sinusoidal and non- are empirically chosen. To identify features represented by a
sinusoidal disturbances is presented. mode, following indices are used:
Fig. 4 shows a simplified two-stage VMD and decision a) Relative Mode Energy Ratio (RMER): Let the
tree based detection and classification method for PQ distur- discrete-time PQ sequence x[n] with N samples is decomposed

1949-3053 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2626469, IEEE
Transactions on Smart Grid
6

into K modes. The relative mode energy ratio (RMER) for the TABLE III: Specifications of PQ disturbances as described in
k th mode is computed as, IEEE PQ standards 1159-2009 [34]
P 2 Disturbance Type Duration Amplitude Spectral Content
Ek n xk [n] Flicker > 0.5 cycle 0.9 - 1.1 pu < 25 Hz
RMERk = =P P 2 (9)
ET k n xk [n] Harmonics Steady state 0 - 0.2 pu 0 - 5 kHz
PN −1 Interharmonics Steady state 0 - 0.02 pu 0 - 6 kHz
where, Ek = n=0 x2k [n] is the energy of k th time domain Transients ns - ms 0 - 8 pu 5 kHz - 5 MHz
PK Sag > 0.5 cycle 0.1 - 0.9 pu -
mode signal and ET = k=1 Ek is total energy of all modes. Swell > 0.5 cycle 1.1 - 1.8 pu -
It is obvious that original signal’s total energy is distributed Interruption > 0.5 cycle < 0.1 pu -
among different modes; the contribution of each mode to Notch / Spike Steady state - -
signal is understood from RMER. Only dominant modes are Noise Steady state 0 - 0.01 pu Broadband
selected for further analysis, while modes having RMERs
below threshold are discarded as noise. Since least dominant
2. The high-frequency (HF) signal xHF [n] composed of the
mode do not contribute significant energy to the signal, they
impulsive and oscillatory transients, spikes, notches and
do not cause considerable disturbance and do not represent
background noises.
important features.
b) Mode’s Center Frequency (ωk ): Since the VMD is Apart from them, the remaining mode signals contain decom-
performed in frequency domain, it results in the center fre- posed harmonics and interharmonic components. The discrimi-
quency for each of the modes without additional spectral esti- nation among harmonics is simply done by center frequencies
mation techniques. Information of ωk is used to discriminate and events like sag, swell and interruption are detected by
signal into fundamental, harmonics, interharmonics and DC thresolding of IA envelope of LF signal as per IEEE Std. 1159-
offset parts. 2009 [34] (Table III). The HF signal is used to further classify
c) Number of zero-crossings: The number of zero- non-stationary transient events based on IA envelope, mean
crossings of k th mode xk [n] is computed as, value of signal, duration and polarity of maximum amplitude.
X The HF signal is thus classified into impulse, notch, oscillatory
ZCk = |sgn (xk [n]) − sgn (xk [n − 1]) | (10) transient and residual noise.
n

It is used to discriminate non-stationary features and noise B. Decision Tree Based Classification
from stationary ones. If for a particular mode, number of zero The proposed disturbance classification algorithm is de-
crossings calculated from mode’s center frequency [Eqn. (3)] picted in the flow chart shown in Fig. 4 (Algorithm is
matches with actual number of zero crossings [Eqn. (10)] (i.e. shown considering fundamental signal frequency to be 60
if ωk ×T = ZCk /2, where T is considered time window), then Hz). It is evident from flow chart that, proposed method can
signal contains stationary features like fundamental, harmonics independently detect presence of single as well as combined
or interharmonics with short/long duration variations (e.g. disturbances and is able to separate noise from disturbance
sag, swell, interruption). Otherwise, mode represents non- features. Here, though a separate detection scheme is not
stationary transient events like impulse, notch, oscillatory shown for DC offset, it can be detected by mode having
transient or noise. It is because of fact that non-stationary center frequency closer to zero Hz, otherwise a separate mode
transients appear for smaller duration within the time window. can be reserved for DC offset detection if there is possibility
d) Instantaneous Amplitude (IA): The instantaneous am- of its occurrence. Further, detection of stationary features
plitude envelope of mode signal is used as a feature for i.e. fundamental frequency and harmonics is based on center
discriminating the short and long duration variations (sags, frequency lying in a soft tolerance band of ±5 Hz. This
swells and interruptions). The instantaneous amplitude (IA) of band takes care of fundamental frequency variation events and
k th mode xk [n] is computed by a moving window of length errors in mode decomposition (if any), particularly with strong
Nk samples as, noise levels. The band is narrowed down to ±2 Hz for flicker
v detection, since flicker frequency can lie in that range.
u 1 m+N Xk −1
u
IAk [m] = t x2k [n] (11)
Nk n=m IV. R ESULTS AND D ISCUSSION
This section elaborates usefulness of proposed method for
where, m = 1, 2, . . . , length (xk [n]) − Nk + 1 and Nk = ωfsk analysing typical signal features that are commonly observed
is the number of samples per fundamental cycle period as in real-time distribution system, with considerable penetration
computed from center frequency ωk . The sampling frequency of distributed generation. For validation purpose, real power
is denoted as fs . quality distortion event signals are taken from IEEE-1159
Based on the aforementioned features, two candidate signals Monitoring Electric Power Quality Database [35] as well as
are constructed: from the GIM-PQE database. [36]
1. The low-frequency (LF) signal xLF [n] containing fun- Most of the renewable energy based DGs implement power
damental component and its short- and long-duration electronic converters for grid-connection. Non-linear loads
variations (sags, swells and momentary interruptions, may such as adjustable speed drives also add power electronics
contain flicker) into conventional distribution system. Impulsive transients are

1949-3053 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2626469, IEEE
Transactions on Smart Grid
7

200
2

Signal
100
Signal
0 0
−100
−2 1.4 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07
0 0.02 0.04 0.06 0.08 0.1

DC offset
40
1.2 20
LF signal

2
Extracted

0
0 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07
1

Magnitude
−2 50

signal
0 0.02 0.04 0.06 0.08 0.1 0.02

LF
0
0.8
0.01 −50
Extracted
HF signal

0.5 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07

0.6 0
0

IA envelope Impulse IA envelope

of impulse (HF signal) of LF signal
50
−0.5 1000 2000 3000
0.4
0 0.02 0.04 0.06 0.08 0.1
0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07
of HF signal
IA envelope

0.5 0.2
40
0 20
0 200 400 600 800 1000 1200 0
0 −20
0 0.02 0.04 0.06 0.08 0.1 Freq. (Hz) 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07
Time (sec.)
40
20
0
Fig. 5: Detection, estimation and localization of notches/spikes 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07
Time (sec.)
by proposed method (real-time event generated signal taken
from [35])
Fig. 6: Detection of simultaneous presence of DC-offset, fun-
damental and impulsive transient by proposed method (real-
typical events that occur due to delay in switching of power time event generated signal taken from [36])
electronic devices, or misfired pulses; also due to lightening or
instant load switchings. A voltage waveform distorted due to
impulsive transients and its decomposition results are shown
in Fig. 5. The fundamental component is fully decomposed as
LF signal and impulse component (i.e. HF signal) is separately
extracted. Spectral content shows broad spectrum decomposed
into several modes however with compact bandwidths. The
noise also constitute such a broad-spectrum, however presence
of impulse is detected based on thresholding of IA envelope of
HF signal. HF signal preserves polarity of impulses, thus they
can be classified into spikes and notches (+ve and -ve polarity
respectively). Also, accurate time localization of impulses
can be obtained based on instances at which IA envelope
crosses threshold value. One may argue that in presence of
strong noise, HF signal shall contain both noise and impulses.
However, it should be noted that, decomposition is done in
enough number of modes (K = 10) and noise is separated Fig. 7: Real-Time Digital Simulator (RTDS) Set-up for Per-
in separate mode(s). Further, RMER feature discards noisy formance Validation of Proposed Method
modes and only remainder of modes are processed.
Another real-time event signal is shown in Fig. 6. Signal
individual racks where several coupled digital signal proces-
consists of an interruption followed by a DC offset and spike
sors (DSPs) perform fast parallel processing [37]. A dedicated
along with fundamental. Decomposition with α = 250, K =
software namely RSCAD has an user-friendly graphical user
10 and classification by decision-tree results in independent
interface (GUI), where a microgrid test system, shown in
detection of each type of disturbance component. Stationary
Fig. 8 is simulated in DRAFT module. Various disturbances
part (i.e. fundamental component) is effectively separated out
are created and obtained voltage and/or current signals are
from non-stationary part (i.e. transient). Also time localization
processed with proposed algorithm in MATLAB running on
of events is possible by thresholding IA envelopes.
personal computer. The signals are sampled uniformly at 6400
Hz with 200 ms time frame. The signals are further added with
A. Validation for Microgrid Test System on Real-Time Digital white Gaussian noise of varying levels in order to validate
Simulator (RTDS) performance of proposed method.
Real time digital simulator (RTDS) is an ideal simulation The microgrid consists of grid-connected a photovoltaic
tool for study of power system, particularly detailed analysis of resource and three doubly-fed induction generators based wind
electromagnetic transient phenomena that occur in real time. turbine. Other parameters are detailed in Table IV.
The randomness in signal as well as detailed transient features A transient event generated in the test system of Fig. 8 is
are effectively and accurately captured by RTDS simulation. shown in Fig. 9. Capacitor bank used at bus-10 for power
RTDS models are based on Dommel’s model of power system factor improvement, is switched at 0.04 sec. An impulsive
components. RTDS has specialized hardware consisting of transient followed by oscillatory response is observed in

1949-3053 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2626469, IEEE
Transactions on Smart Grid
8

PQ signal
1 2
0 0
−1
−2 −2
0 0.02 0.04 0.06 0.08 0.02 0.04 0.06 0.08
2

LF signal
1
0
0
−1 −2
0.02 0.04 0.06 0.08
0 0.02 0.04 0.06 0.08
1

HF signal
1 0.5
0
0
−0.5
−1 0.02 0.04 0.06 0.08
0 0.02 0.04 0.06 0.08
0.5

of HF signal
IA envelope
0.5

0
0 0.02 0.04 0.06 0.08
0 0.02 0.04 0.06 0.08
Time (sec.)
Time (sec.)

(a) Oscillatory transient (b) Impulsive transient

Fig. 9: Detection and classification of impulsive and oscillatory

transient in presence of noise (RTDS signal)

PQ signal
1
0
−1
0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
1

signal
Fig. 8: Microgrid test system consisting of grid-connected PV

LF
0
−1
and wind resources, simulated on RTDS platform 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16

IA envelope
of LF signal
1
TABLE IV: Microgrid Components and their Specifications 0.8
Component Specifications 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
Utility Rated short-circuit MVA=1000, f=60 Hz, 0.2
of HF signal signal
HF

Grid Rated kV=400, Base Voltage=400 kV 0

Rated MVA=20, f=60 Hz, Rated kV=400/33, −0.2
Transformer
R1 & R2 = 0.00375 pu, X1 & X2 = 0.1 pu, 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
IA envelope

T1 0.1
Rm =500 pu, Xm = 500 pu
Transformer Rated MVA=20, f=60 Hz, Rated kV=0.575/33, 0.05
T2, T3, R1 & R2 = 0.00375 pu, X1 & X2 = 0.1 pu, 0
0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
T4, T5 Rm =500 pu, Xm = 500 pu Time (sec.)
DFIG-1, Rated MW=1.5, f=60 Hz, pf=0.9,
DFIG-2, Rated kV=575 V, Rs =0.023 pu,L1s =0.18 pu, Fig. 10: Sag+Impulse+noise signal obtained from sudden
DFIG-3 Rr0 =0.016 pu, L01r =0.16 pu, Lm =2.9 pu load variation in microgrid test system and its detection by
Rated MW=1, Rated Voltage=575 V, f=60 Hz,
Photovoltaic
Vdc =2500 V, No. of cells in series=30, proposed method (RTDS Signal)
Generator
No. of modules in parallel=675
π-section, 5 km each, Vbase =33 kV,
Distribution R0 =0.1153 Ω/km, L0 =3.32 mH/km,
Lines C0 =11.33 µF/km, R1 =0.413 Ω/km, signal. Further, HF signal consists of noise, as well as an
L1 =3.32 mH/km, C1 =5.01 µF/km impulsive transient at the time of load addition is detected.
Loads Total load in the system is 8 MW+j2.72 MVAr
Capacitor (C) QC = 200 kVAR, used for pf improvement at B-10 This example illustrates how stationary, non-stationary features
can be detected and estimated based on proposed VMD based
feature extraction and DT based classification approach.
nearby buses, whereas impulse is observed in remote buses. The islanding event in a microgrid system is common
The two types of transient signals are added with random phenomenon, where main grid disconnects from rest of the
noise and decomposed by VMD. Fig. 9 shows separated distribution system and only DGs have to supply the demand at
fundamental component (as LF signal) and noise plus transient that instant. This results in over/under-loading of DGs thereby
part (as HF signal). Remainder of modes found to contain causing voltage and frequency deviations and waveform distor-
noise (as classified by RMER criterion and DT algorithm) tion. One such event is generated and its decomposition results
and are discarded. The event is time localized by IA envelope are shown in Fig. 11. IA of LF signal indicates deviations
thresholding. This example proves detection capability of in the signal. In spite of fundamental frequency deviations
proposed approach under noisy conditions. during islanding, VMD technique is able to decompose signal
Another example of power quality signal is shown in Fig. into LF fundamental part, harmonic parts (shown in spectrum
10. Here, a sudden load is added into system and removed by different colors) and residual noise. It can thus be noted
at times 0.04 sec. and 0.13 sec. respectively. As evident that, sampling synchronized to fundamental frequency is not
from VMD decomposition results, fundamental component essential for mode decomposition, since small deviation from
(of voltage signal) seems to have a drop in amplitude due spectrum do not cause considerable changes in Eqns. (2) and
to heavy loading, the duration of sag can be observed from (3). Further, DT based classification allows ±5 Hz deviations.
crossing points of threshold (0.9 pu) with IA envelope of LF

1949-3053 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2626469, IEEE
Transactions on Smart Grid
9

TABLE VI: Performance of the classification methods for

PQ signal
1
0 single and mixed PQ disturbances
−1
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
PQ Disturbance ADALINE Fuzzy Dynamics Wavelet Proposed
Type & FFNN [24] & ST [25] & ST [26] [27] Method
LF signal

1 Normal 100 - 100 - 100

0 Osc. Transient 98 94 99 98.67 99.16
−1 Spike 97 100 97 - 99
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 Notch 97 96 98 97.33 99
Flicker 94 - 96 98.67 100
IA envelope

1
of LF signal

Interruption 100 96.99 96 98 99.35

0.8
DC offset - - - - 100
0.6 Sag 100 97.33 99 97.33 100
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 Swell 100 98.66 98 98.67 100
0.2 Harmonics 98 100 99 99.33 100
Reconstructed Residual
noise

0
Sag+transient - 99.33 - 98.18 100
Swell+transient - 100 - 98.18 100
−0.2 Sag+harmonics 98 - 97 98.18 97.4
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
Swell+harmonics 97 96 98 98.18 98.62
1
signal

0
−1
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
Time (sec.) capability since the VMD approach can effectively separate
0.8
the sinusoidal and non-stationary sinusoidal disturbances from
the short-duration non-sinusoidal PQ disturbances as revealed
Magnitude

0.6
0.4 in Figs. 1, 2, 3, 5, 6, 9, 10 and 11.
0.2 it is observed that the average coding time of the proposed
0
0 100 200 300 400 500 600 method using the VMD and decision tree with the four features
Freq.(Hz)
(coded in MATLAB) is 190 ms. In future directions, we extend
Fig. 11: PQ distoriton signal generated due to islanding event a further in-depth study on reducing computational load by
in microgrid test system on RTDS platform implementing the proposed method on real-time embedded
computing platforms, such as DSP and FPGA to further
TABLE V: Classification accuracy of proposed method under validate the feasibility of the proposed method for detection
different noise levels for various PQ signals and classification of single and mixed PQ disturbances in grid-
connected DG systems.
PQ Disturbance Signals Clean 45 dB 40 dB 35 dB 30 dB
Fault (Sag) 100 100 100 99.8 98.2 In the present study, the decomposition is applied to a
Load Reduction/ Capacitive Load (Swell) 100 100 100 98.4 97.6 windowed signal, in block processing manner, where a block
Fault (Interruption) 100 100 99 99 97.2
Harmonics/Interharmonics 100 100 99.4 99.4 98.5 of 10/12 cycles for fundamental frequency 50/60 Hz system
Dynamic Load Variations (Flicker) 99.6 99.6 98.2 97.4 95.6
Fault (DC Offset) 100 100 99.4 98 97.4
(of duration 0.2 sec.) is considered. There can be some non-
Oscillatory Transient 100 100 100 99.5 98.2 stationary events lasting for longer durations, which are not
Sag+Harmonics 99 99 97.28 96 94.1
Swell+Harmonics 99.5 98.4 97.2 95 93.8 captured within this window, those might be considered as
quasi-stationary events by the method. The information by
few past windows needs to be recorded for monitoring of
B. Classification Accuracy of Proposed Method such events. For implementation purpose, proposed method
can be used with moving window processing where window
For different types of simulated PQ disturbances, the classi-
of fixed width takes new samples each time leaving equal
fication rates of the proposed method is presented in Table V
number of samples behind. In such case, initialization of
under both noisy and noise-free environments. The characteris-
center frequencies and spectrum can be done based on those
tics of simulated events and number of events are randomized,
quantities of previous window so as to reduce number of
total number of cases considered is a random number between
iterations required for VMD. The algorithm of VMD involves
100 to 150 in each case. The robustness and accuracy of the
calculation of FFT and inverse FFT of signals, which can be
proposed method is verified with different signal to noise ratio
done by recursive approach from one window to other while
(SNR) values ranging from 45 to 30 dB. Results demonstrate
implementing in real time.
that the proposed method based on the VMD and DT can
Though in proposed method, number of modes (K) is
achieve the better detection and classification rates for single
empirically chosen to be 10, the number will depend upon
and mixed PQ disturbances. The classification rate decreases
how strong or weak the grid is and how distorted the signals
as the noise level increases, but still above 93% for SNR of
can be. With the knowledge of anticipated amount and number
30 dB. It is due to close relationship of VMD’s mode update
harmonics and other disturbances, proposed method can be
equation [Eqn. (2)] with Wiener filtering.
tuned for accurate signal decomposition into modes represent-
For both single and mixed PQ disturbances, the classifica-
ing principal features. Modes can be increased in presence of
tion rates of the proposed method are compared in Table VI
considerable harmonics or noise.
with four methods including, FFNN and ADALINE [24], ST
and fuzzy rule [25], ST and Dynamics [26], and wavelet net-
works [27]. Evaluation results show that the proposed method V. C ONCLUSIONS
has comparatively better performance over the other existing In this paper, we present a new automated method using
methods. The proposed method shows better classification variational mode decomposition and decision tree algorithm

1949-3053 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2626469, IEEE
Transactions on Smart Grid
10

for detection and classification of single and mixed PQ dis- [16] D. Granados-Lieberman, R. J. Romero-Troncoso, R. A. Osornio-Rios, A.
turbances in grid-connected DG system. The accuracy and Garcia-Perez, and E. Cabal-Yepez, ”Techniques and methodologies for
power quality analysis and disturbances classification in power systems:
robustness of the method is tested and compared with four A review”, IET Generat., Transmiss. Distrib., vol. 5, no. 4, pp. 519-529,
methods under various system operating scenarios. Validation Apr. 2011.
with several synthetic and real-time generated signals, and [17] O. P. Mahela, A. G. Shaik, and N. Gupta, ”A critical review of detection
and classification of power quality events”, Renew. Sustain. Energy Rev.,
simulations on microgrid test bed on RTDS platform shows the vol. 41, pp. 495-505, 2015.
potential ability of VMD technique to accurately estimate am- [18] S. Khokhar, A. A. B. M. Zin, A. S. B. Mokhtar, and M. Pesaran, ”A
plitudes, frequencies and phases of the fundamental, harmonic comprehensive overview on signal processing and artificial intelligence
techniques applications in classification of power quality disturbances”,
and interharmonic, and flicker components under mixture of Renew. Sustain. Energy Rev., vol. 51, pp. 1650-1663, 2015.
other non-sinusoidal PQ disturbances, in both noise-free and [19] Y. Önal, D. Gökhan Ece, and Ö Nezih Gerek, ”HilbertHuang Trans-
noisy scenarios. Various studies show that the proposed VMD form Based Approach for Measurement of Voltage Flicker Magnitude
and Frequency”, Electr. Pow. Compo. Sys., vol. 43, no. 2, pp. 167-176,
with decision-tree based algorithm is an effective and efficient 2015
method for detection and classification of single and mixed [20] C.-I. Chen, Y.-C. Chen, Y.-R. Chang, and Y.-D. Lee, ”An accurate
PQ disturbances under noise-free and noisy environments. solution procedure for calculation of voltage flicker components”, IEEE
Trans. Ind. Electron., vol. 61, no. 5, pp. 2370-2377, 2014.
[21] S. Santoso, W. M. Grady, E. J. Powers, J. Lamoree, and S. C. Bhatt,
”Characterization of distribution power quality events with Fourier and
R EFERENCES wavelet transforms”, IEEE Trans. Power Del., vol. 15, no. 1, pp. 247-
254, 2000.
[1] M. Mesbah, P. S. Moses, S. M. Islam, M. A. S. Masoum, ”Digital [22] C. Xue, L. Hui-jin, Z. Quan-ming, Y. Li-ming, and L. Qing-fen,
Implementation of a Fault Emulator for Transient Study of Power ”Power quality disturbances detection and location using mathematical
Transformers Used in Grid Connection of Wind Farms”, IEEE Trans. morphology and complex wavelet transformation”, in 3rd IEEE Conf.
Sustain. Energy, vol.5, no.2, pp.646-654, Apr. 2014. Ind. Electron. Appl. (ICIEA), pp. 2263-2268, Jun. 2008.
[2] S. R. Mohanty, N. Kishor, P. K. Ray, J. P. S. Catalao, ”Comparative [23] N. C. Tse, J. Y. Chan, W.-H. Lau, and L. L. Lai, ”Hybrid Wavelet and
Study of Advanced Signal Processing Techniques for Islanding Detec- Hilbert Transform with Frequency-Shifting Decomposition for Power
tion in a Hybrid Distributed Generation System”, IEEE Trans. Sustain. Quality Analysis”, IEEE Trans. Instrum. Meas., vol. 61, no. 12, pp.
Energy , vol.6, no.1, pp.122-131, Jan. 2015. 3225-3233, 2012.
[3] P. K. Ray, S. R. Mohanty, N. Kishor, J. P. S. Catalao, ”Optimal Feature [24] M. V. Rodriguez, R. J. R. Troncoso, R. A. O. Rios and A. G. Perez,
and Decision Tree-Based Classification of Power Quality Disturbances ”Detection and classification of single and combined power quality
in Distributed Generation Systems”, IEEE Trans. Sustain. Energy, vol.5, disturbances using neural networks”, IEEE Trans. Ind. Electron., vol.
no.1, pp.200-208, Jan. 2014. 61, no. 5, pp. 2473-2482, May 2014.
[4] P. K. Ray, S. R. Mohanty, and N. Kishor, ”Classification of power [25] P. K. Dash and M. Biswal, ”Measurement and classification of simul-
quality disturbances due to environmental characteristics in distributed taneous power signals patterns with an S-transform variant and fuzzy
generation system”, IEEE Trans. Sustain. Energy, vol. 4, no. 2, pp. 302- decision tree”, IEEE Trans. Ind. Inf., vol. 9, no. 4, pp. 1819-1827, 2013.
313, 2013. [26] S. He, K. Li and M. Zhang, ”A real-time power quality disturbances
classification using hybrid method based on S-transform and dynamics”,
[5] P. K. Ray, N. Kishor and S. R. Mohanty, ”Islanding and power quailty
IEEE Trans. Instrum. Meas., vol. 62, no. 9, pp. 2465-2475, Dec. 2013.
disturbance detection in grid-connected hybrid power system using
[27] M. A. S. Masoum, S. Jamali and N. Ghaffarzadeh, ”Detection and clas-
wavelet and S-transform”, IEEE Trans. Smart Grid, vol. 3, no. 3, pp.
sification of power quality disturbances using discrete wavelet transform
1082-1094, 2012.
and wavelet network”, Sci. Meas. Technol., vol. 62, no. 3, pp. 531-544,
[6] S. Suja and J. Jerome, ”Pattern recognition of power signal disturbances
2014.
using S-transform and TT transorm”, Elect. Power Energy Syst., vol. 32,
[28] A. Safaee, D. Yazdani, A. Bakhshai, and P. Jain, ”Three-phase harmonic
no. 1, pp. 37-53, 2010.
detection methods for grid-connected converters,” in 2011 Twenty-Sixth
[7] B. Biswal, M. Biswal, S. Mishra and R. Jala, ”Automatic classification Annual IEEE Applied Power Electronics Conference and Exposition
of power quality events using balanced neural tree”, IEEE Trans. Ind. (APEC), 2011.
Electron., vol. 61, no. 1, pp. 521-530, Jan. 2014. [29] R. Kumar, B. Singh and D. T. Shahani, ”Symmetrical components
[8] M. Sabarimalai Manikandan, S. Samantaray and Innocent Kamwa, based technique for power quality event detection and classification”,
”Detection and classification of power quality disturbances using sparse In IEEE International Conference on Power Electronics, Drives and
signal decomposition on hybrid dictionaries”, IEEE Trans. Instrum. Energy Systems (PEDES), 2014.
Meas., vol. 64, no. 1, pp. 27-38, Jan. 2015. [30] M. Mansor and N. A. Rahim, ”Voltage sag detection-A survey”, in IEEE
[9] S. Shukla, S. Mishra, B. Singh, ”Empirical-mode decompositionn with Int. Conf. Technical Postgraduates (TECH-POS), pp. 1-6, 2009.
Hilbert transform for power-quality assessment”, IEEE Trans. Power [31] D. Kaiser, J. F. Kaiser et al., ”Estimation of power systems amplitudes,
Deivery, vol. 24, no. 4, pp.2159-2165, Oct. 2009. frequencies, and phase characteristics using energy operators”, in Energy
[10] G. Chicco, J. Schlabbach and F. Spertino, ”Experimental assessment of Conversion Congress and Exposition (ECCE), IEEE, pp. 930-937 2012.
the waveform distortion in grid connected photovoltaic installations”, [32] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q.
Solar Energy, vol. 83, no, 7, pp. 1026-1039, Jul. 2009. Zheng, N.-C. Yen, C. C. Tung, and H. H. Liu, ”The empirical mode
[11] J Bialasiewicz and E. Muljadi, ”The wind farm aggregation impact on decomposition and the Hilbert spectrum for nonlinear and non-stationary
power quality”, in Proc. IEEE Annual Conf. Ind. Electron., pp. 4195- time series analysis”, in Proceedings of the Royal Society of London A:
4200, 2006. Mathematical, Physical and Engineering Sciences, vol. 454, no. 1971,
[12] IOAN Şerban and Corneliu Marinescu, ”Power quality issues in a stand- pp. 903-995, 1998.
alone microgrid based on renewable energy”, Rev. Roum. Sci. Techn.– [33] K. Dragomiretskiy and D. Zosso, ”Variational mode decomposition”,
Électrotechn. et Énerg, vol. 53, no. 3, pp. 285-293, 2008. IEEE Trans. Signal Process., vol. 62, no. 3, pp. 531-544, 2014.
[13] Z. Chen and E. Spooner, ”Grid power quality with variable speed wind [34] IEEE Recommended Practice for Monitoring Electric Power Quality,
turbines”, IEEE Trans. Energy Conversion, vol. 16, no. 2, pp. 148-154, IEEE Standard 1159-2009, 2009.
2001. [35] IEEE 1159.2: Power Quality Event Characteriza-
[14] B. Byman, T. Yarborough, R. Schnorr Von Carolsfeld, and J. Van Gorp, tion, IEEE power quality standards, Available at:
”Using distributed power quality monitoring for better electrical system http://www.grouper.ieee.org/groups/1159/2/testwave.html
management”, IEEE Trans. Ind. Appl., vol. 36, no. 5, pp. 1481-1485, [36] Power Quality Events Data Collection, Instituto de Telecomunicaes
Sep. 2000. 2008. Available at: http://www.gim.lx.it.pt/fct57708/data/
[15] Marcelo AA Lima, et. al. ”A novel method for power quality multiple [37] Kuffel, R., J. Giesbrecht, T. Maguire, R. P. Wierckx, and P. McLaren,
disturbance decomposition based on Independent Component Analysis”, ”RTDS-A fully digital power system simulator operating in real time”,
International Journal of Electrical Power & Energy Systems vol. 42, no. In WESCANEX, Communications, Power, and Computing Conference
1, pp. 593-604, 2012. Proceedings, IEEE, vol. 2, pp. 300-305, 1995.

1949-3053 (c) 2016 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2626469, IEEE
Transactions on Smart Grid
11

Pankaj D. Achlerkar received a B.E. in Electrical, M. Sabarimalai Manikandan (M’13) received the
Electronics and Power from Department of Electrical B.E. degree in electronic and communication en-
Engineering, Government College of Engineering, gineering from Bharathiar University, Coimbatore,
Aurangabad, India, in 2013 and M.Tech. in Power India, in 2000, the M.E. degree in microwave and
System Engineering from School of Electrical Sci- optical engineering from Madurai Kamaraj Univer-
ences, Indian Institute of Technology Bhubaneswar, sity, Madurai, India, in 2002, and the Ph.D. degree
India, in 2016. in cardiovascular signal processing from the Depart-
His current research areas of interest include grid ment of Electronics and Communication engineer-
integration impacts of distributed generators, power ing, IIT Guwahati, Guwahati, India, in 2009.
quality estimation and analysis, and microgrids. He served as an Assistant Professor with Amrita
Vishwa Vidyapeetham University, Ettimadai, India.
He was a Chief Engineer with the Advanced Technology Group, Samsung
India Electronic Pvt. Ltd., Noida, India. He is currently an Assistant Professor
with the School of Electrical Sciences, IIT Bhubaneswar, Bhubaneswar, India.
S. R. Samantaray (M’08-SM’10) received a He holds six filed patents in the field of cardiac health monitoring, speech
B.Tech. in electrical engineering from UCE Burla, processing, data security, and video summarization systems. His current
India, in 1999 and a Ph.D. in power system en- research interests include signal detection and analysis, biomedical signal and
gineering from the Department of Electronics and image processing, pattern recognition, multimedia, affective computing, and
Communication Engineering, National Institute of wireless health monitoring systems.
Technology, Rourkela, India, in 2007. Dr. Manikandan was a recipient of the 2012 Outstanding Performance
Dr. Samantaray holds the position of Associate Award during his tenure at Samsung India Electronic Pvt. Ltd.
Professor in the School of Electrical Sciences, In-
dian Institute of Technology Bhubaneswar, India. He
visited the Department of Electrical and Computer
Engineering, McGill University, Montreal, Canada
as a Post-Doctoral Research Fellow and Visiting Professor. His major research
interests include intelligent protection for transmission systems (including
FACTs) and microgrid protection with distributed generation and dynamic
security assessment in large power networks.
He is the recipient of the 2007 Orissa Bigyan Academy Young Scientists
Award, the 2008 Indian National Academy of Engineering Best Ph.D. Thesis
Award, the 2009 Institute of Engineers (India) Young Engineers Award, the
2010 Samanta Chandra Sekhar Award and the 2012 IEEE PES Technical
Committee Prize Paper Award, Excellence in Reviewing Award-2013, Elsevier
Scinece and NASI-SCOPUS Young Scientist Award 2015.
Dr. Samantaray is an Editor of IEEE Transactions on Smart Grid, Associate
Editor of IET, Generation, Transmission & Distribution; Canadian Journal
of Electrical and Computer Engineering and Electric Power Components and
Systems. He is the member of National Academy of Sciences, India.

1949-3053 (c) 2016 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.