detection and classification of power quality disturbances using sparse signal decomposition on...

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT 1

Detection and Classification of Power QualityDisturbances Using Sparse Signal Decomposition

on Hybrid DictionariesM. Sabarimalai Manikandan, Member, IEEE, S. R. Samantaray, Senior Member, IEEE,

and Innocent Kamwa, Fellow, IEEE

Abstract— Several methods have been proposed for detectionand classification of power quality (PQ) disturbances usingwavelet, Hilbert transform, Gabor transform, Gabor–Wignertransform, S transform, and Hilbert–Haung transform. Thispaper presents a new method for detection and classificationof single and combined PQ disturbances using a sparse signaldecomposition (SSD) on overcomplete hybrid dictionary (OHD)matrix. The method first decomposes a PQ signal into detailand approximation signals using the proposed SSD techniquewith an OHD matrix containing impulse and sinusoidal elemen-tary waveforms. The output detail signal adequately capturesmorphological features of transients (impulsive and oscillatory)and waveform distortions (harmonics and notching). Whereasthe approximation signal contains PQ features of fundamental,flicker, dc-offset, and short- and long-duration variations (sags,swells, and interruptions). Thus, the required PQ features areextracted from the detail and approximation signals. Then, ahierarchical decision-tree algorithm is used for classification ofsingle and combined PQ disturbances. The proposed methodis tested using both synthetic and microgrid simulated PQdisturbances. Results demonstrate the accuracy and robustnessof the method in detection and classification of single andcombined PQ disturbances under noiseless and noisy conditions.The method can be easily expanded for compressed sensing basedPQ monitoring networks.

Index Terms— Compressed sensing, disturbance classification,overcomplete dictionary, power quality (PQ) signal analysis,power system monitoring, sparse representation.

I. INTRODUCTION

POWER quality (PQ) monitoring has become an importantpart of power distribution networks to avoid equipment

damage and to determine the cause of the disturbances [1].Several approaches for detection and classification of PQdisturbances have been developed based on the digital filters,morphology operators, short-time Fourier transform (STFT),wavelet transform (WT), wavelet packet transform (WPT),

Manuscript received February 24, 2014; revised May 16, 2014; acceptedMay 18, 2014. The Associate Editor coordinating the review process wasDr. Kurt Barbe.

M. S. Manikandan and S. R. Samantaray are with the School of Elec-trical Sciences, IIT Bhubaneswar, Bhubaneswar 641112, India (e-mail:[email protected]; [email protected]).

I. Kamwa is with Hydro-Qubec/Hydro-Quebec Research Institute, Varennes,QC J3X 1S1, Canada (e-mail: [email protected]).

Color versions of one or more of the figures in this paper are availableonline at http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TIM.2014.2330493

Hilbert transform (HT), Gabor transform (GT), Wigner distrib-ution function (WDF), S-transform (ST), Gabor-Wigner trans-form (GWT), Hilbert–Haung transform (HHT), and hybridtransform based methods [2]–[21].

Lieberman et al. [18] presented an excellent reviewon techniques and methodologies for PQ analysis anddisturbances classification in power systems. The disturbanceclassification method commonly consists of three majorsteps: preprocessing, feature extraction, and classification.Tse et al. [5] summarized major limitations of differenttechniques, such as STFT, WT, WDF, GT, and GWT foranalysis of PQ signals. The discriminative wavelet featuresextraction highly relies on the optimal wavelet, number ofdecomposition level and characteristics subbands. As shownin [5], the discrete WT and WPT exhibits varying spectraleakages in subbands. The ST uses a scalable Gaussianwindow for analyzing and detecting PQ disturbances [17].The optimal parameters of Gaussian window are chosen topreserve temporal–spectral characteristics of PQ disturbances.The HHT by the empirical mode decomposition techniqueis very much influenced by the noise superimposed on thesignal. In our previous work [21], it is observed that a set ofelementary waveforms from a single basis matrix (or multi-wavelet bases) may not flexible enough to capture essentialPQ features of different types of PQ disturbances. Althoughmany decomposition techniques used in the aforementionedclassification methods, it is not enough to capture meaningful,discriminative morphological features for classification ofdifferent PQ disturbances appeared simultaneously [2].

In this paper, we present a new method for detection andclassification of PQ disturbances using sparse signal decompo-sition (SSD) technique with overcomplete hybrid dictionaries(OHDs). The PQ features extracted from both detail andapproximation signals are used for detection and classificationof single and combined of PQ disturbances. A hierarchicaldecision-tree (HDT) algorithm is presented to reduce compu-tational complexity. This paper mainly deals with the signaldecomposition, feature extraction, event detection, and classi-fication. Finally, to demonstrate the effectiveness of the SSDtechnique, the proposed method is tested with different typesof single and combined PQ signals. The results show that theproposed method is proven to be capable of achieving betteraccuracy in detection and classification of PQ disturbances,

0018-9456 © 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.


2 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT

even under noisy conditions. It is noted that the SSD techniqueis a new promising tool for analysis of PQ disturbance signals.

The structure of this paper is as follows. Section IIdescribes the proposed SSD technique with OHDs. Section IIIpresents PQ disturbance detection and classification algorithm.In Section IV, various simulations are carried out to demon-strate the effectiveness of the proposed decomposition tech-nique on detecting and classifying individual and combinedPQ disturbances. Finally, the conclusion is drawn in Section V.

II. SPARSE REPRESENTATION BASED PQ DISTURBANCE

DETECTION AND CLASSIFICATION

The proposed method consists of four major stages: SSD,event detection, PQ feature extraction, and classification.This paper first presents a SSD technique with OHDs fordecomposing PQ signal into two-subsignals: the detail signalcontaining transients and waveform distortions (such as har-monics, interharmonics, spike/impulse, notching, and noise)and the approximation signal containing fundamental andother sinusoidal PQ disturbances (such as flickers, sags, swells,interruptions, dc-offset, and frequency variation). Second, theproposed method extracts PQ indices from the output detailand approximation signals of the SSD technique. Finally, themethod compares the extracted PQ indices with their typicalparameters for detection and classification of different typesof PQ disturbances.

A. Description of the SSD

1) Sparse Representation: In recent years, sparse represen-tation has been successfully used in many signal processingapplications, including audio and arrhythmia classification,denoising, and compression [22]. Let assume that the recordedcomposite signal x is sparse in an N × M OHD matrix � ∈R

N×M , M > N that contains hybrid elementary waveformsfrom different analytical functions. For a predefined OHDmatrix � with M > N , a N × 1 signal x can be expressed as

x = �α =M∑

m=1

αmψm (1)

where α = [α1, α2, α3, . . . . . . αM ] is the sparse transformedcoefficients obtained for OHDs unlike other time-frequencyanalysis techniques using a single basis matrix. In general,an optimal OHD matrix can be constructed based on priorinformation about temporal and spectral characteristics of thesignals of interests. An overcomplete dictionary that is able toprovide better representation of a composite signal, includingboth stationary and nonstationary components can either beconstructed using hybrid elementary waveforms from differentanalytical functions or by adapting the dictionary to a set oftraining recorded signals [22].

2) Design of OHDs: Based on prior information aboutsparseness of the PQ signals of interests, the OHD matrix� may be constructed by different elementary waveformsgenerated from a variety of basic functions, such as impulse,cosines, sines, Fourier, wavelet, Gabor, and so on. In practice,the acquired PQ signal is a mixture of fundamental component,

transients, and background noises. These PQ signals may notexhibit better sparsity in a single basis matrix. For example, thesinusoidal components of the PQ disturbances are compactlyapproximated by elementary sinusoids, which are sparse in thefrequency-domain. Meanwhile, the transient components, suchas impulsive and oscillatory transients, spikes, notches, andother very-high slope portions are compactly approximated byelementary impulse waveforms, which exhibit sparsity in thetime-domain.

Choosing a predefined OHD matrix is appealing becauseit leads to simple and fast algorithms for the sparse repre-sentations of many classes of signals [22]. In this paper, wechoose an OHD matrix including the elementary waveformsfrom impulse, discrete cosine transform (DCT), and discretesine transform (DST) functions to provide better decompo-sition of PQ disturbance signals. The choice of DCT andDST dictionaries is motivated by effective representation offundamental sinusoidal, harmonics, and flicker components.Compared with a learned dictionary approach [22], such pre-defined OHD matrix have the potential to provide an adequatesparse representations of transient and sinusoidal componentsof PQ signals. Here, the OHD matrix with size of N × M isconstructed as

[�]nm = [I | C | S

]N×M (2)

where I denotes impulse dictionary with size of N × N , Cdenotes the discrete cosine dictionary with size of N × L, andS denotes the discrete sine matrix with size of N × L. Theimpulse dictionary I is constructed as

[IN ]

i j =

⎡⎢⎢⎢⎢⎢⎣

1 0 · · · 0 00 1 · · · 0 00 0 1 0 0......

.... . . 0

0 0 0 · · · 1

⎤⎥⎥⎥⎥⎥⎦

N×N

. (3)

The elementary impulse waveforms from the column vec-tors of impulse dictionary matrix I can provide a better approx-imation of impulsive and oscillatory transients, spike, notches,and discontinuity disturbances contained in the electric powersignal. In this paper, the discrete cosine waveforms are fromcolumn vectors of the DCT dictionary, which is computed as

[CN ]

i j =√

2

N

[εi cos

(π(2 j + 1)i

2N

)](4)

i, j = 0, 1, 2, 3, . . . ,L − 1

where εi = 1/√

2 for i = 0, otherwise εi = 1. The discreteelementary sine waveforms are from column vectors of theDST dictionary, which is computed as

[SN ]

i j =√

2

N

[εi sin

(π(2 j + 1)(i + 1)

2N

)](5)

i, j = 0, 1, 2, 3, . . . ,L − 1

where εi = 1/√

2 for i = N − 1, otherwise εi = 1. Theelementary waveforms from the DCT and DST dictionariescan effectively capture short- and long-duration PQ distur-bances including sags, swells, fluctuations, and harmonics.Selection of both DCT and DST dictionaries is appealing


MANIKANDAN et al.: DETECTION AND CLASSIFICATION OF PQ DISTURBANCES 3

because the sparse approximation coefficients can be directlyused for estimating power quantities of sinusoidal components.In addition, in many cases it leads to minimize the recon-struction error at the boundary of analyzing signal block.Depending on the detection, localization, measurement, andclassification problems, the columns of the OHD matrix can bedesigned for decomposition. In our detection and classificationproblem, the size of sinusoidal dictionary is set as 200, whichis computed for the frequency component of 250 Hz withblock duration of 20 ms and sampling rate of 20 000 Hz. Thenumber of columns of sinusoidal dictionary is chosen basedon the range of frequencies of the signal components to bepreserved in the approximated signal.

3) Sparse Coefficients Estimation: For an input signalx ∈ R

N×1 and a predefined or learned dictionaries � ∈ RN×M

with M > N , a sparse coefficients vector α is estimated bysolving the following optimization formulation [22]:

arg minα

‖α‖0 subject to x = �α. (6)

The sparse coefficients vector specifies the weight for each ofthe elementary waveforms from representation dictionary � .‖α‖0 is the �0-norm which measures the number of nonzerovalued coefficients in the vector α. To combat complexityinvolved in the above optimization problem, many nonlinearoptimization algorithms are reported to estimate sparest vectorα that approximates �0-norm by an �1-norm [23], [24]

arg minα

‖α‖1 subject to x = �α. (7)

The above optimization problem can be solved using basispursuit algorithms or nonlinear iterative greedy algorithms,such as matching pursuit and orthogonal matching pursuit [25].In this paper, the convex optimization problem is solved usingan well-known �1-norm minimization algorithm

α̃ = arg minα

{‖�α − x‖22 + λ‖α‖1

}(8)

where ‖�α − x‖22 and ‖α‖1 are known as the reconstruction

error term and the sparsity term, respectively, x is a signalto be decomposed, and λ is a regularization parameter forsparsity that controls the relative importance of the fidelityand the sparsity of vector α. The �1-norm and �2-norm ofthe vector α are defined as ‖α‖�1 = ∑

i |αi | and ‖α‖�2 =(∑

i |αi |2)1/2, respectively. The estimated sparse coefficientsvector α̃ is composed of detail coefficients (or high-frequencycomponents) and approximation coefficients (or low-frequencycomponents) of the PQ signal. Thus, we define an estimatedsparse coefficients vector α̃ as

α̃ = [d c s]T (9)

where d is the detail coefficients vector obtained for theimpulse elementary matrix I with size of N × N , c is the DCTcoefficient vector obtained for the cosine dictionary matrix Cwith size of N×L, and s is the DST coefficient vector obtainedfor the sine dictionary matrix S with size of N × L. Using (1)and (9), the input signal x can be expressed as

x ≈ �α̃ = [I C S] α̃ = Id + Cc + Ss. (10)

Fig. 1. Example of fundamental signal with notch events. (a) PQ signal withSNR of 35 dB. (b) Extracted detail signal. (c) Measured rms envelope of thedetail signal. (d) Extracted approximation signal. (e) Measured rms envelopeof the approximation signal. The pu stands for per unit measurements.

Finally, an input PQ signal x can be represented as a linearcombination of weighted elementary waveforms from OHD

x =N∑

m=1

dm{Im}Nn=1+

L∑

m=1

cm{Cm}Nn=1+

L∑

m=1

sm{Sm}Nn=1. (11)

From the above expression, the detail signal xd[n] correspond-ing to the N × N impulse dictionary matrix I and detailcoefficient vector d is computed as

xd =N∑

m=1

dm{Im}Nn=1 = d1I1 + d2I2 + · · · + dN IN (12)

where the size of mth column vector Im is N ×1. The approx-imation signals for the DCT and DST coefficient vectors withtheir DCT and DST dictionaries C and S are computed as

xc =L∑

m=1

cm{Cm}Nn=1 = c1C1 + c2C2 + · · · + cLCL (13)

xs =L∑

m=1

sm{Sm}Nn=1 = s1S1 + s2S2.+ · · · + sLSL. (14)

Finally, the reconstructed signal x̂[n] is computed as

x̂[n] = xd[n] + xc[n] + xs[n], n = 0, 1, . . . N − 1 (15)

= xd[n] + xa[n]. (16)

B. Performance of SSD Technique in DetectingPQ Disturbances and Measuring Their PQ Indices

In this paper, the detail signal xd[n] and approximationsignal xa[n] are further processed for the detection and local-ization of PQ disturbances. To demonstrate the effectiveness ofthe proposed SSD technique, a wide variety of PQ disturbancesignals are generated by the synthetic model, the experimentalsetup shown in [20, Fig. 11(a)] and the microgrid modelshown in [19, Fig. 18]. The decomposition results are shown



Fig. 2. Example of fundamental component with oscillatory transient.(a) PQ signal with SNR of 45 dB. (b) Extracted detail signal. (c) Mea-sured rms envelope of the detail signal. (d) Extracted approximation signal.(e) Measured rms envelope of the approximation signal.

Fig. 3. Example of fundamental component with 70% voltage sag. (a) PQsignal with SNR of 45 dB. (b) Extracted detail signal. (c) Measured rmsenvelope of the detail signal. (d) Extracted approximation signal. (e) Measuredrms envelope of the approximation signal.

in Figs. 1–4. In all these figures, the analyzing PQ signalis shown in Fig. 1(a). The output detail and approximationsignals are shown in Fig. 1(b) and (d), respectively. Theinstantaneous root mean square (rms) waveforms obtained forthe detail and approximation signals are shown in Fig. 1(c)and (e), respectively. From the output detail signals shown inthe second plot of Figs. 1–4, it is noted that the elementaryimpulse waveforms from the impulse dictionary I can capturePQ features including amplitude, duration, time of occurrence,polarity, and shape of the disturbances. Thus, the output detailsignal can be directly processed for detecting and classifyingthe different types of transients and waveform distortions.From the output approximation signals shown in the fourthplot of Figs. 1–4, it is noted that the sinusoidal elementarywaveforms from C and S dictionaries can capture sinusoidal

Fig. 4. Example of normal PQ signal with frequency of 50 Hz. (a) PQ signalwith SNR of 45 dB. (b) Extracted detail signal. (c) Measured rms envelopeof the detail signal. (d) Extracted approximation signal. (e) Measured rmsenvelope of the approximation signal.

components of the analyzing signal. For the normal PQ signalwith fundamental frequency of 50 Hz, as shown in Fig. 4,it is found that the magnitude of detail samples are verysmall in the order of 10−5. The maximum magnitude ofthe detail signal is 8.413 × 10−5. The approximation signalprovides PQ features including amplitude, frequency, andphase of sinusoidal components. By referring the measuredrms envelopes of the detail and approximation signals shownin Fig. 4, the proposed SSD technique is able to detectthe presence of PQ disturbance by comparing the minimumdisturbance magnitude threshold (DRMS > 0.01 pu (per unit))and the acceptable magnitude range of normal signal (0.9 pu <ARMS > 1.1 pu), as described in Table IV, according to theIEEE 1159-1995 standard [1].

Fig. 5 shows the effectiveness of the proposed SSDtechnique and other popular decomposition techniques(DWT and ST) for PQ signal analysis in electric powersystems. The real-time PQ signals generated by the experi-mental setup shown in [20, Fig. 11(a)] and by the microgridmodel setup, as shown in [19, Fig. 18], are digitized withsampling rates of 25 000 and 20 000 Hz, respectively. Theresults of different decomposition techniques (SSD, DWT, andST) for input PQ signal with large spike events are shownin Fig. 5(a). For comparison purpose, the subbands of DWTare extracted from eight level decomposition structure withdb20 wavelet, as described in [5]. Fig. 5(a) shows the resultsof different decomposition techniques for input PQ signalwith 65% sag and transient events. From the outputs of theproposed SSD technique, it can be noticed that the detail signalcompletely preserves the morphological features (amplitude,duration, shape, and polarity) of short-duration spike event ascompared with other decomposition techniques. Although, theDWT is useful for detecting transient events, it may not be aneffective approach for measuring certain distinctive PQ indicesincluding magnitude, duration, spectral content, polarity ofthe transient events (impulsive/spike, oscillatory), single notch,



Fig. 5. Effectiveness of the proposed SSD technique and other populardecomposition techniques (DWT and ST) for PQ signal analysis in electricpower systems. (a) Results of different decomposition techniques (SSD, DWT,and ST) for input PQ signal generated by the experimental setup shownin [20]. The subbands of DWT are extracted from eight level decompositionstructure with db20 wavelet, as described in [5]. (b) Results of differentdecomposition techniques for input PQ signal containing 65% sag andtransient events at 0.05 s generated by the microgrid model setup, as shownin [19, Fig. 18].

and multiple notches. By referring the results of DWT in 5(a),it is noted that components of transient event are spreadacross subbands. Thus, it is difficult to select characteristicssubbands to estimate the power quantities including amplitude,frequency, and phase of the fundamental component, flickers,harmonics and interharmonics, and also the instantaneous rmsvalues that can be used for discriminating the short- andlong-duration variations (sags, swells, and interruptions) anddc-offset disturbances.

We further studied the effectiveness of the SSD techniquein removal of sinusoidal components of the PQ disturbancesignal that are not applicable for transient disturbance eventdetection and measurement of PQ indices of transient events.Fig. 6(a) shows the input PQ signal generated by the microgridmodel setup, as shown in [19, Fig. 18]. It contains bothshort-duration transient and flicker events. Fig. 6(b1) and (b2)shows the detail and approximation signals obtained using theSSD technique. It shows that the SSD technique with impulsedictionary (or impulse model) is able to preserve amplitude,duration, shape, and polarities of transient more accurately ascompared with the decomposition results of DWT, as shownin Fig. 6(c1)–(c3). The result of the ST is shown in Fig. 6(d).Thus, the impulse dictionary is more suitable to detect andlocalize transients in the signals of interests. Fig. 6(b2) clearlyshows that the approximation signal contains both fundamen-tal and flicker components. Since the impulsive dictionary

Fig. 6. Performance of the proposed SSD technique and other populardecomposition techniques (DWT and ST) for the PQ disturbance signalcontaining both transient and fluctuation events. (a) PQ signal generatedby the microgrid model setup, as shown in [19, Fig. 18]. (b1) and (b2)Detail and approximation signals obtained using the proposed SSD technique,respectively. (c1)–(c3) Subbands extracted from eight level decompositionstructure with db20 wavelet, as described in [5]. (d) Output of ST.

can adequately capture the transients, spikes, notches, andbackground noise components, the essential PQ quantities(amplitude, frequency, and phase) can be directly estimatedfrom the approximation coefficients vector more reliably andaccurately.

Using the SSD technique, the power quantities (amplitude,frequency, and phase) of fundamental, harmonics/interharmonics, frequency variation, and flicker componentscan be estimated by constructing the sinusoidal dictionarysuch that the elementary sinusoids can capture all thesinusoidal components of the PQ signal x[n]. For estimationof power quantities, the approximation coefficients areprocessed to compute amplitude and phase of the PQ signal.

The amplitude Ak is computed as: Ak =√

c2k + s2

k and the

phase is computed as: φk = tan−1( skck), where ck denotes the

kth coefficient value from cosine coefficients vector c andsk denotes the kth coefficient value from sine coefficientsvector s. As shown in [5], the PQ signal is synthesizedfor validating the performance of the SSD technique. Thesynthesized PQ signal contains nine harmonic components[seven integer harmonics and two interharmonics (475and 775 Hz)], as shown in Table I. For the predefinedOHD matrix, including the impulse dictionary with sizeof 4000 × 4000 and the sinusoidal dictionary with size of2250 × 2250, the harmonic estimation approach based onthe SSD technique is able to measure the power quantitiesof harmonics/interharmonics with frequency componentsare below 2500 Hz. The result of harmonic parameterestimation approach is shown in Table I. The performanceof the proposed SSD based approach is comparable with theresults obtained by the hybrid transforms (DWT and HT).



TABLE I

COMPARISON OF HARMONIC ESTIMATION RESULTS

Fig. 7. Performance of harmonic parameter estimation using the proposedSSD technique for the PQ signal containing nine harmonic components (seveninteger harmonics and two interharmonics). (a) Results of SSD techniqueand estimation of power quantities (amplitude, frequency, and phase) forsynthesized harmonic signal with SNR of 20 dB. (b) Results of SSD techniqueand estimation of power quantities for the harmonic signal with SNR of 10 dB.

The estimation results are shown in Fig. 7. It is noted thatthe amplitude, frequency, and phase values are computedwith estimation error of ±0.006 pu, ±1.1 Hz, and ±2°.Furthermore, the estimation performance is further studiedusing the noisy PQ signals with signal-to-noise ratio (SNR) of20 and 10 dB. Results in Fig. 7 shows that the SSD techniqueis able to elucidates the harmonic components under lowSNR conditions. For the noisy PQ signals, the amplitude,frequency, and phase estimation errors are ±0.014 pu,±1.1 Hz, and ±6.4°.

In practice, recorded PQ signals may be corrupted withrandom Gaussian noise introduced by recording devices andcommunication channel. Therefore, we studied the robust-ness of the SSD technique in detecting PQ disturbances andextracting the essential PQ indices (rms and duration values)and power quantities (amplitude, frequency, and phase) under

Fig. 8. Robustness of the proposed SSD technique on extracting theinstantaneous amplitude (rms values) estimation for the PQ signal with 35%swell with duration of 100 ms that is generated by the microgrid model setup,as shown in [19, Fig. 18]. (a) Decomposition and estimation results for thePQ signal with SNR of 30 dB. (b) Results for the PQ signal with SNR of10 dB. (c) Results for the PQ signal with SNR of 5 dB.

low SNR conditions. In this paper, the PQ signals arecorrupted with additive white Gaussian noise at differ-ent SNRs that can emulate the instrumentation and chan-nel noises introduced by the PQ data acquisition devicesand communication channels. Figs. 7 and 8 show theeffectiveness of the SSD technique in detecting and esti-mating sinusoids corrupted by random Gaussian noise.From the detail signals, as shown in the second plots ofFigs. 7(a) and (b) and 8(a)–(c), it is noted that the backgroundnoise components can be adequately removed from the PQsignal using the hybrid dictionaries, including impulse andsinusoidal elementary waveforms. Furthermore, the measuredinstantaneous amplitude rms results, as shown in the thirdplots of Fig. 8(a)–(c) demonstrate that the SSD techniquecan be suitable for characterization of levels of backgroundnoise.

For noisy PQ signals with 35% swell at 61 ms and eventduration of 100 ms, the decomposition results, as shown insecond and fourth plots of Fig. 8(a)–(c) demonstrate that theSSD technique can capture amplitude variation of swell eventunder low SNR condition. The measured PQ indices from theapproximation signal are shown in Table II. For all SNR values(30, 20, 10, and 5 dB), it is noted that the measured disturbanceduration and amplitude values are comparable with thosevalues obtained by processing the clean PQ signal generatedby the microgrid model setup, as shown in [19, Fig. 18].From the instantaneous amplitude measurement results, asshown in the fifth plots of Fig. 8(a)–(c) and Table II, it is notedthat the SSD technique is capable of providing better resultsin detecting and measuring PQ indices, such as instantaneousamplitude and duration of sags/swells, interruptions anddc-offset disturbances under low SNR conditions.

Table III reports the performance of the proposed SSDtechnique for different values of regularization parameter (λ)



TABLE II

MEASURED PQ INDICES, INCLUDING rms VALUE ( ARMS) AND EVENT

DURATION ( AT) UNDER DIFFERENT SNR CONDITIONS

TABLE III

PERFORMANCE OF THE PROPOSED SSD TECHNIQUE FOR DIFFERENT

VALUES OF REGULARIZATION PARAMETER (λ) UNDER DIFFERENT

PQ SIGNALS WITH SNR VALUES OF 40 AND 30 dB

under different PQ signals with SNR values of 40 and 30 dB.To select an optimal λ value, the SNR and maximum absoluteamplitude error metrics are used for assessing the quality ofreconstructed signal. Furthermore, the computational time isalso computed for different values of λ. It is noted that thecomputational time decreases when the λ value increases.Table III shows the most important PQ indices are measuredfrom the detail and approximation signals. From these results,it is noted that the proposed SSD technique with λ of 0.1 canprovide better noise reduction without significantly distortingthe PQ morphological features.

III. PQ DISTURBANCE DETECTION AND

CLASSIFICATION ALGORITHM

After the decomposition of the PQ signals, the PQ featuresand decision rules are used to detect and classify PQ dis-turbances. The detection, feature extraction, and classificationstages are described in detail in the following subsections.

A. Detail Signal Features for Detection and Classificationof Transients and Waveform Distortions

In this paper, distinctive PQ features are extracted from theoutput detail and approximation signals obtained using theSSD technique. The signal block length (Bdur) is equal to10 or 12 Hz (or a block duration of 200 ms) for both 50- and60-Hz systems, respectively. Different types of feature can beextracted from the detail and approximation signals dependingon the number of single and combined PQ disturbancesconsidered in the detection and classification problem in hand.However, one may believe that selection of optimal featurescan improve accuracy of classification systems with significantreduction in computation load.

The PQ features are extracted from the output detail signalto detect and distinguish between different types of transients(such as impulsive and oscillatory) and types of waveformdistortions (such as harmonics, interharmonics, single/multiplenotch, and noise). According to the IEEE 1159-1995 PQstandard [1], these disturbances exhibit certain distinctive char-acteristics, such as duration (or width), magnitude, polarity,and spectral, content that can be used for their detection andclassification. Therefore, the proposed method extracts PQfeatures, including maximum rms (Dmax), duration (DT),number of isolated events (DNE), and peak polarity of event(DSIGN) from the rms envelope obtained for the output detailsignal xd[n] of the proposed SSD technique. In this paper, rmsenvelope is computed using rectangular window with durationof 2 ms and window shift of 1 sample. The rms envelope xdrmsis computed as

xdrms[ j ] =

√√√√√ 1

N

j+N−1∑

n= j

x2d [n], j = 0, 1, 2, .....N − 1 (17)

where N is the number of samples per half cycle of thefundamental, xd [n] is the nth sample of the detail signal,and xdrms[ j ] is the j th sample of rms value. In this paper,the rms envelope xdrms is further processed for measuring theduration, time-localization, magnitude, and number of isolatedevents within a signal block with duration of 200 ms. Fordifferent PQ disturbance signals, the rms envelopes of theirdetail signals are shown in the third plot of Figs. 1–4. Theresults demonstrate that the output detail signal of the proposedSSD technique can be directly used to identify the type oftransient and waveform disturbances and to determine its PQparameters, such as magnitude, duration, time of occurrence,and polarity. The details of PQ features extracted from theoutput detail signal are given below.

1) DF1: Maximum rms (Dmax): In this paper, the magni-tude (Dmax) of the events is determined as the maximum



value of the rms envelope xdrms feature that is usedto detect the presence of disturbance in an analyzingsignal. The maximum rms value (Dmax) is computed asDmax = maxN−1

j=0 xdrms[ j ].Based upon typical magnitude ranges of PQ distur-bances described in the IEEE 1159-1995 standard [1]and decomposition results under noisy conditions, themagnitude threshold (εdrms) is set to 0.01 pu to detectthe presence of disturbances including transients andwaveform distortions. Thresholding is then applied onthe resulting rms envelope xdrms to detect the distur-bances (Figs. 1–4). Results reported in Table III showthat by adjusting the (εdrms) level, the sensitivity of thedisturbance detection method can easily be controlledunder noiseless and noisy conditions.

2) DF2: Duration (DT): The durations of detected eventswith a predefined magnitude threshold of 0.01 puare used to distinguish between individual types ofspike/notch (DT < 5 ms), oscillatory transients(5 ms < DT < 50 ms), and harmonics/interharmonics(DT > 50 ms) according to the IEEE 1159-1995PQ standard. The duration of the event is determinedas the time during which xdrms continuously exceedsthe magnitude threshold level (εdrms). If the durationof the event is shorter than the duration-threshold of50 ms, the detected event is a transient group, includingimpulsive/spike, oscillatory, and single/multiple notchevents. Otherwise, the event is a waveform distortionclass, including harmonic, interharmonic, and noise.However, in the classification stage, the detail signal isfurther processed to extract the autocorrelation featurefor discriminating harmonics events from noise.

3) DF3: Peak polarity (DSIGN): According to the IEEE1159-1995 standard, a transient can be a unidirectionalimpulse of either polarity or a damped oscillatory wavewith both positive and negative polarity. In additionwith a duration of event, the polarity of the event canbe used to identify whether the detected disturbanceevent belongs to impulsive/spike, notch, or oscillatorytransient class. Unlike other decomposition techniques,the proposed SDD technique can effectively preservesthe polarity of the event (Figs. 1–3).

4) DF4: Number of isolated events (DNE): Based onduration of events detected, total numbers of isolatedevents are computed by comparing with typical durationthresholds and polarity. The number of isolated events(DNE) can be used to identify whether the event belongsto single and multiple transient class (Figs. 1–3).

5) DF5: Time of occurrence of events (DTOC): Based upondecomposition results obtained for disturbances includ-ing sag, swell, interruption, and dc-offset, short-durationtransients may be found at the start and end of thoseevents (Fig. 3). Therefore, the time of occurrence oftransients with duration of less than 5 ms are measuredby processing the rms envelope xdrms.

6) DF6: Autocorrelation of detail signal (DACF): Fordetecting and characterizing the harmonics and thenoise, the index of minimum value of autocorrelation of

TABLE IV

TYPICAL SPECIFICATIONS OF DISTURBANCES AS DEFINED IN

IEEE PQ STANDARDS 1159-2009 [1]. THE pu STANDS

FOR PER UNIT MEASUREMENTS

detail signal are determined. This feature will be com-puted when the detected event duration (DT) equals(Bdur) for the magnitude threshold (Dmax) > 0.01.

B. Approximation Signal Features for Detection andClassification of Short- and Long-Duration Variations

From the results shown in Figs. 1–4, the proposed SSD tech-nique demonstrate that the output approximation signal can beused to identify the type of short- and long-duration variations(including sag, swell, interruption, flicker, and dc-offset) andits amplitude, frequency, and phases of sinusoidal components.The results further show that the proposed method performsbetter noise-reduction. According to the IEEE 1159-1995standard [1], typical duration of those disturbances ranges from0.5 Hz up to more than 1 min, whereas the typical magnitudesrange from 0.1 up to 1.8 pu. The most distinctive PQ featureof these disturbances is the change of rms value of the PQsignal during the disturbance. Table IV shows the typicalparameters of individual short- and long-duration variations.To distinguish between individual types of short- and long-duration variations, the minimum and maximum magnitudesof the xarms envelope that is computed every 10 ms through arectangular window over on the approximation signal. If themagnitude of event is below 0.1 pu, the detected event is aninterruption. The detected events with magnitudes between 0.1and 0.9 pu are classified as sags, whereas the detected eventswith magnitudes between 1.1 and 1.8 pu are classified asswells. The analyzing signals are marked as normal PQ signalswhen their magnitudes between 0.9 and 1.1 pu and maximummagnitude (Dmax) of detail signal is below 0.01 pu (Figs. 1–4).All aforementioned decision rules can accurately detect andclassify different types of single PQ disturbances but it isnot enough to detect and classify many different disturbancesappear simultaneously. Therefore, six PQ features are extractedfrom the xarms envelope for classifying correctly the combinedPQ disturbances. The details of PQ features extracted fromthe rms envelope xarms of the approximation signal are givenbelow.

1) AF1: Mean value of the complete rms envelope xarms.2) AF2: Standard deviation of the complete rms envelope.3) AF3: Mean value of the rms envelope xarms during the

detected event portion.



4) AF4: Standard deviation value of the rms envelope xarmsduring the detected event portion.

5) AF5: Duration value of the detected event that is deter-mined as the time during which rms envelope xarmscontinuously exceeds the upper magnitude thresholdARMSUTH of 1.1 pu and lower magnitude thresholdARMSLTH of 0.9 pu.

6) AF6: Frequency value measured using the index ofminimum value of autocorrelation of the rms envelopexarms with sampling rate of the signal. The AF6 value isbelow 25 Hz for flickers whereas AF6 value is between48 and 52 for dc-offset events.

In this paper, the features AF1 and AF2 are used to detectthe normal, flicker, and dc-offset events. For the Dmax valueless than 0.01, the analyzing signals are marked as normal PQsignals if the AF1 value is between 0.9 and 1.1 pu and the AF2is less than 0.01 pu. The detected event is classified as a flickerif the AF2 value is between 0.01 and 0.1 pu and the AF6 valueis between 0.01 pu and 0.1. The detected event is classifiedas a dc-offset if the AF2 value is between 0.01 pu and 0.1 puand the AF6 value is between 48 and 52 Hz. In the nextsection, the details of decision rules of the HDT algorithm fordetecting and classifying the both individual and combined PQdisturbances based on all aforementioned features extractedfrom the detail and approximation signals.

C. HDT Algorithm for Classification

In this paper, the construction of HDT is different accordingto the one or two features extracted at the each stage forclassification PQ disturbances. The decision rules of HDTalgorithm are presented in Table V. The decision rules forclassification are constructed based on three major classes thatcan be characterized using detail, approximation, and bothdetail and approximation signals: 1) transient and waveformdistortions class (impulsive, oscillatory, notch, and harmonics);2) short- and long-duration variations class (including sag,swell, interruption, flicker, and dc-offset); and 3) combinedPQ disturbances class.

IV. RESULTS AND DISCUSSION

In this section, we study the effectiveness of the proposedSSD technique for detection and classification of individualand combined PQ disturbances. In the case studies, the PQsignals including impulsive and oscillatory transients, swell,sag, harmonics, flicker, interruption, notching, and dc-offsetare generated using both synthetic and microgrid modelsaccording to the IEEE recommended PQ standards [1]. ThePQ signals are generated with sampling rates of 4, 6.4, 10, 20,and 60 kHz, and the resolution of 12 and 16 bits [2], [5], [15].In this paper, the PQ signals are acquired at sampling rate of20 kHz and quantized with resolution of 12 bits.

The results reported in Figs. 1–4 demonstrate the advantageof the proposed SSD technique in detection and classificationof both individual and combined PQ events under noiselessand noisy conditions. The proposed SSD-based method is usedto detect and classify of 32 classes of PQ disturbances (includ-ing 11 individual and 21 combined disturbances). The number

TABLE V

DECISION RULES OF HDT ALGORITHM: DF3 = −1: NEGATIVE

POLARITY; DF3 = +1: POSITIVE POLARITY; DF3 = 0: BOTH

POLARITY PEAKS; AND DF5 = 1: TIME INSTANTS OF

TRANSIENTS ARE COINCIDENCE WITH START AND

END TIMES OF SAG, SWELL, INTERRUPTION, AND

DC-OFFSET FOUND IN THE APPROXIMATION SIGNAL

of PQ events corresponding to each class is 150. Differenttypes of PQ disturbance signals are generated according tothe magnitude, duration, and spectral information describedaccording to the IEEE recommended PQ standards [1]. Here,the fundamental frequency of signal is varied between 48 and52 Hz to study the effectiveness and robustness of the proposedSSD technique.

Based upon results in Table III, it is noted that proper choiceof the regularization parameter plays a significant role in noisereduction and detection of minimum amplitude disturbances.By considering the computational time and noise-reductioncapability, the regularization parameter λ is set to 0.1, whichis suitable enough to detect disturbances within minimummagnitude range described in [1].

The classification accuracy is defined as the ratio of cor-rectly classified events to that of the total number of events.



TABLE VI

PERCENTAGE OF CORRECT CLASSIFICATION RESULTS

OBTAINED FOR VARIOUS SINGLE AND COMBINATION

OF PQ DISTURBANCES WITH DIFFERENT SNRS

The percentage of classification accuracy for 32 types of PQdisturbances is shown in Table VI. The classification resultsdemonstrate that the selected distinctive PQ features fromboth detail and approximation signals are found to be suitableenough to detect and classify PQ disturbances with highlevel of accuracy. The result shows that some degradation ofclassification accuracy for combined PQ disturbances includ-ing harmonics with flicker, sag, swell, and transient event.The performance can be further improved by combining withother decomposition techniques for extracting the distinctivefeatures of these combined disturbances. Table VI shows theimpact of SNR on the performance of the proposed methodand demonstrates the robustness of the method with differentSNRs varying from 45 to 30 dB. The proposed method pro-vides satisfactory classification results by selecting a suitableregularization parameter.

The correct classification results of the proposed methodare compared with other recently reported methods, includingADALINE and FFNN [2], ST and fuzzy decision-tree [3],ST and dynamics [4], and wavelet networks [15], which havebeen tested with both single and combined PQ disturbances.The classification results are summarized in Table VII. Results

TABLE VII

PERFORMANCE COMPARISON IN TERMS OF PERCENTAGE

OF CORRECT CLASSIFICATION RESULTS

show that the proposed method achieves significantly betterclassification results as compared with other existing methods,except for harmonic event with other flicker, sag, and swellevents. From the decomposition and measurement results, asshown in Figs. 1–6, it is noted that the SSD technique canadequately discriminate transients (such as impulsive/spikeand oscillatory) and waveform distortions (such as harmonics,interharmonics, single/multiple notch, and noise) from thevarious types of short- and long-duration variations (includingsag, swell, interruption, flicker, and dc-offset) under bothsingle and combined PQ disturbance conditions. Thus, detailsignal features [DF1–DF6] and approximation signal features[AF1–AF6] as described in Sections III-A and III-B canlead to provide better classification accuracy for single andcombined PQ disturbances. Due to the irregular shape of theinstantaneous amplitude (xarms) obtained for highly varyingamplitude harmonic components in, as shown in Fig. 7(a),and the small magnitude flicker, sag, and swell events buriedin the harmonic events, the method achieves a classificationaccuracy of 82%–90% for the combined disturbances, suchas flicker + harmonic (CL30), sag + harmonic (CL31), andswell + harmonic (CL32).

Based upon results, it is noted that the computation timeof the proposed sparse decomposition technique is high ascompared with the other decomposition techniques. However,computational complexity can be improved using advancedembedded digital signal processor. On the other hand, theproposed detection and classification method can be effectivelyused for further research development in compressed sensingbased PQ monitoring networks.

V. CONCLUSION

This paper presents a new method based on SSD techniquefor detection and classification of single and combined PQdisturbances. The method consists of SSD technique with



OHD matrix, event detection, features extraction, and HDTalgorithms. The decomposition results reported here showthat the output detail signal of the SSD technique can beused directly for extracting the PQ features, including mag-nitude, duration, polarity, time of occurrence, and shape ofthe transients, whereas the approximation signal can able toprovide amplitude, frequency, and phases of the sinusoidalcomponents. The PQ features extraction combining with deci-sion rules of the HDT algorithm can detect and classifydifferent types of PQ events. The performance of the proposedmethod is validated using a large number of PQ events withdifferent noise levels. Results show that the proposed methodachieves significantly better classification results as comparedwith other existing methods, except for harmonic plus otherdisturbances, such as flicker, sag, and swell. At this point,we believe that the proposed method has a great potential incompressed sensing based PQ monitoring applications.

REFERENCES

[1] IEEE Recommended Practice For Monitoring Electric Power Quality,IEEE Standard 1159–2009, 2009.

[2] M. Valtierra-Rodriguez, R. D. J. Romero-Troncoso,R. A. Osornio-Rios, and A. Garcia-Perez, “Detection and classificationof single and combined power quality disturbances using neuralnetworks,” IEEE Trans. Ind. Electron., vol. 61, no. 5, pp. 2473–2482,May 2014.

[3] P. K. Dash and M. Biswal, “Measurement and classification ofsimultaneous power signal patterns with an S-transform variant andfuzzy decision tree,” IEEE Trans. Ind. Inf., vol. 9, no. 4, pp. 1819–1827,Nov. 2013.

[4] S. He, K. Li, and M. Zhang, “A real-time power quality disturbancesclassification using hybrid method based on S-transform and dynam-ics,” IEEE Trans. Instrum. Meas., vol. 62, no. 9, pp. 2465–2475,Dec. 2013.

[5] N. C. F. Tse, J. Y. C. Chan, W.-H. Lau, and L. L. Lai, “Hybridwavelet and Hilbert transform with frequency-shifting decomposition forpower quality analysis,” IEEE Trans. Instrum. Meas., vol. 61, no. 12,pp. 3225–3233, Dec. 2012.

[6] C.-C. Liao, H.-T. Yang, and H.-H. Chang, “Denoising techniqueswith a spatial noise-suppression method for wavelet-based power qualitymonitoring,” IEEE Trans. Instrum. Meas., vol. 60, no. 6, pp. 1986–1996,Jun. 2011.

[7] S.-H. Cho, G. Jang, and S.-H. Kwon, “Time-frequency analysis ofpower-quality disturbances via the Gabor–Wigner transform,” IEEETrans. Power Del., vol. 25, no. 1, pp. 494–499, Jan. 2010.

[8] T. Radil, P. M. Ramos, F. M. Janeiro, and A. C. Serra,“PQ monitoring system for real-time detection and classification ofdisturbances in a single-phase power system,” IEEE Trans. Instrum.Meas., vol. 57, no. 8, pp. 1725–1733, Aug. 2008.

[9] D. Gallo, C. Landi, and M. Luiso, “Accuracy analysis of algorithmsadopted in voltage dip measurements,” IEEE Trans. Instrum. Meas.,vol. 59, no. 10, pp. 2652–2659, Oct. 2010.

[10] J. Barros and R. I. Diego, “Analysis of harmonics in power systems usingthe wavelet-packet transform,” IEEE Trans. Instrum. Meas., vol. 57,no. 1, pp. 63–69, Jan. 2008.

[11] J. Barros and E. Perez, “Automatic detection and analysis of voltageevents in power systems,” IEEE Trans. Instrum. Meas., vol. 55, no. 5,pp. 1487–1493, Oct. 2006.

[12] L. Peretto, M. Artioli, G. Pasini, and R. Sasdelli, “Performance analysisand optimization of a robust algorithm for voltage transients detec-tion,” IEEE Trans. Instrum. Meas., vol. 55, no. 6, pp. 2244–2252,Dec. 2006.

[13] D. G. Ece and O. N. Gerek, “Power quality event detection using joint2-D-wavelet subspaces,” IEEE Trans. Instrum. Meas., vol. 53, no. 4,pp. 1040–1046, Aug. 2004.

[14] J. L. J. Driesen and R. J. M. Belmans, “Wavelet-based powerquantification approaches,” IEEE Trans. Instrum. Meas., vol. 52, no. 4,pp. 1232–1238, Aug. 2003.

[15] M. A. S. Masoum, S. Jamali, and N. Ghaffarzadeh, “Detection andclassification of power quality disturbances using discrete wavelettransform and wavelet networks,” IET Sci. Meas. Technol., vol. 4, no. 4,pp. 193–205, Jul. 2010.

[16] U. D. Dwivedi and S. N. Singh, “Enhanced detection of power-qualityevents using intra and interscale dependencies of wavelet coefficients,”IEEE Trans. Power Del., vol. 25, no. 1, pp. 358–366, Jan. 2010.

[17] P. K. Dash, B. K. Panigrahi, and G. Panda, “Power qualityanalysis using S-transform,” IEEE Trans. Power Del., vol. 18, no. 2,pp. 406–411, Apr. 2003.

[18] D. Granados-Lieberman, R. J. Romero-Troncoso, R. A. Osornio-Rios,A. Garcia-Perez, and E. Cabal-Yepez, “Techniques and methodologiesfor power quality analysis and disturbances classification in powersystems: A review,” IET Generat., Transmiss. Distrib., vol. 5, no. 4,pp. 519–529, Apr. 2011.

[19] S. Kar and S. R. Samantaray, “Time-frequency transform-baseddifferential scheme for microgrid protection,” IET Generat., Transmiss.Distrib., vol. 8, no. 2, pp. 310–320, Feb. 2014.

[20] P. K. Ray, N. Kishor, and S. R. Mohanty, “Islanding and power qualitydisturbance detection in grid-connected hybrid power system usingwavelet and S-transform,” IEEE Trans. Smart Grid, vol. 3, no. 3,pp. 1082–1094, Sep. 2012.

[21] P. Kathirvel, M. S. Manikandan, P. Maya, and K. P. Soman, “Detectionof power quality disturbances with overcomplete dictionary matrix and�1-norm minimization,” in Proc. Int. Conf. Power Energy Syst. (ICPS),Dec. 2011, pp. 1–6.

[22] M. Aharon, M. Elad, and A. Bruckstein, “K -SVD: An algorithm fordesigning overcomplete dictionaries for sparse representation,” IEEETrans. Signal Process., vol. 54, no. 11, pp. 4311–4322, Nov. 2006.

[23] E. J. Candes, J. Romberg, and T. Tao, “Robust uncertaintyprinciples: Exact signal reconstruction from highly incomplete frequencyinformation,” IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489–509,Feb. 2006.

[24] D. L. Donoho and M. Elad, “Optimally sparse representation fromovercomplete dictionaries vial L1-norm minimization,” Proc. Nat. Acad.Sci., vol. 100, no. 5, pp. 2197–3002, 2002.

[25] S. S. Chen, D. L. Donoho, and M. A. Saunders, “Atomic decompositionby basis pursuit,” SIAM J. Sci. Comput., vol. 20, no. 1, pp. 33-–61, 1998.

M. Sabarimalai Manikandan (M’13) receivedthe B.E. degree in electronic and communicationengineering from Bharathiar University, Coimbatore,India, in 2000, the M.E. degree in microwave andoptical engineering from Madurai Kamaraj Univer-sity, Madurai, India, in 2002, and the Ph.D. degreein cardiovascular signal processing from the Depart-ment of Electronics and Communication engineer-ing, IIT Guwahati, Guwahati, India, in 2009.

He served as an Assistant Professor with AmritaVishwa Vidyapeetham University, Ettimadai, India.

He was a Chief Engineer with the Advanced Technology Group, SamsungIndia Electronic Pvt. Ltd., Noida, India. He is currently an Assistant Professorwith the School of Electrical Sciences, IIT Bhubaneswar, Bhubaneswar, India.He holds six filed patents in the field of cardiac health monitoring, speechprocessing, data security, and video summarization systems. His currentresearch interests include signal detection and analysis, biomedical signal andimage processing, pattern recognition, multimedia, affective computing, andwireless health monitoring systems.

Dr. Manikandan was a recipient of the 2012 Outstanding Performance Awardduring his tenure at Samsung India Electronic Pvt. Ltd.



S. R. Samantaray (M’08–SM’10) received theB.Tech. degree in electrical engineering from theUniversity College Of Engineering, Burla, India,in 1999, and the Ph.D. degree in power systemengineering from the Department of Electronics andCommunication Engineering, National Institute ofTechnology, Rourkela, India, in 2007.

He is an Assistant Professor with the School ofElectrical Sciences, IIT Bhubaneswar, Bhubaneswar,India. He was a Post-Doctoral Research Fellow andVisiting Professor with the Department of Electrical

and Computer Engineering, McGill University, Montréal, QC, Canada. Hiscurrent research interests include intelligent protection for transmission sys-tems (including FACTs) and microgrid protection with distributed generationand dynamic security assessment in large power networks.

Dr. Samantaray is an Editor of IET Generation, Transmission and Distribu-tion, Canadian Journal of Electrical and Computer Engineering, and ElectricPower Components and Systems. He was a recipient of the 2007 Orissa BigyanAcademy Young Scientists Award, the 2008 Indian National Academy ofEngineering Best Ph.D. Thesis Award, the 2009 Institute of Engineers (India)Young Engineers Award, the 2010 Samanta Chandra Sekhar Award, and the2012 IEEE PES Technical Committee Prize Paper Award.

Innocent Kamwa (S’83–M’88–SM’98–F’05)received the Ph.D. degree from Laval University,Québec, QC, Canada, in 1988.

He then joined Hydro-Québec’s Research Institute(IREQ), where he is currently the Project Managerof Power Grid Control and Automation. He isalso the Chief Scientist of Hydro-Québec’s SmartGrid. He is a Professional Engineer and an AdjunctProfessor of Power Systems Engineering withMcGill University, Montréal, QC, Canada, andLaval University.

Dr. Kamwa serves on many IEEE/PES technical committees as a memberand an officer, including the Fellow Evaluation, Electric Machinery,and Power System Stability committees. He is an Editor of the IEEETRANSACTIONS ON POWER SYSTEMS and the Co-Editor-in-Chief of IETGeneration, Transmission and Distribution. He was a recipient of the IEEEPES Prize Paper Award in 1998, 2003, and 2009.

detection and classification of power quality disturbances using sparse signal decomposition on...

Documents