Documente Academic
Documente Profesional
Documente Cultură
i=1
c
i
(t) +r(t) (1)
Often, r(t) is regarded as c
n+1
(t) [20].
The sifting process should be applied with care, for carrying the
process to an extreme could make the resulting IMF a pure fre-
quency modulated signal of constant amplitude. To guarantee that
theIMF components maintainenoughphysical senseof bothampli-
tude and frequency modulations, a criterion should be dened for
the sifting process to stop. This can be accomplished by limiting
size of the standard deviation (SD), it is similar to the Cauchy con-
vergence test, computed from the two consecutive sifting results
as
SD =
T
t=0
_
d
i1
(t) d
i
(t)
2
d
2
i1
(t)
_
(2)
where d
i
(t) is proto (candidate)-mode function.
A typical value for SD can be set between 0.2 and 0.3. As a com-
parison the two Fourier spectra, computed by shifting only ve out
of 1024 points from the same data, can have an equivalent SD of
0.20.3calculatedpoint-by-point. Therefore, SDvalueof 0.20.3for
the sifting procedure is a very rigorous limitation for the difference
between siftings [20].
In Fig. 2, Extrema <3 indicates that the residue becomes a mono-
tonic function fromwhich no more IMFs can be extracted.
2.2. EEMD
One of the major drawbacks of the EMD algorithm is the
appearance of mode mixing. Therefore, the EEMD algorithm was
introduced. The algorithm denes the IMF set for an ensemble of
trials, each one obtained by applying EMD to the signal of inter-
est with added independent identically distributed white noise of
the same standard deviation. The problemof mode mixing can be
overcome by taking into account the properties of white noise [21].
The steps for the EEMD algorithmare as follows:
i. Add a white noise series n(t) to the targeted signal, named x
1
(t)
in the following description, and x
2
(t) =x
1
(t) +n(t).
ii. Decompose the data x
2
(t) with the EMD algorithm.
iii. Repeat Steps (i) and (ii) until the trial numbers, regardless of
added white noise, consistently register the same power. The
new IMF combination C
ij
(t) is achieved when i is the iteration
number and j is the IMF scale.
iv. Estimate the mean (ensemble) of the nal IMF of the decompo-
sitions as the desired output:
EEMD c
j
(t) =
n
i
i=1
c
ij
(t) (3)
where n
i
denotes the trial (ensemble) numbers.
n
=
n
i
(4)
wheren
i
is thenumber of ensemblemembers, is thenal standard
deviation of error, which is dened as the difference between the
input signal and the corresponding IMFs [22].
In practice, the number of ensemble members is often set to 100
and the standard deviation of white noise series is set to 0.1 or 0.2.
2.3. Comparison of EMD and EEMD decomposition for feature
extraction and selection
The extracted IMFs by EMD and EEMD are shown in Fig. 3a and
b. The low-level IMF contains high-frequency components; while
the high-level IMF contains low-frequency components.
In Fig. 3a, the IMFs with EMD algorithm have side effects that
reduce the accuracy of the classication algorithms. However, the
EEMDalgorithmdoes not present side effects andis, therefore, used
in this study. Another criterion is the Pearson product-moment
correlation coefcient (R). It is a measure of the correlation (lin-
ear dependence) between two variables (X and Y), giving a value
between +1 and 1 inclusive.
O. Ozgonenel et al. / Electric Power Systems Research 95 (2013) 192199 195
Fig. 3. (a) Feature vectors (IMFs) for a voltage sag signal with EMD. (b) Feature vectors (IMFs) for a voltage sag signal with EEMD.
An equivalent expression gives the correlation coefcient as the
mean of the products of the standard scores. Based on a sample
of paired data (X
i
, Y
i
), the sample Pearson correlation coefcient is
dened in Eq. (5).
R =
1
n 1
n
i=1
__
X
i
X
_ _
Y
i
Y
__
(5)
where (X
i
X)/
X
,
X and
X
are the standard score, sample mean,
and sample standard deviation of X. Table 1 shows the R values for
EMD and EEMD.
Table 1 also proves that EEMDhas better resolution for the ana-
lyzed PQ events.
2.4. Hilbert-Huang Transformation Method (HHT)
HHT represents the original signal X(t) into the time frequency
domain by combining the empirical mode decomposition with the
Hilbert transform. The advantage of using HHT for feature extrac-
tion procedure can be summarized in Table 2.
EMD can decompose signal into a nite number of n IMFs
(C
j
, j =1,2,3...,n), which extract the energy associated with various
intrinsic time scales and residual r
n
; this step is called the sifting
Table 1
Relation with rst IMFs and current power signal.
Cross-Correlation EMD EEMD
IMFhealthy state 0.9762 1.0000
IMFvoltage sag 0.8819 0.9081
IMFvoltage swell 0.9300 0.9685
IMFvoltage icker 0.9700 0.9950
IMFharmonics (3, 5, 7) 0.7855 0.7861
IMFtransient 0.5685 0.6104
IMFdc component 0.9244 0.9357
IMFEMI 0.5334 0.6429
IMFoutage 0.9313 0.9856
196 O. Ozgonenel et al. / Electric Power Systems Research 95 (2013) 192199
Table 2
Comparison of frequently used mathematical methods for feature extraction [20,21].
Fourier STFT Wavelet HHT
Basis Non-adaptive Non-adaptive Non-adaptive Adaptive
Frequency Convolution: Global Convolution: Regional Convolution: Regional Differentiation: Local
Presentation Energy-frequency Energy-time-frequency Energy-time-frequency Energy-time-frequency
Non-linear No No No Yes
Non-stationary No No Yes Yes
Feature Extraction No Discrete: No Continuous: Yes Discrete: No Continuous: Yes Yes
Theoretical base Theory complete Theory complete Theory complete empirical
process described as in Eq. (6). IMFs, as a class of functions, satisfy
two denitions. In the whole data set, the number of extrema and
the number of zero-crossings must be either equal or differ at most
by one. At any point, the mean value of the envelope dened by the
local maxima and minima is zero.
The Hilbert transformis then applied to each IMF component C
j
:
v
j
(t) =
1
_
+
c
j
()
t
d (6)
where c
j
(t) and v
j
(t) are real part and imaginary part of an analytic
signal z
j
(t):
z
j
(t) = c
j
(t) +jv
j
(t)
z
j
(t) = A
j
(t) exp(jw
j
(t))
(7)
With amplitude, (IA), and phase dened by the expressions:
IA(t) =
_
c
j
(t)
2
+v
j
(t)
2
j
(t) = arctan
_
v
j
(t)
c
j
(t)
_
(8)
Therefore, the instantaneous frequency (IF), was given by:
IF(t) =
d
j
(t)
dt
(9)
Thus, the original data can be expressed in the following form:
x(t) =
n
j=1
A
j
(t) exp
_
i
_
w
j
(t) dt
_
(10)
Figs. 4 and 5 show IA and IF curves for healthy and voltage sag
signals. Fig. 6 shows the IA values for all analyzed PQ events.
3. Singular value decomposition (SVD)
SVD is used for calculating singular values and increases the
accuracy of the classication algorithm. Singular values give more
Fig. 4. Instantaneous amplitude (IA) for healthy and voltage sag signal.
information for identication of the system. SVD is a part of hierar-
chical classication on multiple feature groups in this study.
For any real mn matrix A, there exist orthogonal matrices
U = [u
1
, u
2
, ..., u
m
] R
mm
,
V = [v
1
, v
2
, ..., v
m
] R
nn
,
(11)
such that:
A = UV
T
(12)
where = diag(
1
,
2
, ...,
min(m,n)
)) R
mn
and
1
2
...
min(m,n)
0.
The
i
is the ith singular value of A in non-increasing order and
the vectors u
i
and v
i
are the ith left singular vector and the ith right
singular vector of A for i =min(m,n), respectively [22]. The singular
values of matrix A are unique, while the singular vectors corre-
sponding to distinct singular values are uniquely determined up to
the sign.
Fig. 5. Instantaneous frequency (IF) for healthy and voltage sag signal.
Fig. 6. IA corresponding to rst IMF for all disturbances.
O. Ozgonenel et al. / Electric Power Systems Research 95 (2013) 192199 197
Fig. 7. Singular values for training data set.
Fig. 7 shows singular values for analyzed power signals. Fig. 7
only shows 3 sets of data over 90 training data sets. In Fig. 7, SV1
represents singular values for the healthy signal, SV2 for voltage
sag, SV3 for voltage swell, SV4 for icker, SV5 for harmonics, SV6
for transients, SV7 for DC component, SV8 for EMI and nally SV9
for the voltage interruption signal.
4. Support vector machines (SVMs)
SVMs are a class of learning machines that nd optimal hyper-
planes among different classes of input data or training data in a
high dimensional feature space F, and new test data can be clas-
sied using the individual hyper-planes. The optimal hyper-plane,
found during a training phase, makes the smallest number of train-
ing errors [23]. Fig. 8 illustrates an optimal hyper-plane for two
classes of training data.
Let {x
i
,y
i
}, i =1,2,. . .,N be N training data vectors x
i
with class
label y
i
. Given an input vector x, an SVM constructs a classier of
the form:
f (x) = sign
_
N
i=1
a
i
y
i
K(x
i
, x) +b
_
(13)
where {a
i
} are non-negative Lagrange multipliers, each of which
corresponds to a training data, b is a bias constant, and K(.,.)
is a kernel satisfying the conditions of Mercers theorem [24,25].
Fig. 8. Optimal hyperplane.
Frequently used kernel functions are the polynomial kernel
K(x
i
, x
j
) = (x
i
.x
j
+1)
d
and Gaussian Radial Basis Function (RBF),
K(x
i
, x
j
) = e
|x
i
x
j
|
2
/2
2
.
The problem of nding the optimal hyperplane is specied by
the following quadratic programming problem:
Minimizing:
W(a) =
N
i=1
a
i
+
1
2
N
i=1
N
j=1
a
i
a
j
y
i
y
j
K(x
i
, x
j
) (14)
subject to:
N
i=1
a
i
y
i
= 0 (15)
0 a
i
C, i = 1, 2, ..., N
The above quadratic programming problemcan be solved with
traditional optimization techniques. The vectors for which a
i
>0
after optimization are called support vectors.
4.1. One-against-all method
For a k-class problem, the one-against-all method constructs k
SVMmodels. The ith SVMis trained with all of the training exam-
ples inthe ithclass withpositive labels andthe others withnegative
labels. The nal output of the one-against-all method is the class
that corresponds to the SVMwith the highest output value. Given
a set of training examples D{(x
1
, y
1
), (x
2
, y
2
), ..., (x
N
, y
N
)}, where
the ith sample x
i
R
n
and y
j
{1, ..., k} is the class of x
j
the ith SVM
solves the following optimization problem [2628] (where c >0 is
a regularization parameter for the tradeoff between model com-
plexity andtraining error, and
i
measures the (absolute) difference
between f (x) = (w
i
)
T
(x) +b
i
and y
i
.):
min
1
2
(w
i
)
T
w
i
+c
N
i=1
i
j
(16)
198 O. Ozgonenel et al. / Electric Power Systems Research 95 (2013) 192199
Table 3
Comparisons of the performances of the one-against-all (OAA) and the one-against-one (OAO) SVMclassiers.
Total number of training/testing segments Classication (SVM) One Against All (OAA)One Against One (OAO)
90/25 HS VSG VSW VF VH VTR VDC EMI VI
HS 32 0 0 0 0 0 0 0 0
VSG 1 22 0 0 0 0 0 0 0
VSW 0 0 22 1 0 0 0 0 0
VF 0 0 0 32 0 0 0 0 0
VH 0 0 0 0 33 0 0 0 0
VTR 0 0 0 0 0 43 0 1 0
VDC 0 0 0 0 0 0 33 0 0
EMI 0 0 0 0 0 1 0 32 0
VI 0 0 0 0 0 0 0 0 22
HS: healthystate, VSG: voltagesag, VSW: voltageswell, VF: voltageicker, VH: voltageharmonics, VTR: voltagetransients, VDC: DCvoltage, EMI: electromagnetic interference.
Table 4
Number of measurement point [2931].
Number of measurement points
Voltage level Load type
Heavy Industry Industry + urban Urban only Total
NB NF NB NF NB NF NB NF
33kV 8 16 17 28 14 16 39 60
154 kV 3 5 10 13 9 9 22 27
380 kV 1 1 4 7 0 0 5 8
Total 12 22 31 48 23 25 66 95
NB: number of buses, NF: number of feeders.
subject to
(w
i
)
T
(x
j
) +b
i
1
i
j
, y
j
= i
(w
i
)
T
(x
j
) +b
i
>1 +
i
j
, y
j
/ = i
j
i
0; j = 1, ..., N
(17)
The decision function of the ith SVMis
f (x) = (w
i
)
T
(x) +b
i
(18)
A point x is in the class that corresponds to the largest value of
the decision functions: the class of
x = argmax
i=1,...,k
(w
i
)
T
(x) +b
i
(19)
A comparative study is carried out in the next section that com-
pares the performances of the OAO (one-against-one) and OAA
(one-against-all) decomposition methods in training the proposed
multiclass SVMs. The results of the application of the proposed
algorithmto the data set are presented in Table 3.
Diagonal elements represent the correctly classied PQ
disturbances while non-diagonal elements represent incorrect
classication results (Table 3).
5. Experimental results and discussions
Toevaluate the performance of the proposedpower qualityclas-
sication algorithm, a total of 115 PQ disturbance data were used.
The PQ signals in each class were divided into 90 training sets and
25 test sets. The polynomial kernel is used as the kernel function
and the parameter d is chosen as 2 in the SVMclassier. The exact
number of the training and testing segments for each PQ event
is shown in Table 2. Testing data were obtained from TUBITAK.
This is an ongoing project and is devoted to taking the nationwide
power quality gures of the Turkish Electricity Transmission Sys-
temthrougheldmeasurements carriedout byapplication-specic
mobile monitoring equipment. In this project, PQ measurements
are being carried out according to IEC 610004-30, except for a
continuous seven-day sampling rate in the selected locations. In
the transmission system, most of the bus bars have more than one
incoming and outgoing feeder. Therefore, the number of measure-
ment points can be taken as the number of feeders for currents and
as the number of bus bars for voltages. The classication of these
points as heavy industry, urban + industrial, and urban sites are
given in Table 4, with respect to voltage level. It is noted that there
are more than 95 incoming and outgoing feeders at 33 measured
transformer substations [2931]. The real-time data are acquired
in 3s, which corresponds to 150 periods.
The most accurate multi classiers (OAA) for each PQ dataset
are indicated by red font. As seen from Table 2, the OAA method
(Error =0%, Precision=100%) is superior to the OAO method
(Error =16%, Precision=84%). Therefore, the OAA decomposition
method is chosen for the multiclass SVMin this study. According to
the results of the test data, the effectiveness of the proposed SVM
algorithmis madesuitablebyincreasingtheRBFkernel parameters
standard deviation (sigma 0.1 to 1). The polynomial kernel is worse
than the RBF kernel due to the CPU time. It has been observed that
additional properties of the feature vectors such as singular values
increase the performance of classication algorithm. The perfor-
mance of SVMalgorithmwith OAA and OAO is calculated as 100%
and 84%, respectively. Nevertheless, the performance of SVMalgo-
rithm decreases dramatically without using singular values of IF
and IA values and is calculated as 80% with OAA and 56% with OAO,
respectively. Table 5 gives a comparison of the different classiers
on the training and testing data.
Table 5
Comparison of performances of the different classiers.
Classier Error Accuracy Time (s)
SVM-Linear 0.4000 0.6000 2.136
SVM-Poly d=2 0.0000 1.000 12.711
SVM-RBF sigma =0.1 0.2800 0.7200 8.208
SVM-RBF sigma =1 0.0000 1.000 10.686
Gauss Mixture Model 0.2555 0.7444 1.493
AdaBoost +Naive Bayes 0.0400 0.9600 3.529
Bagging+Naive Bayes 0.1600 0.8400 2.577
O. Ozgonenel et al. / Electric Power Systems Research 95 (2013) 192199 199
6. Conclusion
In this paper, a different PQ monitoring approach based on
EEMD, HHT, and SVMwas proposed. EEMDwas used for signal con-
ditioning technique and two distinctive features, IA and IF, were
obtained by HHT. Then a series of statistical calculations was per-
formed to form feature matrix. The algorithm is found capable of
classifying many PQ disturbances with a high accuracy.
SVM with radial basis (sigma =1) kernel function was used for
the decision-making unit and presented high accuracy among the
other classiers. The proposed algorithm was applied to two sets
of data, one for generated data on a computer and the other for
the data obtained from TUBITAK. Practical applications showed
that the proposed hybrid algorithm can be used in power quality
monitoring software.
The analysis and results presented in this work clearly showthe
potential capability of the suggestedhybridalgorithminclassifying
the distorted PQ waveforms.
Acknowledgement
Theauthors wouldliketothankTUBITAKfor supplyingreal-time
data.
References
[1] M.D.L. Angrisani, P. Daponte, A. Testa, A measurement method based on the
wavelet transform for power quality analysis, IEEE Transactions on Power
Delivery 13 (4) (1988) 990998.
[2] D. Granados-Lieberman, R.J. Romero-Troncoso, R.A. Osornio-Rios, A. Garcia-
Perez, E. Cabal-Yepez, Techniques and methodologies for power quality
analysis and disturbances classication in power systems: a review, IET Gen-
eration, Transmission & Distribution 5 (4) (2011) 519529.
[3] C. Muscas, Power quality monitoring in modern electric distribution
systems, IEEE Instrumentation and Measurement Magazine 13 (2010)
1927.
[4] P.J. Loughlin, J.W. Pitton, E. Les, Atlas, Proper time-frequency energy distribu-
tions and the Heisenberg uncertainty principle, in: Proceedings of The IEEE-SP
International Symposium, Victoria, BC, Canada, 1992, pp. 151154.
[5] I. Daubechies, Ten Lectures on Wavelets, SIAM, Philadelphia, PA, 1992.
[6] S. Mallat, A theory for multiresolution signal decomposition: the
wavelet representation, IEEE Transactions on Pattern Analysis 11 (1989)
674693.
[7] R.R. Coifman, M.V. Wickerhauser, Entropy-based algorithms for best basis
selection, IEEE Transactions on Information Theory 38 (1992) 713718.
[8] P. Kailasapathi, D. Sivakumar, Methods to analyze power quality disturbances,
European Journal of Scientic Research 47 (1) (2010) 616.
[9] P.K. Dash, R.K. Jena, M.M.A. Salama, Power quality monitoring using an inte-
grated Fourier linear combiner and fuzzy expert system, International Journal
of Electrical Power 21 (1999) 497506.
[10] S. Suja, Jovitha Jerome, Pattern recognition of power signal disturbances using
S Transform and TT Transform, International Journal of Electrical Power 32
(2010) 3753.
[11] P.K. Dash, S.K. Panda, A.C. Liew, B. Mishra, R.K. Jena, Anewapproachtomonitor-
ing electric power quality, Electric Power Systems Research 46 (1998) 1120.
[12] M. Oleskovicz, D.V. Coury, O.D. Felho, W.F. Usida, A. driano, A.F.M. Carneiro,
Leandro R.S. Pires, Power quality analysis applying a hybrid methodology with
wavelet transforms and neural Networks, Electric Power Systems Research 31
(2009) 206212.
[13] S.A. Soliman, M.E. El-Hawary, Measurement of power systems voltage and
icker levels for power quality analysis: a static LAV state estimation based
algorithm, International Journal of Electrical Power 22 (2000) 447450.
[14] S.A. Soliman, A.H. Mantaway, M.E. El-Hawary, Simulated annealing optimiza-
tion algorithm for power systems quality analysis, International Journal of
Electrical Power 26 (2004) 3136.
[15] M. Aiello, A. Cataliotti, S. Nuccio, A Chirp-Z transform-based synchronizer for
power systemmeasurements, IEEE Transactions on Instrumentation and Mea-
surement 54 (2005) 10251032.
[16] G. Zheng, Shi Mei-Xiang, D. Liu, J. Yao, MiaoZhu-Mei, Power quality disturbance
classication based on rule-based and wavelet-multi-resolution decomposi-
tion, Proc. 2002 Int. Conf. Mach. Learn. Cybernetics 4 (2002) 21372141.
[17] M. Uyar, S. Yildirim, M. Tunay Gencoglu, An effective wavelet-based feature
extraction method for classication of power quality disturbance signals, Elec-
tric Power Systems Research 78 (2008) 17471755.
[18] S. Kaewarsa, K. Attakitmongcol, T. Kulworawanichpong, Recognition of power
quality events by using multiwavelet based neural Networks, Electric Power
Systems Research 30 (2008) 254260.
[19] T. Jayasree, D. Devaraj, R. Sukanesh, Power quality disturbance classica-
tion using Hilbert transform and RBF Networks, Neurocomputing 73 (2010)
14511456.
[20] N.E. Huang, Z. Shen, S.R. Long, M.C. Wu, H.H. Shih, Q. Zheng, N.C. Yen, C.C. Tung,
H.H. Liu, The empirical mode decomposition and the Hilbert spectrumfor non-
linear and non-stationary time series analysis, Proceedings of the Royal Society
A-Mathematical Physical and Engineering Sciences 454 (1998) 903995.
[21] Z. Wu, N.E. Huang, Ensemble empirical mode decomposition: a noise-assisted
data analysis method, Advances in Adaptive Data Analysis 1 (2009) 141.
[22] G.H. Golub, C.F.V. Loan, Matrix Computations, The Johns Hopkins University
Press, Baltimore, MD, 1996.
[23] C. Cortes, V. Vapnik, Support-vector networks, Machine Learning 20 (1995)
273297.
[24] V.N. Vapnik, Statistical Learning Theory, Wiley, NewYork, 1998.
[25] Chih-Wei Hsu, Chih-Jen Lin, A comparison of methods for multiclass support
vector machines, IEEE Transactions onNeural Networks 13 (2) (2002) 415425.
[26] J.C. Platt, N. Cristianini, J. Shawe-Taylor, Large margin DAGs for multi-class
classication, Advances in Neural Information Processing Systems 12 (2000)
547553.
[27] C.W. Hsu, C.J. Lin, A comparison of methods for multi-class support vector
machines, IEEE Transactions on Neural Networks 13 (2002) 415425.
[28] R. Rifkin, A. Klautau, In defense of one-vs-all classication, Journal of Machine
Learning Research 5 (2004) 101141.
[29] E. Ozdemirci, Y. Akkaya, et al., MobilemonitoringsystemtotakePQsnapshots of
Turkishelectricity transmissionsystem, in: InstrumentationandMeasurement
Technology Conference Proceedings, British Columbia, Canada, 2008, pp. 16.
[30] T. Demirci, A. Kalaycioglu, et al., National power quality monitoring network
and assessment center for the Turkish electricity transmission system: recent
developments, in: Signal Processing, Communication and Applications Confer-
ence, Antalya, Turkey, 2008, pp. 14.
[31] http://www.guckalitesi.gen.tr/en/about/scope.php (last accessed 22.07.2012).