soumya r. mohanty,senior member, ieee prakash k. ray ...webx.ubi.pt/~catalao/tste2362797.pdf ·...
TRANSCRIPT
IEEE
Proo
f
IEEE TRANSACTIONS ON SUSTAINABLE ENERGY 1
Comparative Study of Advanced Signal ProcessingTechniques for Islanding Detection in a Hybrid
Distributed Generation System
1
2
3
Soumya R. Mohanty, Senior Member, IEEE, Nand Kishor, Senior Member, IEEE,Prakash K. Ray, Member, IEEE, and João P. S. Catalão, Senior Member, IEEE
4
5
Abstract—In this paper, islanding detection in a hybrid dis-6tributed generation (DG) system is analyzed by the use of hyper-7bolic S-transform (HST), time–time transform, and mathematical8morphology methods. The merits of these methods are thoroughly9compared against commonly adopted wavelet transform (WT)10and S-transform (ST) techniques, as a new contribution to ear-11lier studies. The hybrid DG system consists of photovoltaic and12wind energy systems connected to the grid within the IEEE 30-bus13system. Negative sequence component of the voltage signal is14extracted at the point of common coupling and passed through the15above-mentioned techniques. The efficacy of the proposed meth-16ods is also compared by an energy-based technique with proper17threshold selection to accurately detect the islanding phenomena.18Further, to augment the accuracy of the result, the classifica-19tion is done using support vector machine (SVM) to distinguish20islanding from other power quality (PQ) disturbances. The results21demonstrate effective performance and feasibility of the proposed22techniques for islanding detection under both noise-free and noisy23environments, and also in the presence of harmonics.24
Index Terms—Distributed generation (DG), islanding detection,25power quality (PQ), signal processing, support vector machine26(SVM).27
I. INTRODUCTION28
A LTHOUGH there is a huge expansion of transmission and29
distribution of power supply, it is still inadequate to meet30
the scarcity of the power demand. In this context, penetration of31
renewable energy resources-based distributed generation (DG)32
is gradually increasing, which seems a suitable alternative.33
However, due to the intermittency in wind speed and solar34
radiation intensity, they can be integrated with fuel cells along35
Manuscript received February 16, 2014; revised June 17, 2014 andSeptember 11, 2014; accepted October 05, 2014. The work of J. P. S.Catalão was supported by FEDER funds (European Union) throughCOMPETE and from Portuguese funds through FCT-Portugal, under ProjectsFCOMP-01-0124-FEDER-020282 (PTDC/EEA-EEL/118519/2010) and PEst-OE/EEI/LA0021/2013. Paper no. TSTE-00069-2014.
S. R. Mohanty and N. Kishor are with the Motilal Nehru National Instituteof Technology, Allahabad 211004, India (e-mail: [email protected];[email protected]).
P. K. Ray is with the International Institute of Information Technology,Bhubaneswar 751003, India (e-mail: [email protected]).
J. P. S. Catalão is with the University of Beira Interior, 6201-001 Covilha,Portugal, with the Instituto de Engenharia de Sistemas e Computadores—Investigação e Desenvolvimento (INESC-ID), 1000-029 Lisbon, Portugal, andalso with the Instituto Superior Técnico (IST), University of Lisbon, 1049-001Lisbon, Portugal (e-mail: [email protected]).
Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSTE.2014.2362797
with other storage systems to constitute a hybrid system for reli- 36
able and improved power quality (PQ) [1], [2]. On the other 37
hand, high penetration of such systems raises serious issues 38
related to islanding occurrence events and PQ problems that 39
need to be addressed [3]. 40
Islanding is a phenomenon that occurs when DG resources 41
feed to the local load and the utility grid is disconnected. 42
Accurate and early detection of these disturbances is very 43
important to improve the operation of DG systems. In the past 44
few years, a number of research papers have been reported. 45
Islanding detection techniques can be categorized into active 46
and passive methods [4]. Usually, active methods are based on 47
the injection of a small disturbance into the system and ana- 48
lyzing the change in output parameters for islanding detection. 49
Some of the popular active detection techniques are Sandia fre- 50
quency shift (SFS), active frequency drift (AFD) [5], automatic 51
phase shift (APS), and slip mode frequency shift (SMS) [6]. 52
Although these methods provide less nondetection zone (NDZ), 53
SFS may lead to poor PQ because of positive feedback, while 54
AFD, APS, and SMS may show higher NDZ with an increase 55
of reactive load. Passive methods require monitoring of system 56
parameters and the selection of a suitable threshold in order to 57
detect an islanding event. In fact, challenges exist in selecting 58
the most significant parameters and an appropriate threshold 59
value. The use of voltage unbalance/total harmonic distortion 60
(VU/THD) methodology may cause undesirable trip signal to 61
the circuit breaker due to accumulated sum over the perfor- 62
mance indices within one cycle of load switching, which may 63
be misinterpreted as islanding [7], [8]. The rate of change of 64
frequency (ROCOF) [9], [10] has also been considered for the 65
detection of islanding events. Nonetheless, with the variation 66
of real and reactive power imbalance, this method may lead to 67
NDZ [11]. Intelligent-based techniques for islanding detection 68
are discussed in [12] based on the variations in system param- 69
eter indices and data-mining technology. Then, the transient 70
characteristic of disturbance signals is utilized for detecting 71
power islands [13], [14]. 72
Nevertheless, to minimize NDZ in the above-mentioned 73
methods, signal processing techniques like short-time Fourier 74
transform (STFT), wavelet transform (WT), S-transform (ST), 75
hyperbolic S-transform (HST), and TT transform (TTT) Q176
[15]–[18] are being considered to improve detection perfor- 77
mance. Brief time–frequency characteristics can be conve- 78
niently achieved by using STFT, but the transient signals cannot 79
1949-3029 © 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.
IEEE
Proo
f
2 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY
be adequately described due to its fixed window size. Again,80
wavelet is a strong candidate for classification scheme to ana-81
lyze the signal signature, but is oversensitive to noise signals82
and involves computational complexity. The extension of WT83
with phasor information, known as ST [17], [18], was devel-84
oped for the detection of various islanding and PQ disturbances,85
based on moving a varying and scalable localizing Gaussian86
window [19], [20]. However, in some nonstationary signals87
accompanied with transients, the detection capability of ST also88
degrades [21].89
Therefore, considering the review of available techniques90
reported in the literature, the study in this paper considers a91
modified version of ST, i.e., HST, to compare with WT and92
ST for islanding detection. TTT [22]–[24] that represents time–93
time resolution is also used in the comparative study. These94
techniques are applied to extract more distinctive and pertinent95
statistical features, such as standard deviation (STD), energy96
and cumulative-sum (CUSUM) with selection of an appro-97
priate threshold value, so that islanding events are localized.98
Moreover, mathematical morphology with simple addition and99
subtraction operator [25]–[27] is also explored in this paper for100
islanding detection.101
Hence, as a new contribution to earlier studies, a plethora of102
techniques (HST, TTT, and mathematical morphology, against103
WT and ST) along with support vector machine (SVM) is104
proposed and studied in a comprehensive way for island-105
ing detection and classification of disturbances under various106
operating scenarios.107
This paper is organized as follows. The DG-based hybrid108
IEEE 30-bus system is introduced in Section II. Islanding detec-109
tion methods are given in Section III. The simulation results110
with the comparative study of detection methods performance111
are provided in Section IV. Finally, the conclusion drawn from112
the study is presented in Section V.113
II. DG-BASED HYBRID SYSTEM114
Alternative energy resources, such as photovoltaic, wind115
energy systems, and fuel cells have gained momentum with116
penetration in the main grid to meet the scarcity of power to117
a greater extent. But, the intermittent nature of both wind speed118
and solar radiation intensity affects the reliability of power119
supply.120
In this context, integration of the resources to form a hybrid121
DG system becomes a suitable alternative in order to minimize122
the problems. The system may be connected to the grid to share123
the excess or deficit power, as per situation demands.124
Fig. 1 shows the IEEE 30-bus hybrid DG system to study125
the effect of wind/photovoltaic penetration on the detection of126
islanding event.127
This system consists of three wind generators denoted as128
DG1 to DG3, and a photovoltaic system denoted as DG4. The129
modeling of the wind energy system, and photovoltaic system,130
with their specifications, in the current study is referred in [2]131
and [18].132
The overall specifications of this system are as follows.133133
1) Generator: rated short-circuit 1200 kVA, 50 Hz, and134
11 kV.135
Fig. 1. IEEE 30-bus hybrid system with wind and photovoltaic penetration. F1:1
2) DG: 136136
a) DG1, DG2, and DG3: wind generators with 200 kW 137
each, 500 V, doubly-fed induction generator type. 138
b) DG4: photovoltaic system with 100 kW, 500 V. 139
c) Boost converter: switching frequency of 8 kHz, 140
input dc voltage of 200 V, output ac of 500 V, 141
line-to-line. 142
d) Loads: 150 kW (at DG1, DG2, and DG3), 80 kW 143
(at DG4). 144
III. ISLANDING DETECTION METHODS 145
This section presents a brief description of the signal pro- 146
cessing techniques used for islanding detection. 147
The application of wavelet and ST was already demonstrated 148
in [18] and [19] and thus not discussed here. 149
The following sections describe HST, TTT, and mathemat- 150
ical morphology that are explored in this paper for islanding 151
detection. Fig. 2 shows the flowchart for islanding detection. 152
A. Hyperbolic S-Transform 153
The phasor information on contrary to WT is provided for 154
the detection of signatures in time–frequency plane. Yet, the 155
localization is carried out by fixed Gaussian window that has 156
no parameter variation with respect to time/frequency and thus 157
sometimes degrade its performance [21]. 158
On the other hand, ST with Gaussian window fails to localize 159
disturbances in time domain, while its variant known as HST 160
that has pseudo-Gaussian hyperbolic window offers greater 161
control over the window function, providing better time and 162
frequency resolutions at low and high frequencies. The hyper- 163
bolic window provides frequency dependence shape along with 164
its width and height. A higher asymmetry of the window at 165
low frequencies leads to an increase in width in the frequency 166
domain, with consequent interference between dominant noise 167
frequencies [21]. 168
IEEE
Proo
f
MOHANTY et al.: COMPARATIVE STUDY OF ADVANCED SIGNAL PROCESSING TECHNIQUES FOR ISLANDING DETECTION 3
Fig. 2. Flowchart for islanding and PQ disturbance detection using ST, HST,and TTT.
F2:1F2:2
In this study, HST contour is used for islanding detection169
where the hyperbolic window is defined as170
Whb =2fs√
2π (αhb + βhb)exp
{−−f2
sX2
2
}(1)
where171
X =(αhb + βhb)
2αhbβhb(τ − t− ξ)
+(αhb − βhb)
2αhbβhb
√(τ − t− ξ)2 + λ2
hb
0 < αhb < βhb and ξ =
√(βhb − αhb)
2λ2hb
4αhbβhb. (2)
The discrete version of HST can be calculated, and172
G(mF , nF ) denotes Fourier transform of hyperbolic window173
G(mF , nF ) =2fs√
2π(αhb + βhb)exp
{−−f2
sX2D
2
}(3)
where174
XD =(αhb + βhb)
2αhbβhbt+
(αhb − βhb)
2αhbβhb
√t2 + λ2
hb. (4)
H[mF , nF ] is the frequency-shifted Fourier transform,175
H[mF ] which is given by176
H[mF ] =1
N
N−1∑
mF=0
h(k) exp(−i2πnF k) (5)
S[nF , j] =
N−1∑
mF=0
H(mF + nF )G(mF , nF ) exp(−i2πmF j)
(6)
where N is the total number of samples and G(mF , nF )177
represents the Fourier transform of the hyperbolic window.178
B. TTT 179
The computation of inverse Fourier transform of discrete ST 180
provides discrete TTT that maps single-dimensional signal in 181
time domain to two-dimensional signal in time domain. 182
Since each of these is associated with a particular window 183
position on the time axis, collectively they give a time–time 184
distribution. The selection of different level as in WT is not 185
required in the TTT. Further, high-frequency component of the 186
signal is better localized in TTT and its output. In this tech- 187
nique, among the different frequency components, the high- 188
frequency components have more energy concentration, thus 189
leading to better localization property. One of the major utility 190
of the TTT is the time-local view, through the scaled windows 191
of the primary time series that makes it an appropriate approach 192
for disturbance detection [22]–[24]. 193
The Fourier transform of the signal is taken as a win- 194
dow sliding along the time axis. Hence, it is designated as 195
the two-dimensional representation of the signal. The general 196
expression of short-time Fourier transform is given by 197
STFT(t, f) =
∞∫
−∞h(τ)w(t− τ) exp(−2πifτ)dτ. (7)
By applying the inverse FT of (7) brings 198
[h(τ)w(t− τ)] =
∞∫
−∞STFT(t, f) exp(−2πifτ)df. (8)
When all values of t are considered, the window function 199
h(τ)w(t− τ) becomes a two-dimensional function denoted by 200
STFTTT, given by 201
STFTTT(t, τ) =
∞∫
−∞STFT(t, f) exp(2πifτ)df. (9)
TTT is obtained from ST as 202
TT(t, τ) =
∞∫
−∞S(t, f) exp(2πifτ)df. (10)
The discrete form appears as 203
TT(kΔt, jΔt) =
N/2−1∑
n=N/2
S(nΔf, jΔt) exp(i2πnk/N) (11)
h(ktΔt) =
N−1∑
j=0
TT(ktΔt, jΔt) (12)
where Δt is the time interval, N is the sampling number, kt 204
is the time index, kt = 0, 1, . . . , N − 1, j is the time shifting 205
index, and nF is the frequency index. 206
C. Mathematical Morphology 207
Morphological filters are nonlinear signal transformation 208
tools that modify shapes of signals [25]. Mathematical mor- 209
phology is developed from set theory and integral geometry, 210
IEEE
Proo
f
4 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY
and derives its name because it deals with the shape of signals.211
It is a nonlinear signal processing approach applied to the study212
on power system fault/disturbance.213
The wavelet, HST, and TTT are all used as integral trans-214
form for detection objective. In fact, computational complexity215
is higher for transients and harmonics. Further, it assumes216
the periodicity of the signal by which accuracy of detection217
degrades. On the contrary, mathematical morphology deals with218
simple addition and subtraction of signals without undergoing219
multiplication and division. With a smaller data window, i.e.,220
with the knowledge of a small portion of the signal, it is able221
to detect the abnormalities quickly and accurately. Dilation and222
erosion are two basic operations with which generalized open223
close operation can be explored. Erosion is a kind of shrink-224
ing transform, which can make the target signal contract with225
holes enlarging. Dually, dilation is an expanding process, which226
realizes the target signal, enlarging together with holes con-227
tracting. Generally, erosion and dilation are not reversible each228
other. So the conjugation of them can form new morphologi-229
cal operators, named by opening and closing, which are also230
important operators in mathematical morphology. Further, the231
choice of structural element should be optimum to enable better232
performance for the morphological filter [26], [27].233
Consider ds(n) as a disturbance voltage signal, defined in234
domain Dds= {x0, x1, . . . , xn}, and gs(m) be the structur-235
ing element (SE) defined in domain Dgs = {y0, y1, . . . , ym},236
where n and m are integers, such that n > m. Then dilation237
(D) and erosion (E) of ds(n) by gs(m) are given as [25]238
yD(n) = (ds ⊕ gs)(n) = max
{ds(n−m) + gs(m)
0 ≤ (n−m) ≤ n,m ≥ 0
}
(13)
yE(n) = (ds Θ gs)(n) = min
{ds(n+m)− gs(m)
0 ≤ (n+m) ≤ n,m ≥ 0
}.
(14)
Now, with the above dilation and erosion, two composite239
operations [opening (O) and closing (C)] are given as240
yO(n) = (ds ◦ gs)(n) = ((ds Θ gs)⊕ gs)(n) (15)
yC(n) = (ds • gs)(n) = ((ds ⊕ gs)Θ gs)(n). (16)
An SE is a function used as a probe to extract features,241
keeping details, and reducing noise from a signal. There is no242
specific rule for appropriate choice of SE. The length of SE243
depends upon the sampling period of the signal. If sampling244
rate increases, accordingly the length of SE will also increase,245
not strictly linearly [28]. Islanding being an online application,246
speed is always an important factor so that a quick response247
is governed in the detection of system disturbances. In a con-248
ventional opening–closing (OC) and closing–opening (CO)249
filter, there exists serious statistical deflection that influences250
noise suppression of morphological filter directly. Because of251
dissimilarity of expansion process in opening operation and252
contraction process in closing operation, overall filtering per-253
formance may not be good. In this context, in order to suppress254
noise and achieve better performance, a different SE is taken.255
To overcome this, generalized morphological filter is imple-256
mented for the detection application. The output of such filter257
Fig. 3. Hyperplane for SVM. F3:1
depends upon proper choice of SE along with morphological 258
transformation 259
FGOC(ds(n)) = ds ◦ gs1 • gs2(n) (17)
FGCO(ds(n)) = ds • gs1 ◦ gs2(n) (18)
where gs1 and gs2 are two SEs such that gs1 ⊆ gs2 . In the 260
present study, gs1 is linear and gs2 is semicircular having length 261
seven. This paper employs the difference in the generalized 262
opening and closing, open, close erosion, and dilation, in order 263
to detect any abnormality present in the signal 264
detGCO(n) = FGCO(ds(n))− FGOC(ds(n))
detOC(n) = FCO(ds(n))− FOC(ds(n))
detDE(n) = FD(ds(n))− FE(ds(n))
⎫⎪⎬
⎪⎭. (19)
Thus, the difference in generalized open and close operation 265
is implemented to detect islanding and PQ disturbances. 266
D. SVM 267
SVM utilizes a statistical learning technique which provides 268
very good accuracy in high-dimensional feature spaces. The 269
classification of patterns is based on structural risk minimiza- 270
tion method. SVM provides improved generalization perfor- 271
mance than other classical techniques, such as ANN and Bays 272
classifier because its training is supported on a sequential mini- 273
mization technique. This makes it suitable for the classification 274
of islanding and different PQ disturbances [17]. 275
For n-dimensional inputs si(i = 1, 2, . . . ,M), M is the sam- 276
ple number fitting to class1 or class2 with outputs oi = 1 for 277
class1 and oi = −1 for class2, correspondingly. The hyperplane 278
is given as 279
f(s) = wT s+ b =
n∑
j=1
wjsj + b = 0 (20)
where w is the n-dimensional vector and b is a parameter. The 280
separating hyperplane position is defined by the w and b values 281
as depicted in Fig. 3. 282
The constraints are f(si) ≥ 1, if oi = 1 and f(si) ≥ −1, if 283
oi = −1, thus 284
oif(si) = oi(wT s+ b) ≥ +1, for i = 1, 2, . . . ,M. (21)
IEEE
Proo
f
MOHANTY et al.: COMPARATIVE STUDY OF ADVANCED SIGNAL PROCESSING TECHNIQUES FOR ISLANDING DETECTION 5
Fig. 4. Islanding detection event for the grid-connected wind energy system.(a) Negative sequence voltage at PCC of bus 13. (b) Detection using dB4wavelet. (c) Detection using S-transform. (d) Detection using TTT. (e) STDprofile using TTT. (f) Detection using HST. (g) Detection using mathematicalmorphology.
F4:1F4:2F4:3F4:4F4:5
Fig. 5. Islanding detection event for the grid-connected wind energy systemwith 20-dB signal to noise ratio. (a) Negative sequence voltage. (b) Detectionusing dB4 wavelet. (c) De-noised negative sequence voltage. (d) Detectionusing dB4 wavelet de-noising. (e) Detection using S-transform. (f) Detectionusing TTT. (g) STD profile using TTT. (h) Detection using HST. (i) Detectionusing mathematical morphology.
F5:1F5:2F5:3F5:4F5:5F5:6
IEEE
Proo
f
6 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY
Fig. 6. CUSUM-based detection of VU and islanding events.F6:1
Fig. 7. Energy-based detection of islanding events.F7:1
The geometrical distance is given as ||w||−2. The opti-285
mal hyperplane can be driven by the following optimization286
problem [25]:287
Minimize1
2||w||−2 + C
M∑
i=1
ξi (22)
subject to288
oi(wT s+ b) ≥ 1− ξi, for i = 1, 2, . . . ,M (23)
ξi ≥ 0, for all i. (24)
289
The optimal bias value b∗ is given as290
b∗ = −1
2
∑
SVs
oiα∗i (v
T1 si + vT2 si) (25)
where v1 and v2 are random SVM for class1 and class2,291
respectively.292
The decision function corresponds to293
f(s) =∑
SVs
αioisTi s+ b∗. (26)
Unidentified data sample s is categorized as294
s ∈{Class− 1, f(s) ≥ 0
Class− 2, otherwise
}
. (27)
The disturbances classification is accomplished applying a295
kernel function, such as Gaussian radial basis kernel function.296
IV. RESULTS AND DISCUSSION297
Simulation results for the detection of islanding events in the298
hybrid DG system, using the above discussed techniques, are299
presented in this section under various operating scenarios. The 300
operating scenarios for the grid-connected hybrid DG system 301
are created in MATLAB/Simulink. The sampling rate in the 302
simulations is 5 kHz with 50 Hz frequency. 303
The voltage signal at PCC, at different DG locations, is 304
retrieved and processed through sequence analyzer to obtain 305
its negative sequence component. The signals are then passed 306
through the algorithms in order to detect and localize the island- 307
ing instants. The results of contour analysis by HST, TTT, 308
and mathematical morphology are compared with the results 309
obtained by using Daubechies 4 (dB4) as the mother wavelet 310
and ST. 311
The quantitative analysis in terms of energy content, STD 312
and CUSUM, of the transformed signal is also carried out 313
in order to demonstrate the efficacy of the proposed detec- 314
tion techniques. The determination of the energy content of 315
the transformed signal is based on “Parseval’s theorem,” which 316
states that the energy of a signal ds(t) remains the same whether 317
it is computed in a signal domain (time) or in a transform 318
domain (frequency) [29]. Thus, the energy and STD of a signal 319
is expressed as 320
Esignal =1
T
T∫
0
|ds(t)|2dt =N∑
n=0
|ds[n]|2 (28)
where T and N are the time period and the length of the signal, 321
respectively, and ds[n] is the Fourier transform of the signal. 322
Similarly, STD can be calculated from 323
STD =
[1
n
n∑
i=1
(ei − e)
]1/2
(29)
where n is the number of sample data and ei, e are the voltage 324
and mean of the voltage signal. CUSUM is computed by the 325
sum of the consecutive samples of the energy of the voltage 326
signal after being passed through different transforms. CUSUM 327
of voltage signal is calculated as 328
CUSUMv(p) =
r∑
1
Evq(r) (30)
where Evq is the spectral energy of the negative sequence volt- 329
age signal at PCC and “r” is the length of signal over which 330
energy and CUSUM are calculated, and “p” is the entire length 331
of the signal. 332
The islanding event is detected when energy content/ 333
STD/CUSUM becomes greater than a threshold value. The 334
results presented in the following sections will clearly reflect 335
the advantages of HST, TTT, and mathematical morphology 336
over ST and WT in the detection of disturbances. 337
A. Islanding Detection for Grid-Connected Energy Resources 338
This subsection describes the detection using above 339
described signal processing techniques. 340
Fig. 4(a) shows the extracted negative sequence voltage sig- 341
nal at PCC13 (i.e., at bus 13), where DG3 is connected, when 342
IEEE
Proo
f
MOHANTY et al.: COMPARATIVE STUDY OF ADVANCED SIGNAL PROCESSING TECHNIQUES FOR ISLANDING DETECTION 7
TABLE IT1:1CLASSIFICATION OF ISLANDING AND PQ DISTURBANCES BASED ON SVMT1:2
the wind energy system is suddenly disconnected from the grid.343
It is observed that islanding event occurs at around 350 sec. The344
detection using dB4 wavelet is shown in Fig. 4(b).345
The detection using ST, TTT, HST, and mathematical mor-346
phology is shown in Fig. 4(c), (d), (f), and (g), respectively.347
These results clearly show a sudden increase in the magni-348
tudes at the instant of islanding event, thus clearly localizing349
and detecting the disturbance.350
The islanding detection in the same system with 20-dB signal351
to noise ratio is shown in Fig. 5.352
Negative sequence voltage signal at PCC, at bus 13 where353
DG3 is connected, with 20-dB noise is shown in Fig. 5(a). It354
can be observed from the graphical results of Fig. 5(b) that the355
WT completely fails in the presence of noise, whereas, Fig. 5(c)356
and (d) show the de-noised negative sequence voltage signal357
and detection using WT, respectively. After de-noising, it is358
observed that the signal contains still some noise and as a result359
the detection using WT shows some deterioration. While ST360
shows the detection of islanding with an increased contour in361
Fig. 5(e), it is accompanied with scattered characteristic, i.e.,362
deterioration in localization.363
Hence, with ST, the detection process gets affected with364
some degradation with the increase in noise level in the voltage365
signal. The presence of noise obstructs an accurate detection,366
so these two techniques (wavelet and ST) are suggested to be367
noise sensitive.368
Instead, as shown in Fig. 5(f)–(i), TTT, HST, and mathemati-369
cal morphology are capable to detect and localize the islanding370
instant effectively. The transient noise characteristic leads to371
the appearance of bulge along the TT contours. Its STD value372
shows an apparent rise at the time of the islanding event.373
The performance can also be presented quantitatively with374
the computation of performance indices: STD, energy content,375
and CUSUM [30], [31]. These features are the most distin-376
guished ones for discriminating islanding events with those of377
nonislanding scenarios. A threshold is selected by comparing 378
these indices for normal operating condition with that of the 379
islanding event. The threshold value for the configuration in 380
which all the resources (wind energy system, fuel cell, and pho- 381
tovoltaic system) are connected to the grid, is selected as 1.0145 382
for STD, 2.2314 p.u. for energy and 3.2546 for CUSUM by 383
using HST. When the STD is more than 1.0145, energy content 384
is more than 2.2314 p.u. and CUSUM is more than 3.2546, then 385
the islanding event is detected; otherwise it corresponds to a PQ 386
disturbance [18], [19]. 387
The CUSUM-based detection is also tested on the IEEE 388
30-bus system, with disturbances in the form of islanding along 389
with reduced 40% and 60% load change at PCC1 of DG1 loca- 390
tion. Fig. 6 shows the discrimination of a 2% VU from islanding 391
of DG1 location at bus 1 (including load) based on CUSUM, 392
with selection of 0.411 as threshold value. Similarly, energy is 393
considered as a common performance index in Fig. 7, which 394
displays the performance of the proposed transforms/MM along 395
with conventional techniques such as THD and VU. 396
It is clear that indices value in case of THD and VU come 397
nearer to the selected threshold, thereby increasing the chance 398
of misdetecting islanding event, whereas, the indices value in 399
case of proposed techniques come much higher with respect to 400
the threshold value, proving the improvement in discriminating 401
islanding from PQ disturbances. ST and HST show oscilla- 402
tions because of variations of frequency during disturbances 403
and being time–frequency resolution analysis. 404
To augment this, the proposed transforms and SVM-based 405
technique are used to classify islanding and different PQ dis- 406
turbances, as shown in Table I. The voltage signal is passed 407
through the transforms ST, HST, TTT, and MM, to find the fea- 408
ture dataset to be trained and tested by SVM in order to classify 409
all the disturbances. The features such as STD, energy, kurto- 410
sis, and skewness are selected and the procedure and parameters 411
cited in [1] and [17] is followed for the classification objective. 412
IEEE
Proo
f
8 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY
Fig. 8. Boundary plot for discriminating islanding and PQ. (a) Islanding versussag with harmonics and swell with harmonics. (b) Islanding versus flicker andoscillatory transients.
F8:1F8:2F8:3
The disturbances presented in Table I are created at differ-413
ent DG locations of the IEEE 30-bus system. Again, to support414
the classification of disturbances, boundary plots are shown in415
Fig. 8, which justify the discrimination of islanding from other416
PQ disturbances. Similarly, other disturbance’s classification417
are studied using the above hybrid technique. A feature data418
matrix equal to 500× 8 is formulated for all the disturbances,419
out of which half of the dataset is used for training and the420
remaining half for the testing of the SVM classifier.421
The classification results given in Table I demonstrate that422
TTT and MM (being time–time transforms) show better accu-423
racy than the time–frequency multiresolution analyses in the424
form of ST and HST. The detection time taken by different425
classification techniques, computed on Pentium-IV, 2.8 GHz,426
1 GB RAM computer in MATLAB environment, is presented in427
Table II. It is observed that the detection time follows IEEE Std.428
1547 in all cases. The islanding event occurs at bus 1, having429
DG1 connected and 2% VU in load is considered. In addition,430
detection time involved for other disturbances such as inrush,431
flicker, oscillatory transient, sag with harmonics are also given432
in that table.433
B. Effect of Harmonics on Islanding Detection434
This section describes the effect of harmonics presence in435
the voltage signal at PCC on the islanding detection (Fig. 9).436
The harmonics in the voltage waveform may be caused due437
to the connection of nonlinear loads in the hybrid system.438
Harmonics, being a vital effect, must be eliminated because439
TABLE II T2:1DETECTION TIME OF DIFFERENT TECHNIQUES T2:2
they lead to malfunctioning and failure of equipment, over- 440
heating, misdetection and misclassification of islanding and PQ 441
problems, among others. 442
To test this effect, a nonlinear load is connected at PCC1, 443
having DG1 connected, followed by islanding at the 825 sam- 444
ple. The corresponding voltage signal is as shown in Fig. 9(a). 445
The increase in voltage in the form of a spike is observed due to 446
islanding occurrence in addition to the presence of harmonics. 447
This voltage signal is processed through WT, ST, HST, TTT, 448
and mathematical morphology to detect the islanding event, 449
shown in Fig. 9(b)–(d), (f), and (g), respectively. Fig. 9(e) 450
represents the variation in STD due to islanding using TTT. 451
The results clearly show again that islanding detection com- 452
pletely fails using WT. Similarly, the detection is degraded to 453
some extent in the case of ST. Instead, using HST, TTT, and 454
mathematical morphology, islanding detection is observed to 455
be much more effective and accurate even in the presence of 456
harmonics. 457
With increased penetration of DG, the islanding and PQ 458
issues become common occurrences that need to be ana- 459
lyzed by means of advanced signal processing techniques, as 460
IEEE
Proo
f
MOHANTY et al.: COMPARATIVE STUDY OF ADVANCED SIGNAL PROCESSING TECHNIQUES FOR ISLANDING DETECTION 9
Fig. 9. Effect of harmonics on islanding detection. (a) Negative sequencevoltage. (b) Detection using dB4 wavelet. (c) Detection using S-transform.(d) Detection using TTT. (e) STD profile using TTT. (f) Detection using HST.(g) Detection using mathematical morphology.
F9:1F9:2F9:3F9:4
discussed in the introduction of the paper. The fast emergence 461
of real-time processors these advanced signal processing tech- 462
niques can be explored to accomplish disturbance detection. 463
Accordingly, various hardware platforms, such as DSP, FPGA, 464
and/or LABVIEW, can be adopted in order to validate the cor- 465
rectness of the proposed method for the detection of islanding 466
and PQ disturbance. Subsequently, the corrective action can 467
be initiated to enhance the reliability of the system, which is 468
indispensable. 469
V. CONCLUSION 470
This paper has presented a new study on the detection of 471
islanding events in a grid-connected hybrid DG system using 472
a plethora of techniques (WT, ST, HST, TTT, and mathematical 473
morphology), under various operating scenarios. The perfor- 474
mance of these approaches for islanding detection was tested 475
and compared under different operating conditions. The con- 476
tour analysis of HST/TTT and mathematical morphology tech- 477
niques demonstrated their capability to clearly detect islanding 478
events under both noise and noise-free conditions. Moreover, 479
the presence of harmonics in the voltage signal confirmed 480
the merits of HST, TTT, and mathematical morphology over 481
WT and ST for the detection and localization of islanding 482
events, which is an important characteristic. Further classifica- 483
tion is accomplished with use of SVM in order to discriminate 484
islanding with other PQ disturbances. 485
ACKNOWLEDGMENT 486
The authors acknowledge the support and facility available 487
from the Department of Electrical Engineering, Motilal Nehru 488
National Institute of Technology, Allahabad, India, for carrying 489
out the research work. 490
REFERENCES 491
[1] P. K. Ray, N. Kishor, and S. R. Mohanty, “Classification of power quality 492disturbances due to environmental characteristics in distributed genera- 493tion system,” IEEE Trans. Sustain. Energy, vol. 4, no. 2, pp. 302–313, 494Apr. 2013. 495
[2] O. C. Onara, M. Uzunoglu, and M. S. Alama, “Modeling, control and 496simulation of an autonomous wind turbine/photovoltaic/fuel cell/ultra- 497capacitor hybrid power system,” J. Power Sources, vol. 185, no. 2, 498pp. 1273–1283, 2008. 499
[3] A. G. Tsikalakis and N. D. Hatziargyriou, “Operation of microgrids with 500demand side bidding and continuity of supply for critical loads,” Eur. 501Trans. Elect. Power, vol. 21, no. 2, pp. 1238–1254, 2011. 502
[4] J. A. Pecas Lopes, N. Hatziargyriou, J. Mutale, P. Djapic, and N. Jenkins, 503“Integrating distributed generation into electric power systems: A review 504of drivers, challenges and opportunities,” Elect. Power Syst. Res., vol. 77, 505no. 9, pp. 1189–1203, 2007. 506
[5] H. Wang, F. Lui, Y. Kang, J. Chen, and X. Wei, “Experimental inves- 507tigation on non detection zones of active frequency drift method for 508anti-islanding,” in Proc. 33rd Annu. Conf. IEEE Ind. Electron. Soc., Nov. 5095–8, 2007, pp. 1708–1713. 510
[6] F. Liu, X. Lin, Y. Kang, Y. Zhang, and S. Duan, “An active islanding 511detection method for grid-connected converters,” in Proc. 3rd IEEE Int. 512Conf. Ind. Electron. Appl., Jun. 3–5, 2008, pp. 734–737. 513
[7] V. Menon and M. H. Nehrir, “A hybrid islanding detection technique 514using voltage unbalance and frequency set point,” IEEE Trans. Power 515Syst., vol. 22, no. 1, pp. 442–448, Feb. 2007. 516
[8] S. Jang and K. Kim, “An islanding detection method for distributed gen- 517eration algorithm using voltage unbalance and total harmonic distortion 518of current,” IEEE Trans. Power Del., vol. 19, no. 2, pp. 745–752, Apr. 5192004. 520
IEEE
Proo
f
10 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY
[9] H. H. Zeineldin, E. F. El-Saadany, and M. M. A. Salama, “Impact of DG521interface control on islanding detection and non-detection zones,” IEEE522Trans. Power Del., vol. 21, no. 3, pp. 1515–1523, Jul. 2006.523
[10] H. H. Zeineldin and S. Kennedy, “Sandia frequency-shift parameter524selection to eliminate non-detection zones,” IEEE Trans. Power Del.,525vol. 24, no. 1, pp. 486–487, Jan. 2009.526
[11] H. H. Zeineldin and M. M. A. Salama, “Impact of load frequency depen-527dence on the NDZ and performance of the SFS islanding detection528method,” IEEE Trans. Ind. Electron., vol. 58, no. 1, pp. 139–146, Jan.5292011.530
[12] K. El-Arroud, G. Joos, I. Kamwa, and D. T. McGillis, “Intelligent based531approach to islanding detection in distributed generation,” IEEE Trans.532Power Del., vol. 22, no. 2, pp. 828–835, Apr. 2007.533
[13] N. W. A. Lidula and A. D. Rajapakse, “A pattern recognition approach534for detecting power islands using transient signals—Part I: Design and535implementation,” IEEE Trans. Power Del., vol. 25, no. 4, pp. 3070–3077,536Oct. 2010.537
[14] N. W. A. Lidula and A. D. Rajapakse, “A pattern recognition approach538for detecting power islands using transient signals—Part II: Performance539evaluation,” IEEE Trans. Power Del., vol. 27, no. 3, pp. 1071–1080, Jul.5402012.541
[15] C.-T. Hsieh, J.-M. Lin, and S.-J. Huang, “Enhancement of islanding-542detection of distributed generation systems via wavelet transform-based543approaches,” Elect. Power Energy Syst., vol. 30, no. 10, pp. 575–580,5442008.545
[16] M. Hanif, M. Basu, and K. Gaughan, “Development of EN50438 com-546pliant wavelet-based islanding detection technique for three-phase static547distributed generation systems,” IET Renew. Power Gener., vol. 6, no. 4,548pp. 289–301, Jul. 2012.549
[17] S. R. Mohanty, P. K. Ray, N. Kishor, and B. K. Panigrahi, “Classification550of disturbances in hybrid DG system using modular PNN and SVM,”551Elect. Power Energy Syst., vol. 44, no. 1, pp. 764–777, 2013.552
[18] P. K. Ray, S. R. Mohanty, and N. Kishor, “Disturbance detection in grid553connected distributed generation system using wavelet and S-transform,”554Elect. Power Syst. Res., vol. 81, no. 3, pp. 805–819, 2011.555
[19] P. K. Ray, N. Kishor, and S. R. Mohanty, “Islanding and power qual-556ity disturbance detection in grid-connected hybrid power system using557wavelet and S-transform,” IEEE Trans. Smart Grid, vol. 3, no. 3,558pp. 1082–1094, Sep. 2012.559
[20] P. K. Dash, B. K. Panigrahi, and G. Panda, “Power quality analysis using560S–transform,” IEEE Trans. Power Del., vol. 18, no. 2, pp. 406–411, Apr.5612003.562
[21] B. Biswal, P. K. Dash, and B. K. Panigrahi, “Non-stationary power sig-563nal processing for pattern recognition using HS-transform,” Appl. Soft564Comput., vol. 9, no. 1, pp. 107–117, 2009.565
[22] C. Simon, M. Schimmel, and J. J. Dañobeitia, “On the TT-transform566and its diagonal elements,” IEEE Trans. Signal Process., vol. 56, no. 11,567pp. 5709–5713, Nov. 2008.568
[23] S. Suja and J. Jerome, “Pattern recognition of power signal disturbances569using S transform and TT transform,” Elect. Power Energy Syst., vol. 32,570no. 1, pp. 37–53, 2010.571
[24] C. R. Pinnegar, “Time-frequency and time-time filtering with S572Transform and TT transform,” Digit. Signal Process., vol. 15, no. 6,573pp. 604–620, 2005.574
[25] P. Li, J. Liu, X. Li, C. Tian, and J. Jie, “Detection of power qual-575ity disturbances in micro-grid based on generalized morphological filter576and backward difference,” in Proc. 4th Int. Conf. Elect. Utility Deregul.577Restruct. Power Technol. (DRPT), Jul. 6–9, 2011, pp. 1063–1066.578
[26] S. Gautam and S. M. Brahma, “Overview of mathematical morphology in579power systems-a tutorial approach,” in Proc. IEEE Power Eng. Soc. Gen.580Meeting (PES’09), Calgary, AB, Canada, Jul. 26–30, 2009, pp. 1–7.581
[27] S. Gautam and S. M. Brahma, “Out of step blocking function in a distance582relay using mathematical morphology,” IET Gener. Transm. Distrib.,583vol. 6, no. 4, pp. 313–319, Apr. 2012.584
[28] C. Xue et al., “Power quality disturbance detect ion and location using585mathematical morphology and complex wavelet transform,” in Proc. 3rd586IEEE Conf. Ind. Electron. Appl., Jun. 3–5, 2008, pp. 2263–2268.587
[29] A. M. Gargoom, N. Ertugrul, and W. L. Soong, “Automatic classification588and characterization of power quality events,” IEEE Trans. Power Del.,589vol. 23, no. 4, pp. 2417–2425, Oct. 2008.590
[30] S. R. Mohanty, A. K. Pradhan, and A. Routray, “Cumulative sum based591fault detector for power system relaying application,” IEEE Trans. Power592Del., vol. 23, no. 1, pp. 79–86, Jan. 2008.593
[31] S. R. Samantaray, A. Samui, and B. C. Babu, “Time-frequency transform-594based islanding detection in distributed generation,” IET Renew. Power595Gener., vol. 5, no. 6, pp. 431–438, Nov. 2011.596
597Soumya R. Mohanty (M’08–SM’13) received the 597Ph.D. degree from Indian Institute of Technology Q2598(IIT), Kharagpur, India. 599
Currently, he is an Assistant Professor with 600the Department of Electrical Engineering, Motilal 601Nehru National Institute of Technology (MNNIT), 602Allahabad, India. His research interests include dig- 603ital signal processing applications in power system 604relaying and power quality, and pattern recognition 605applications to distributed generation-based system. 606
607Nand Kishor (SM’13) received the Ph.D. degree Q3607from Indian Institute of Technology (IIT), Roorkee, 608India. 609
Currently, he is an Associate Professor with 610the Department of Electrical Engineering, Motilal 611Nehru National Institute of Technology (MNNIT), 612Allahabad, India. Since August 2012, he has been 613working as a Marie Curie Experienced Researcher 614(Marie Curie Fellow) with the Electrical Engineering 615Department, Aalto University, Espoo, Finland. His 616research interests include artificial intelligence (AI) 617
applications in power system, distributed generations, wireless sensor network, 618and digital signal processing applications in power system. 619
620Prakash K. Ray (S’10–M’12) received the Ph.D. Q4620degree from Motilal Nehru National Institute of 621Technology (MNNIT), Allahabad, India. 622
Currently, he is working as an Assistant Professor 623with the Department of Electrical and Electronics 624Engineering, International Institute of Information 625Technology (IIIT), Bhubaneswar, India. His research 626interests include distributed generations, renewable 627energy resources, digital signal processing, and soft 628computing applications in power system. 629
630João P. S. Catalão (M’04–SM’12) received the 630M.Sc. degree from the Instituto Superior Técnico Q5631(IST), Lisbon, Portugal, in 2003, and the Ph.D. degree 632and Habilitation for Full Professor (“Agregação”) 633from the University of Beira Interior (UBI), Covilha, 634Portugal, in 2007 and 2013, respectively. 635
Currently, he is a Professor with UBI and 636Researcher at the Instituto de Engenharia de Sistemas 637e Computadores—Investigação e Desenvolvimento 638(INESC-ID), Lisbon, Portugal. He is the Primary 639Coordinator of the EU-funded FP7 project 640
SiNGULAR (“Smart and Sustainable Insular Electricity Grids Under Large- 641Scale Renewable Integration”), a 5.2 million euro project involving 11 industry 642partners. He has published more than 95 journal papers, 175 conference 643proceedings papers and 12 book chapters, with an h-index of 21 (according 644to Google Scholar), having supervised more than 25 post-docs, Ph.D., and 645M.Sc. students. He is the Editor of the book entitled Electric Power Systems: 646Advanced Forecasting Techniques and Optimal Generation Scheduling 647(CRC Press, 2012) and translated into Chinese in January 2014. Currently, 648he is editing another book entitled Smart and Sustainable Power Systems: 649Operations, Planning, and Economics of Insular Electricity Grids (CRC 650Press, forthcoming). His research interests include power system operations 651and planning, hydro and thermal scheduling, wind and price forecasting, 652distributed renewable generation, demand response, and smart grids. 653
Prof. Catalão is the Editor of the IEEE TRANSACTIONS ON SUSTAINABLE 654ENERGY, Editor of the IEEE TRANSACTIONS ON SMART GRID, and 655Associate Editor of the IET Renewable Power Generation. He was the Guest 656Editor-in-Chief for the Special Section on “Real-Time Demand Response” 657of the IEEE TRANSACTIONS ON SMART GRID, published in December 6582012, and he is currently Guest Editor-in-Chief for the Special Section on 659“Reserve and Flexibility for Handling Variability and Uncertainty of Renewable 660Generation” of the IEEE TRANSACTIONS ON SUSTAINABLE ENERGY. He is 661the recipient of the 2011 Scientific Merit Award UBI-FE/Santander Universities 662and the 2012 Scientific Award UTL/Santander Totta. 663