You are on page 1of 11

www.ietdl.

org

Published in IET Generation, Transmission & Distribution


Received on 17th April 2008
Revised on 18th October 2008
doi: 10.1049/iet-gtd:20080190

ISSN 1751-8687

Optimal feature selection for classication of


power quality disturbances using wavelet
packet-based fuzzy k-nearest neighbour
algorithm
B.K. Panigrahi V.R. Pandi
Department of Electrical Engineering, IIT, Delhi 110016, India
E-mail: bkpanigrahi@ee.iitd.ac.in

Abstract: The authors present an automatic classication of different power quality (PQ) disturbances using
wavelet packet transform (WPT) and fuzzy k-nearest neighbour (FkNN) based classier. The training data
samples are generated using parametric models of the PQ disturbances. The features are extracted using
some of the statistical measures on the WPT coefcients of the disturbance signal when decomposed upto
the fourth level. These features are given to the fuzzy k-NN for effective classication. The genetic algorithm-
based feature vector selection is done to ensure good classication accuracy by selecting 16 better features
from all 96 features generated from the WPT coefcients. The necessity of selecting the best feature is to
remove the redundant or irrelevant features, which may reduce the performance of the classication. It also
reduces the computation time since it uses only 16 features instead of 96 features. The experimental analysis
for the validation of the proposed algorithm is carried out in two stages. At the rst stage, the sample data
set is generated by varying the parameters in models in regular intervals and the proposed algorithm is
applied to select the best features to obtain high accuracy. In the second stage, a new data set is generated
by choosing the parameter values, which are not used in the rst case and used to test the accuracy of the
classier with the same selected features as in stage one. The noisy and practical signals are also considered
for the classication process to show the effectiveness of the proposed method.

1 Introduction failure or mal-operation of the electrical equipments. Hence,


the performance and life expectancy from the electrical
Recently, power quality (PQ) has become a signicant issue equipments may differ from the actual design values.
for both utilities and customers. PQ issues [1] and the
resulting problems are the consequences of the increasing In order to improve electric PQ, the sources and causes of
use of solid state switching devices, nonlinear and power such disturbances must be known before appropriate
electronically switched loads, unbalanced power systems, mitigating action can be taken. However, in order to
lighting controls, computer and data processing equipment determine the causes and sources of disturbances, one must
as well as industrial plant rectiers and inverters. These have the ability to detect and localise these disturbances. In
electronic type loads cause quasi-static harmonic dynamic most of the recent reported literatures, researchers have
voltage distortions, inrush, pulse type current phenomenon attempted to use efcient and appropriate signal processing
with excessive harmonics and high distortion. A PQ problem and computational intelligence techniques for the
usually involves a variation in the electric service voltage or detection, localisation and classication of PQ disturbances.
current, such as voltage dips and uctuations, momentary Wavelet transform (WT) [2 4] is widely used for the PQ
interruptions, harmonics and oscillatory transients causing assessment. Some of the reported works [5 10] at an early

296 IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 3, pp. 296 306
& The Institution of Engineering and Technology 2009 doi: 10.1049/iet-gtd:20080190
www.ietdl.org

stage explores the applicability of WT in analysing non- are carried out using fuzzy k-nearest neighbour (FkNN)
stationary signals in power systems. Wavelet-based online algorithm [28]. To improve the classication accuracy and to
disturbances detection for PQ applications are discussed in nd the optimal feature set, genetic algorithms (GAs) are
[11, 12]. Some of the recent literatures [13, 14] further used. The GAs are search algorithm based on the concept of
emphasise the wide application of WT for the PQ analysis. natural selection [29, 30] and are used for many nonlinear
Although WT has been extensively used for detection of optimisation problems. GAs are exploratory search methods
PQ disturbances, but in many of the cases the effect of based on mechanics of selection and survival of the ttest.
electrical noise is not adequately considered. A de-noising
scheme [15] is proposed for enhancing wavelet-based PQ
monitoring system. In [16] detection and classication of 2 Wavelet packet feature
PQ disturbances is presented where WT is combined with extraction
short-time correlation transform.
Analysing the PQ data using WPT involves three steps,
selection of the type of mother wavelet, the order of mother
In order to identify the type of disturbances present in the wavelet and the level of decomposition. A number of wavelet
power signal more effectively, several authors have presented families with unique properties have been proposed in the
different methodologies based on combination of WT and signal processing literature, but the most appropriate family is
articial neural network (ANN). Using the properties of generally application-dependant. After a thorough literature
WT, the features of the disturbance signal are extracted at review, it was found that the fourth-order Daubechies wavelet
different resolution levels and then the features are trained (dB4) yields the best performance for analysing PQ
with different ANN algorithms. By this method, it is disturbances. Thus, in the proposed work the fourth-order
possible to extract important information from a Daubechies wavelet was chosen as the mother wavelet for the
disturbances signal and determine what type of analysis. The selection of the level of decomposition depends
disturbances has caused a PQ problem to occur. In [17, on the desired frequency resolution. To obtain the best
18], the application of neural networks to perform frequency resolution, the fourth level decomposition was
automated waveform identication was discussed in detail. chosen as illustrated in Fig. 1.
In [19, 20], the powerful wavelet multi-resolution analysis
method is applied for signal decomposition and the PQ The wavelet packet approach becoming popular for extracting
events are classied based on standard deviation and root the useful features, because it is like a generalisation of wavelet
mean square value. In [21], the author has demonstrated decomposition, which gives a richer range of possibilities for
the classication of seven types of PQ events by using signal analysis. In wavelet analysis, a signal is split into
wavelets and probabilistic neural network (PNN). Energy approximation and detail coefcients by passing it through
distribution at 13 decomposition levels of wavelet and time low-pass and high-pass lters respectively in the rst level of
duration of disturbances are taken as features and these 14 decomposition. The approximation is then split into a
features are applied to PNN for classication. In [22], a self second-level approximation and detail, and the process is
organising learning array system based on WT for repeated. In wavelet packet analysis, a signal is split into an
classication of seven types of PQ problems has been approximation and a detail in the rst level. Then in the
proposed. Here, 11 decomposition levels of wavelets are used second level both detail and approximation are decomposed
for the classication. Also, 22 types of wavelets are tested to further which forms binary tree like structure as shown in
get effective classication by a particular type of wavelet. Fig. 1. This process is recursively applied on both detail and
approximation upto level k, which depends on frequency
Wavelet packet transform (WPT) has been widely used for resolution. For each level j [ k, there are 2j number of nodes
machine fault diagnosis [2325], biomedical signal analysis available after wavelet packet decomposition. In general
[26] and also to the PQ analysis [27]. In this paper, we have notation W pj denotes the pth node occurring at the jth level.
tried to decompose the PQ disturbances signals up to the
fourth level and extracted the useful statistical features for each In kth
k1
level wavelet packet decomposition there are more
level. The classication of different types of PQ disturbances than 22 different ways to encode the signal. For instance,

Figure 1 Tree structure of the wavelet packet decomposition at level 4

IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 3, pp. 296 306 297
doi: 10.1049/iet-gtd:20080190 & The Institution of Engineering and Technology 2009
www.ietdl.org

Figure 2 Implementation of the wavelet packet transforms

wavelet packet analysis allows the representation of original decomposition level. Here N is the number of coefcients
signal W 00 in the form of W 01 W 22 W 63 W 14
4 W4 .
15
in each decomposed data. The term E(Dij 2 mi)k in
This is an example of a representation that is not possible skewness and kurtosis calculation is the expected value of
with ordinary wavelet analysis. Choosing one out of all the quantity, also called as kth moment about the mean.
these possible encodings presents an interesting problem. Thus in the present case the decomposition of the signal
In this wavelet packet analysis, we use the Shannon upto fourth level yields 16 different nodes corresponding to
entropy-based criterion to select the most suitable different frequency sub-bands. Selecting the above
decomposition of a given signal. The implementation of mentioned features for each node we will obtain the feature
complete wavelet packet analysis upto the second level vector of length 96. The feature vector is denoted by
using HP and LP lter banks is shown in Fig. 2.
Feature [ED1 ED2    EDl s1 s2    sl m1 m2 . . . ml
KRT1 KRT2    KRTl SK1 SK2    SKl
3 Feature extraction ENT1 ENT2    ENTl ] (7)
The coefcient Dij at fourth decomposition level is used to
extract the features. Statistical features like energy, standard
deviation, mean, kurtosis, skewness and Shannon entropy
of the decomposition coefcients Dij are calculated by 4 Feature selection by GA
using the following equations GAs are stochastic search algorithm by which we can explore
the search space to obtain optimal solution. They operate
N 
X 
 2 on string structures called chromosomes, typically a
Energy EDi Dij  i 1, 2 . . . l (1)
j1
concatenated list of binary digits representing the encoding
!1=2 of the control parameters of a given problem. It works with
N  2
1 X a population of individuals and the decisions taken are
Standard deviation si Dij  mi , based on probabilistic rules. The PQ disturbance signal is
N  1 j1
decomposed upto fourth level using wavelet packet
i 1, 2 . . . l (2)

1X N
Mean mi D i 1, 2 . . . l (3)
N j1 ij
 4
E Dij  mi
Kurtosis KRTi i 1, 2 . . . l (4)
s4i
 3
E Dij  mi
Skewness SKi i 1, 2 . . . l (5)
s3i
X
N   Figure 3 Feature selection by GA
Entropy ENTi  Dij2 log Dij2 i 1, 2 . . . l
j1
a Chromosome structure for selecting the optimal features at
(6) fourth level
b Two point crossover operation
where i 1, 2, . . . , l is the number of nodes at the fourth c Mutation operation

298 IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 3, pp. 296 306
& The Institution of Engineering and Technology 2009 doi: 10.1049/iet-gtd:20080190
www.ietdl.org

Figure 4 Power quality waveforms for


a Voltage sag
b Voltage swell
c Harmonic distortion
d Flicker
e Oscillatory transient
f Sag with harmonics
g Swell with harmonics
h Interruption
i Spike
j Notch

decomposition. The feature vector is formed by applying may have an overlapping nature between two classes, which
some of the statistical measures like energy, entropy, leads to misclassication. In this work, we have tried to
standard deviation and so on, the resulting WPT nd the optimal feature vector by applying GA that leads
coefcients. The classication process using all 96 feature to good classication accuracy. The GA encoding schema
vectors may lead to less accuracy, since some of the features of chromosome is made up of a binary string, which

IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 3, pp. 296 306 299
doi: 10.1049/iet-gtd:20080190 & The Institution of Engineering and Technology 2009
www.ietdl.org

represents the selection of one of the feature from the until the mating pool is lled. The best solution in the
available six features in all 16 nodes. It leads us to have 16 current population may be inserted more than once.
control parameters in the chromosome string with each Similarly, the worst solution in the current population may
parameter represented by 3 bits as shown in Fig. 3a. While have disappeared. In this manner, any solution in the
decoding of these chromosome structures, we employed current population may have zero, one, two or many copies
discrete type of decoding schema as mentioned in [29]. in the new population.
During the search process, the tness of each chromosome
is evaluated using FkNN classication accuracy so that our The crossover and mutation operations are performed after
objective, become maximisation of classication accuracy. reproduction. The crossover operator is mainly responsible
for the global search property of the GA. There are many
Generally, GAs consists of three basic operations, namely ways to do the crossover operation, like single point
reproduction, crossover and mutation [30]. Reproduction crossover, two point crossover and uniform crossover. In
comprises forming a new population, usually with the same this work two point crossover is considered. Two parent
total number of chromosomes, by selecting from members strings are selected from the population at random. In the
of the current population following a particular scheme. The two-point crossover the crossover points are generated
higher the tness, the more likely it is that the chromosome randomly between string lengths. The information between
will be selected for the next generation. There are several these two crossover points is exchanged and thus the two
strategies for selecting the individuals, example roulette- child strings are created. This is shown in Fig. 3b. Two
wheel selection, ranking methods and tournament selection. child strings are then generated from the parent strings in
Here we use tournament selection. In tournament selection, the process of crossover by complementing the child strings
the size of the participant in the tournament n is chosen at selected bit positions in order to exchange the already
rst. The n individuals are selected at random from the existing information.
current population and the better solution among the n is
selected and inserted into the new population. Then another Mutation is then applied on some of the strings to introduce
n individual is selected randomly and the better one is new information in the mating pool, with small probability as
inserted into the new population. This procedure is repeated shown in Fig. 3c. The bitwise mutation operator changes

Table 1 Equations and parameter variations for PQ signals

PQ disturbances Model Parameters variation


voltage sag x(t) A(1 2 a(u(t 2 t1) 2u(t 2 t2))) sin(vt) 0.05  a  0.85
t1, t2 , u(t) 1 if t 0, T  t2 2 t1  7T
0 if t , 0
voltage swell x(t) A(1 a(u(t 2 t1) 2u(t 2 t2))) sin(vt) 0.05  a  0.8
t1, t2 , u(t) 1 if t 0, T  t2 2 t1  7T
0 if t , 0
harmonics x(t) A(a1 sin(vt) a2 sin(2vt) a3 sin(3vt) a5 a1 1.0,
sin(5vt) a7 sin(7vt)) a2 a7 (0.0 2 0.3)
icker x(t) A(1 af sin(bfvt)) sin(vt) 0.01  af  0.25
2 Hz  ff  8 Hz
oscillatory x(t) A(sin(vt) b eg (tt1 ) sin(vtr(t 2 t1))) 22  b  2,
transient 50  g  100,
500 Hz  ftr  1500 Hz
sag with x(t) A(12 a(u(t 2 t1) 2u(t 2 t2))) (a1 sin(vt) a2 a1 1.0,
harmonics sin(2vt) a3 sin(3vt) a5 sin(5vt) a7 sin(7vt)) 0.0  a2 , a3, a5 and a7  0.3
0.05  a  0.85
T  t2 2 t1  7T
swell with x(t) A(1 a(u(t 2 t1) 2u(t 2 t2))) (a1 sin(vt) a2 a1 1.0,
harmonics sin(2vt) a3 sin(3vt) a5 sin(5vt) a7 sin(7vt)) 0.0  a2 , a3, a5 and a7  0.3
0.05  a  0.8
T  t2 2 t1  7T
interruption x(t) A(12 a(u(t 2 t1) 2u(t 2 t2))) sin(vt) 0.88  a  1.0
T  t2 2 t1  7T

300 IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 3, pp. 296 306
& The Institution of Engineering and Technology 2009 doi: 10.1049/iet-gtd:20080190
www.ietdl.org

from 1 to 0 and vice versa. The mutation operation helps to keep where the denominator is a normalising factor. The value of
the diversity in the population. In order to maintain better scaling factor m should be chosen in order to have better
solutions during the crossover and mutation process, there is inuence of distance samples. As the value of m approaches
some probability called crossover probability and mutation innity, the result of the classication is same as that of the
probability. Usually, crossover probability is chosen between crisp k-NN algorithm result. Here in this paper, we used
0.6 and 0.9 and mutation probability is taken between 0.001 m 2, so as to represents the Euclidean distance between
and 0.1. In order to maintain the best chromosomes elitism is the samples as used in [28].
performed. It always maintains the best chromosome called
elites in the mating pool.

6 Results and discussion


5 FkNN classier 6.1 Data preparation
Machine learning techniques like support vector machines as The wavelet-based FkNN classier presented in this work is
well as various types of neural networks has been applied to designed to identify ten types of PQ events including voltage
classify the PQ event classication problem. In this paper, sag, voltage swell, momentary interruption, harmonic
we tried to apply the well-known classier called FkNN distortion, voltage icker, oscillatory transient, sag with
algorithm originally proposed by James M. Keller et al. harmonics, swell with harmonics, spike and notch.
[28]. The nearest neighbour algorithm works to nd a Generating the sampled signals of these types from
correlation between unknown samples and training parametric models was advantageous in getting a wide
samples. The training samples are denoted by n range of signals in a controlled manner. The parametric
dimensional numeric attributes. The k-nearest neighbour data set is prepared with a sampling frequency of 3.2 kHz.
(k-NN) algorithm searches the k training samples that are The data length of each class of disturbance is 12 cycles.
adjacent to the unknown sample given for the classication
process. The closeness is measured in terms of Euclidean The typical disturbance waveforms of these kinds are
distances and can be calculated as shown in Fig. 4. The detailed summary of all types of
disturbances along with the variation in parameters and
s models are given in Table 1. Here, the parameter a
X n  2
D X , Y xi  yi (8)
i1

The unknown sample is classied according to the most


occurring class among the kNNs. In a special case if k 1,
it simply classies the unknown sample to its nearest
neighbour. The k values are chosen as odd numbers in the
case of binary classication, so as to avoid the tied votes.
The advantage of this algorithm over the neural-based
classication schemes is that it does not generate the pre-
dened model during training. But it will take more
computation time if the number of training samples is
much higher because each time it has to compute the
distance vector. So it is better to use the k-NN algorithm
when the number of training samples are less. The k-NN
algorithm was modied using fuzzy logic in order to get
better classication. The drawbacks in k-NN like precision
and correctness during tie condition are overcome by
FkNN algorithm. Moreover, this FkNN algorithm is
getting idea from Bayes rule, which provides optimal
solution to the classication problem, that is, it gives the
highest assurance that the classication will be correct.

The membership value of sample x in class i is given by


 2=(m1)

Pk  
u
j1 ij 1=  x  x j
ui (x)  2=(m1)
for all i (9)
Pk   Figure 5 Flowchart for the GA-based feature vector
j1 1=x  xj 
selection

IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 3, pp. 296 306 301
doi: 10.1049/iet-gtd:20080190 & The Institution of Engineering and Technology 2009
www.ietdl.org

Table 2 Optimally selected features by GA

Node Selected feature Node Selected feature Node Selected feature Node Selected feature
0 mean 4 kurtosis 8 energy 12 standard deviation
1 kurtosis 5 standard deviation 9 kurtosis 13 energy
2 skewness 6 entropy 10 energy 14 standard deviation
3 mean 7 kurtosis 11 mean 15 kurtosis

Table 3 Comparison of classication accuracy results

Case Energy Entropy Energy and All (96) GA selected


(16) (16) entropy (32) feature (16)
I. pure signal (10-fold avg) 97.90 97.90 98.08 99.03 99.3857
II. intermediate signal 93.3383 93.7424 95.7962 94.3816 96.2703
III. intermediate signal with different 94.17 94.83 96.06 94.62 97.81
magnitude and same point on wave
IV. intermediate signal with same magnitude 93.45 93.27 95. 74 93.64 96.94
and different point on wave

represents the level of sag or swell in the rst two types of sag with harmonics 170, swell with harmonics 160,
disturbances. The unit step function u(t) in the whole table momentary interruption 52, spike 168 and notch 168)
provides the duration of disturbances present in the pure according to the range of parameter variation in each class.
sine waveform. During the generation of the disturbance We are also interested in generating other data sets called
signal from the parametric model, the value of a and the intermediate signals for testing the algorithm after selecting
position of u(t) has been varied suitably, so that a large the best features by GA. This intermediate data set is
number of signals can be obtained with varying magnitude prepared corresponding to the parameter values that are not
(by changing a) on different points on the wave (by used for generating the rst data set.
changing the parameters t1 and t2) and the duration of the
disturbance (t2 2 t1). The point on the wave is the instant The wavelet packet algorithm is applied to decompose the
on the sinusoid when a disturbance begins and is controlled original signal up to fourth level of decomposition. In this
by the position of the unit step function u(t). As the real
work, we used Shannon entropy-based wavelet packet
PQ disturbance signals may have any point on the wave decomposition that is available in the MATLAB Wavelet
which is beyond control, hence we have generated a variety toolbox [31]. The WPT decomposes the signal into 16 equal
of disturbances having different points on the wave
duration of disturbance and magnitudes. The harmonic
signal consists of a combination of second-, third-, fth-
and seventh-order harmonics. The icker signal is
generated by varying the icker frequency ff and its
amplitude af . Oscillatory transient is denoted by its
amplitude b, frequency ftr and the inverse of the time
constant of decay g. The signal having both sag and swell
along with harmonics are considered as separate classes in
our study. Finally, the momentary interruption with
parameter a is taken for varying the amplitude during
interruption. The remaining two PQ events considered in
this paper are spike and notch, which are of short duration
as compared to sag or swell.

Using the parametric model of each event we generated


totally 1147 different cases (sag 85, swell 80,
Harmonics 64, icker 68, oscillatory transient 132, Figure 6 Classication accuracy for ve types of analysis

302 IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 3, pp. 296 306
& The Institution of Engineering and Technology 2009 doi: 10.1049/iet-gtd:20080190
www.ietdl.org

parts of 100 Hz at the fourth level. The statistical features like the measurements such as nearest neighbour classication [32].
energy, standard deviation, mean, kurtosis, skewness and This vector normalisation has been done to avoid the inuence
entropy corresponding to all the 16 nodes present in the of high-ranging feature vectors over lower ranging ones.
fourth level and is carried out to form the feature vector set. In
this way we obtain 96 features (6  16) corresponding to each
signal. After calculating the features for the complete data
6.2 Results of GA-based selection
covering all types of disturbance, the feature vectors are The complete simulation study was carried out using
normalised between [0, 1] by taking the minimum and MATLAB 7.0 in IBM 3.0 GHz personal computer with
maximum values of the corresponding feature vectors as the 256 MB of RAM. The best feature vector selection is
base. Normalisation is one of the important steps of pre- carried out by GA initially. This is done by combining all
processing of the data before classication by any of the the data sets into a single data set and making it 10-fold,
classication algorithms involving neural networks or distance so that each fold contains samples of all classes. Now the

Table 4 Comparison of classication accuracy results for intermediate signal


Class Type of Energy Entropy Energy and entropy All (96) GA selected feature
number disturbances (16) (16) (32) (16)
1 voltage sag 100.00 100.00 100.00 100.00 97.647
2 voltage swell 97.647 97.647 97.647 94.118 94.118
3 harmonics 95.833 93.75 95.833 100.00 100.00
4 icker 72.222 89.683 84.127 68.254 83.248
5 oscillatory transient 91.406 90.625 92.969 98.438 94.531
6 sag with harmonics 100.00 97.778 100.00 97.222 96.10
7 swell with 92.941 92.941 92.941 94.118 97.059
harmonics
8 interruption 83.333 75.00 94.444 91.667 100.00
9 spike 100.00 100.00 100.00 100.00 100.00
10 notch 100.00 100.00 100.00 100.00 100.00
average 93.3383 93.7424 95.7962 94.3816 96.2703

Table 5 Comparison of classication accuracy results for noisy signal

Class number Type of disturbances 40 dB noise 35 dB noise 30 dB noise 25 dB noise 20 dB noise


1 voltage sag 100.0000 92.7928 93.2830 83.0189 56.3025
2 voltage swell 100.0000 100.0000 95.8506 85.9504 74.0000
3 harmonics 91.3580 83.7838 81.8919 81.8182 85.2272
4 icker 100.0000 100.0000 100.0000 96.1538 98.7500
5 oscillatory transient 97.7143 93.7107 91.8333 96.4497 94.7712
6 sag with harmonics 100.0000 100.0000 99.5516 100.0000 100.0000
7 swell with harmonics 98.5437 100.0000 100.0000 99.5516 100.0000
8 interruption 100.0000 100.0000 100.0000 86.4407 81.5384
9 spike 87.0466 84.9785 86.8113 98.9899 100.0000
10 notch 84.9741 84.4221 86.0226 98.1651 100.0000
average 95.9637 93.9688 93.5244 92.6538 89.0589

IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 3, pp. 296 306 303
doi: 10.1049/iet-gtd:20080190 & The Institution of Engineering and Technology 2009
www.ietdl.org

chromosomes are encoded such that it contains the one will select all the features having the feature set
information of which feature should be selected in each dimension 96, the overall average accuracy further increases.
node as explained previously. The objective of the GA is But the computation time becomes more because of the
maximising the 10-fold average accuracy. The population larger dimension of the feature set. In the rst case, pure
size and number of generation is selected as 20 and 40, signal data set is considered and the 10-fold average of the
respectively. The crossover and mutation probability of GA classication accuracy is presented in Table 3. In this case,
is selected as 0.8 and 0.1, respectively. The owchart for the training as well as testing data set contains the feature
selecting the best feature by GA is shown in Fig. 5. set pertaining to the disturbance signals having different
magnitude, durations and points on the wave. A bar chart
The best features for all the 16 nodes are reported in is provided for the classication accuracy of each individual
Table 2. The classication accuracy obtained by the class in Fig. 6. In the second case, the original data set is
proposed method using the best selected features is considered as training data. The intermediate signals are
reported in Table 3. In this table, for a comparative study generated with different disturbance magnitudes, duration
the classication accuracy using only energy or entropy as a and point on wave values that are different from those used
feature for all the 16 nodes is also reported. It is observed for generating the training signal. The features extracted
that either energy or entropy as a feature for the from these intermediate signals are considered as testing
classication purpose nearly yields the same percentage of data and the classication accuracy is reported in the same
average accuracy. When both energy and entropy are Table 3, case II. The detailed results showing the
selected as features thereby increasing the feature dimension individual class accuracy for these intermediate signals is
to 32, it is observed that the average accuracy increases. If reported in Table 4. However, to show the performance of

Figure 8 Transient signal captured and its detection using


WPT
a Signal
Figure 7 Laboratory setup to acquire practical PQ data b Wavelet packet decomposition level 1 (D1)

Table 6 Comparison of classication accuracy results for practical data

Class number Type of disturbances Training/testing data set Accuracy with GA selected feature (16)
1 voltage sag 50/30 96.67
2 voltage swell 50/30 100.00
3 harmonics 50/30 86.67
4 icker 50/30 100.00
5 oscillatory transient 50/30 93.33
6 sag with harmonics 50/30 100.00
7 swell with harmonics 50/30 100.00
8 interruption 50/30 100.00
9 spike 50/30 93.33
10 notch 50/30 93.33
average 96.33

304 IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 3, pp. 296 306
& The Institution of Engineering and Technology 2009 doi: 10.1049/iet-gtd:20080190
www.ietdl.org

the proposed algorithm, case II is again subdivided into two algorithm is used for automatic classication of the
different cases. In case III, we have considered the disturbances using the feature sets obtained thereby. The
intermediate signals generated with the different wavelet packet algorithm along with the FkNN classier
disturbance magnitudes (a) and the same point on wave as can be used as an efcient tool to recognise the type of
considered in the generation of training data. Case IV power disturbance occurring at a particular point of time.
describes the classication accuracy of the proposed The implementation of GA-based feature selection makes
algorithm with intermediate signals generated with same the classication accuracy better. The proposed algorithm
disturbance magnitudes (as considered in the training data has been tested with the data obtained from the parametric
preparation) but with different points on the wave. It is model of the PQ disturbances, noisy PQ data and the data
observed that the optimal features selected for the 16 nodes set obtained from the laboratory setup.
of the WPT gives a comparatively better result for the
classication of PQ disturbances. Further, it is observed
that the execution time for the classication of 1-fold data
(training: 1000 data and testing: 100 data) is 0.3317 s for
8 Acknowledgments
all 96 features and 0.0633 s for 16 selected features. The authors acknowledge the nancial support of
Department of Science and Technology, Govt. of India for
the research project SR/S3/EECE/46/2007.
6.3 Results of performance under noisy
environment
In an electrical power distribution network, the practical data
9 References
consists of noise; therefore the proposed approach has to be
[1] BOLLEN M.H.J.: Understanding power quality problems:
analysed under noisy environment. Gaussian white noise is
Voltage Sags and Interruptions (IEEE Press, 1999)
widely considered in the research of PQ issues. We have
obtained the noisy signals for all the ten classes having
[2] DAUBECHIES I.: The wavelet transform, time/frequency
different signal to noise ratio (SNR). The features of the
location and signal analysis, IEEE Trans. Inf. Theory, 1990,
noisy signal are extracted using WPT as described in the
36, pp. 961 1005
previous section. Features extracted for noisy data
consisting of SNR of 20, 30 and 40 dB and tested with 20,
[3] MALLAT S.G.: A theory of multi resolution signal
25, 30, 35 and 40 dB noise levels. The classication
decomposition: the wavelet representation, IEEE Trans.
accuracy for this case is reported in Table 5.
Pattern Anal. Mach. Intell., 1989, 11, (7), pp. 674 693

6.4 Results of accuracy for practical data [4] MEYER Y. : Wavelets and operators (Cambridge
University Press, London, UK, 1992)
In this section, we have demonstrated the applicability of the
proposed technique to the practical data. The PQ data are [5] SANTOSO S., POWERS E.J., GRADY W.M., HOFMANN P.: Power
obtained from the laboratory. The laboratory setup is given quality assessment via wavelet transform analysis, IEEE
in Fig. 7. The electrical system consists of a source (400 V), Trans. Power Deliv., 1996, 11, (2), pp. 924 930
which supplies power through a prototype model of a
transmission line. The loads are of linear (static PQ load) [6] SANTOSO S., POWERS E.J. , GRADY W. : Power quality
and nonlinear (converter load) nature. Different PQ disturbance identication using wavelet transformers and
disturbances are simulated by making faults in the line, articial neutral network. The 1996 Int. Conf. Harmonic
additional load switching, capacitor switching and varying and Quality of Power, Las Vegas, NV, USA, 1996,
the ring angle of the converter load. A transient signal pp. 615 618
obtained by capacitor switching is presented in Fig. 8 along
with the detection of this event using WPT. The terminal [7] RIBEIRO P.F.: Wavelet transform: an advanced tool for
voltage waveforms are acquired through voltage sensors analyzing non-stationary harmonic distortions in power
(Lv-20-P). Output of the voltage sensor is limited between systems. The IEEE Int. Conf. Harmonics in Power Systems,
+10 V to provide compatibility with subsequent signal Bologna, Italy, 1994
processing systems. These signals are captured by using
dSPACE 1104 kit and monitored in computer. The [8] RIBEIRO P.F., ROGERS D.A.: Power electronics, power quality
sampling frequency is set as 3.2 kHz. The classication and modern analytical tools: the impact on electrical
result of the practical data is reported in Table 6. engineering education. The 24 Annual Conf. Frontier in
Education, San Jose, CA, USA, 1994, pp. 448 451
7 Conclusion [9] PILLAY P., BHATTACHARJEE A.: Application of wavelets to
This paper presents the application of WPT to extract the model short-term power system disturbances, IEEE Trans.
useful features from the PQ disturbances signals. FkNN Power Syst., 1996, 11, (4), pp. 2031 2037

IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 3, pp. 296 306 305
doi: 10.1049/iet-gtd:20080190 & The Institution of Engineering and Technology 2009
www.ietdl.org

[10] ANGRISANI L., DAPONTE P., DAPUZZO M., TESTA A.: A classication and measurement, IEE Proc., Gener. Transm.
measurement method based on the wavelet transform Distrib., 2002, 149, pp. 310 318
for power quality analysis, IEEE Trans. Power Deliv., 1998,
12, (4), pp. 990 998 [21] GAING Z.L.: Wavelet-based neural network for power
disturbance recognition and classication, IEEE Trans.
[11] KARIMI M., MOKHTARI H., IRAVANI M.R.: Wavelet based online Power Deliv., 2004, 19, (4), pp. 1560 1568
disturbance detection for power quality applications, IEEE
Trans. Power Deliv., 2000, 5, pp. 1212 1220 [22] HE H., STARZYK J.A. : A self-organizing learning array
system for power quality classication based on wavelet
[12] MOKHTARI H., GHARTEMANI M.K., IRAVANI M.R.: Experimental transform, IEEE Trans. Power Deliv., 2006, 21, (1),
performance evaluation of a wavelet-based on-line pp. 286 295
voltage detection method for power quality application,
IEEE Trans. Power Deliv., 2002, 17, pp. 161 172 [23] LIU B.: Selection of wavelet packet basis for rotating
machinery fault diagnosis, J. Sound Vib., 2005, 284,
[13] CHEN S., ZHU H.Y.: Wavelet transform for processing pp. 567 582
power quality disturbances, EURASIP J. Adv. Signal
Process., 2007, (1), pp. 1 20 [24] ZAREI J., POSHTAN J.: Bearing fault detection using
wavelet packet transform of induction motor stator
[14] BOLLEN M.H.J., GU I.Y.H., AXELBERG P.G.V. , STYVAKTAKIS E.: current, Tribol. Int., 2007, 40, pp. 763 769
Classication of underlying causes of power quality
disturbances: deterministic versus statistical methods, [25] LI X., QU L., WEN G., LI C.: Application of wavelet packet
EURASIP J. Adv. Signal Process., 2007, Article ID 79747, 17 analysis for fault detection in electro-mechanical systems
pages, 2007. doi:10.1155/2007/79747 based on torsional vibration measurement, Mech. Syst.
Signal Process., 2003, 17, (6), pp. 1219 1235
[15] YANG H.T., LIAO C.C.: A de-noising scheme for enhancing
wavelet-based power quality monitoring systems, IEEE [26] TING W., GUO-ZHENG Y., BANG-HUA Y., HONG S.: EEG feature
Trans. Power Deliv., 2001, 16, pp. 353 360 extraction based on wavelet packet decomposition for
brain computer interface, Measurement, 2008, 41, (6),
[16] ZANG H., LIU P., MALIK O.P.: Detection and classication pp. 618 625
of power quality disturbances in noisy conditions,
IEE Proc. Gener. Transm. Distrib., 2003, 150, (5), [27] HU G.-S., ZHU F.-F. , REN Z. : Power quality disturbance
pp. 567 572 identication using wavelet packet energy entropy and
weighted support vector machines, Expert Syst. Appl.,
[17] SANTOSO S., POWERS E.J. , GRADY W. , PARSONS A. : Power 2008, 35, (1 2), pp. 143 149
quality disturbance waveform recognition using wavelet-
based neural classier, Part 1: theoretical foundation. [28] KELLER J.M., GRAY M.R., GIVENS J.A. JR.: A fuzzy k-nearest
The 1997 IEEE/PES Winter Meeting, New York, NY, USA, neighbor algorithm, IEEE Trans. Syst. Man Cybern., 1985,
1997 15, (4), pp. 580 585

[18] SANTOSO S., POWERS E.J., GRADY W., PARSONS A.: Power quality [29] GOLDBERG D.E. : Genetic algorithms in search
disturbance waveform recognition using wavelet-based optimization and machine learning (Addison-Wesley,
neural classier, Part 2: application. The 1997 IEEE/PES Reading, MA, 1989)
Winter Meeting, New York, NY, USA, 1997
[30] BAKIRTZIS A.G., BISKAS P.N., ZOUMAS C.E., PETRIDIS V.: Optimal
[19] GAOUDA A.M. , SALAMA M.M.A. , SULTAN M.K. , CHIKHANI A.Y. : power ow by enhanced genetic algorithm, IEEE Trans.
Power quality detection and classication using wavelet- Power Syst., 2002, 17, (2), pp. 229 236
multiresolution signal decomposition, IEEE Trans. Power
Deliv., 1999, 14, (4), pp. 1469 1476 [31] www.mathworks.com

[20] GAOUDA A.M. , KANOUN S.H., SALAMA M.M.A., CHIKHANI A.Y.: [32] HAN J., KAMBER M. : Data mining: concepts and
Wavelet base signal processing for disturbance techniques (Morgan Kaufmann Publishers, 2001)

306 IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 3, pp. 296 306
& The Institution of Engineering and Technology 2009 doi: 10.1049/iet-gtd:20080190

You might also like