IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 18, NO.

2, APRIL 2003 595

Power Quality Disturbance Data Compression,

Detection, and Classification Using Integrated
Spline Wavelet and S -Transform
P. K. Dash, B. K. Panigrahi, D. K. Sahoo, and G. Panda

Abstract—In this paper, power quality transient data are Spline wavelets [4] stand apart in the general theory of wavelet
compressed and stored for analysis and classification purposes. transform. Their construction starts with the specification of
From the compressed data set, original data are reconstructed the underlying multiresolution function spaces. Thus, spline
and then analyzed using a modified wavelet transform known as
-transform. Compression techniques using splines are performed wavelets can be characterized explicitly; this is in contrast with
through signal decomposition, thresholding of wavelet transform most other constructions where the scaling function is specified
coefficients, and signal reconstruction. Finally, we present com- indirectly through a two-scale relation.
pression results using splines and examine the application of The main advantage of an explicit construction is that
splines compression in power quality monitoring to mitigate one does not need to worry about the delicate issues of the
against data-communication and data-storage problems. Since
-transform has better time frequency and localization property, convergence of the iterated filter bank. After reconstructing
power quality disturbances are detected and then classified in a the power quality data using spline wavelets, -transform
superior way than the recently used wavelet transform. (a modified form of wavelet transform) [5] is used to detect
Index Terms—Data compression, pattern classification, power and classify the type of disturbance. The -transform uses
quality, spline wavelets, -transforms. a scalable Gaussian window that dilates and translates. The
-transform is unique in that it provides frequency-dependent
resolution while maintaining a direct relationship with the
I. INTRODUCTION
Fourier spectrum. Several examples are chosen to use spline

P OWER-SUPPLY quality issues and resulting problems

are consequences of the increasing use of solid-state
switching devices, nonlinear and power electronically switched
wavelets for data compression and reconstruction and -trans-
form for detection and classification of disturbance type. The
-transform contours resemble the nature of the disturbance
loads, unbalanced power systems, lighting controls, computer unlike the wavelet transform [6], [7] and, hence, directly pro-
and data-processing equipment, as well as industrial plant vides a visual classification for low-frequency short-duration
rectifiers and inverters. These electronic-type loads cause disturbances and high-frequency transients as well. A simple
quasistatic harmonic dynamic voltage distortions, inrush, automated classification procedure is also described in the
pulse-type current phenomena with exciting harmonics, and paper for different types of power quality problems.
high distortion.
Power quality problems can cause system equipment mal-
function; computer data loss and memory malfunction of sen- II. BASIC PRINCIPLES OF SPLINE WAVELETS
sitive loads such as computer, programmable logic controller Among all wavelet families, splines [8] appear to have the
(PLC) controls, protection, and relaying equipment; and erratic best approximation property in the sense that the magnitude of
operation of electronic controls. the constant is minimum. This means that, in the asymptotic
Therefore, it is necessary to monitor these disturbances. regime where the error is small, we can apply a coarser sam-
Since transient disturbances occur in the order of microseconds, pling step if we use splines as opposed to other basis functions
a single captured event recorded using this type of monitoring (or wavelet) with the same order . The potential reduction in
system produces megabytes of data. As a result, the volume sampling density can be quite significant. For instance, it is ob-
of the recorded data increases significantly, necessitating the served that splines at half the resolution could provide a better
development of a technique to compress the data volume approximation than Daubechies wavelets.
[1], [2]. It is the purpose of this paper to introduce spline for Splines are piecewise polynomials with pieces that are
compressing power quality disturbance signals. Splines-based smoothly connected together. The joining points of the polyno-
compression schemes have been used especially in image mials are called knots. For a spline of degree , each segment
processing. In 1946, Schoenberg [3] showed how one could is a polynomial of degree , which would suggest that we need
use splines to interpolate equally spaced samples of a function. coefficients to describe each piece.
Splines are uniquely characterized in terms of a B-spline ex-
Manuscript received September 29, 2000; revised April 4, 2002 pansion as
P. K. Dash is with Multimedia University, Cyberjaya, Selangor, Malaysia.
B. K. Panigrahi, D. K. Sahoo, and G. Panda are with the National Institute of
Technology, Rourkela, India. (1)
Digital Object Identifier 10.1109/TPWRD.2002.803824

0885-8977/03$17.00 © 2003 IEEE

596 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 18, NO. 2, APRIL 2003

Fig. 1. Centered B-spline of degree 0 to 3. (a)

B-spline Coefficients. The B-splines of degree 0 to 3 are

shown in Fig. 1.
The following closed-form representation of the cubic
B-spline is obtained as

(b)
Fig. 2. (a) First-scale signal decomposition into c (n) into c (n) and d (n).
(2) (b) First-scale signal reconstruction from c (n) and d (n) into c (n).

. , respectively; contains a detailed version of

Here, we consider the center-symmetric B-spline of degree , and higher-frequency components (i.e., sharp edges transition)
for compression and reconstruction of power quality than a smoothed signal . Mathematically this can be
data. expressed as
(6)
(3)
(7)

A. Optimal Time-Frequency Localization This is shown in Fig. 2(a). In each level, the length of the de-
If denotes the B-spline wavelet of degree , than we composed signal is reduced to half of the original signal. Similar
have the following approximate formula (Cosine-modulated higher-order decomposition is performed.
Gaussian) The reconstruction of can be performed by a pyramidal
scheme similar to decomposition Fig. 2(b) shows the recon-
struction of from decomposed signals and .
Mathematically, this can be expressed as
(4) (8)

with Higher-scale reconstruction is performed in a similar manner.

(5)
IV. DATA COMPRESSION USING SPLINE
The quality of this Gabor approximation improves rapidly with
increasing ; for 3 the approximation error is less than Data compression is another area where splines can be used
3%. Implication is that there are spline wavelet bases that can for their symmetry, short support, and excellent approximation
be optimally localized in time and frequency. In other words, properties. Splines offers advantages, especially in higher di-
we can get as close as we wish to the time limit and frequency mensions where the overhead with respect to wavelets becomes
localization limit specified by the uncertainty principle. negligible. This is a powerful idea for the implementation of it-
erative algorithms using a coarse-to-fine iteration strategy. The
III. SIGNAL DECOMPOSITION AND RECONSTRUCTION approach is reasonably fast because it makes the best use of its
iterations: good starting conditions with an optimizer that is ex-
In splines, multiresolution processing in time domain is con- tremely efficient near the optimum.
verted to time-scale domain which is known as forward trans- In splines compression method, a threshold value is fixed to
form. This process of transform of time domain to time-scale do- discard some coefficients and to eliminate noise in the process
main is known as forward transform (“Wavelet transform”) and of decomposition. Accordingly, some coefficients are discarded
technically known as signal decomposition because a particular which results in reduction of the original file size by three to six
signal is decomposed into several other signals with a different times. Thresholding process is explained below.
level of resolution. The reverse transform from time-scale do-
main to time domain is known as inverse Wavelet transform or
V. THRESHOLDING OF DECOMPOSED COEFFICIENTS
signal reconstruction.
A recorded digitized time signal (which is a sampled Thresholding in wavelets are broadly classified as
version of the original signal) is decomposed into its detailed 1) soft thresholding;
and smoothed signals, using filters and 2) hard thresholding.
DASH et al.: POWER QUALITY DISTURBANCE DATA COMPRESSION USING INTEGRATED SPLINE WAVELET 597

Soft thresholding is adopted for denoising and hard thresh-

olding is adopted for data compression. Particularly abnormal
spikes of noises get reduced when soft thresholding is applied
because large coefficients are made zero by this. We also find
that in case of hard thresholding, small coefficients are made
zero (i.e., coefficients below threshold) are set to zero. This
leads to the compression of the number of coefficients. It should
also be noted that denoising itself leads to compression, but to a
(a)
lesser degree. Particularly in the case of the power system signal,
this scheme is effective since we have only a single frequency
extract.
The threshold is based on the absolute maximum value of
the wavelet transform coefficients at the associated scale
which can be expressed as

where is the threshold value, lies in the range 0 1

and is the wavelet transform coefficients at the associated
scales and is given by
for
for (8a)
After thresholding, it is clear that that are smaller than
are discarded and those that are larger are stored. These spline
transform coefficients and their temporal positions in the signal
are stored together. Consider an actual disturbance signal
with its wavelet transform at scale 1 , where has
1600 sample points, has 800 sample points. It is clear
that is half of . Fig. 3(b) shows the plot of and
its first-level decomposition signal . From Fig. 3(b)(iii),
it is clear that contains the disturbance signal compared
with . Taking 0.9, 0.1
0.0459. With this threshold value, the number of nonzero ele-
ments in signal is only three. Both magnitude and posi-
tion are to be stored and, therefore, there are 6 (2 3) sample
points to be stored. Performing “one-scale compression,” the
most smoothed signal is , as shown in Fig. 3(b)(iii). It has
only 800 sample points. Hence, the compressed data now has
806 (3 2 800) sample points, compared to 1600 sample
points of the original signal.

A. Compression Ratio
It is the ratio of original file size to the compressed file size.
For the example just presented, the compression ratio is 5.0633.
By choosing a different threshold value, a different compression
ratio can be obtained
original file size
Compression ratio
compressed file size

B. Energy Retained by the Compressed Signal Fig. 3. (a) Schematic distribution network. (b) One-scale spline construction
technique (i) original signal, (ii) distorted signal at scale 1, (iii) smoothed signal
It is the ratio of vector norm of the compressed signal to the at scale 1, (iv) distorted signal: scale 1 after thresholding, and (v) reconstructed
vector norm of original signal. signal.

Energy retained VI. NORMALIZED MEAN SQUARE ERROR

vector norm of compressed data
vector norm of original data
In order to enhance the compression ratio, we now decompose NMSE
the impulsive transient shown in Fig. 4(a) into three time scales. in the order of to (9)
598 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 18, NO. 2, APRIL 2003

where is the time period; the choice of unity for the constant in
eq. (12) makes the Gaussian window in eq. (10) the narrowest in
the time domain. Now the continuous -transform is modified
to

(13)
The signal can be written in discrete form as , where
varies from 0 to 1 and is known as discrete time series
of the signal . Discrete Fourier transform of the time series
can be expressed as

(14)

where 0, 1, 1 and the inverse Discrete Fourier

transform is

(15)

-transform in discrete case is the projection of the vector de-

fined by the time series onto a spanning set of vectors. S
transform of the discrete time series is

Fig. 4. (a)–(e) Spline decomposition and reconstruction for impulse, (f)

S -transform of reconstructed signal. (16)

where is the original signal, and is the reconstructed VIII. RESULTS

signal. To obtain a high-quality reconstructed signal, normalized
mean square error (NMSE) should be small and of the order of The proposed techniques are used to analyze the power
10 . From the example mentioned in Fig. 3(b), NMSE is found quality disturbance transients in a power distribution system
as 2.5088 10 . simulated using a power system block set. A sample distri-
bution system is shown in Fig. 3(a). The sampling rate for
VII. -TRANSFORM the collection of power quality transient data is taken equal
to 10 kHz. Different distorted data are generated by varying
-transform is an extension of the idea of the Gabor trans- different parameters in the power distribution system. The
form and wavelet transform, and is based on a moving and scal- following case studies are presented in this paper: the result is
able localizing Gaussian window. -transform falls within the obtained for a disturbed signal in three scales taking different
broad range of multiresolution spectral analysis and it is impor- compression ratios. The threshold selection is based on the
tant to note that it is used here specifically for a translatable and absolute maximum value. Here, a three-scale spline wavelet
scalable Gaussian window where the standard deviation is an in- decomposition scheme is applied to the distorted signal.
verse function of the frequency. With the introduction of dilation For all three signal thresholds , and 10% of abso-
parameter into the localizing Gaussian function defined as lute maximum value of spline coefficients at the corresponding
scale. Then from the compressed data, the original signal is re-
(10)
constructed using spline reconstruction filter with a mean square
where is the standard deviation. The Gaussian function is error of the order of 10 to 10 . Splines have better com-
chosen as the window because it is the minimum uncertainty pression capability but they fail to classify. Hence, we have per-
window. The multiresolution -transform is defined by formed -transform on compressed data for classification pur-
pose.
(11)
A. Data Compression Using Spline Wavelets
The primary purpose of the dilation (or scaling) parameter Case 1: Fig. 4(a) shows an impulsive transient at 22 ms,
is to increase the “width” of the window function for Fig. 4(b)–(d) shows a three-scale decomposition using spline
lower frequency and vice versa, and is controlled by selecting a and Fig. 4(e) shows the reconstructed signal. Here we obtained
specific functional dependency of with the frequency . a compression ratio of 7.1429. -transform is then applied to
and (12) the reconstructed coefficients for detection and classification of
DASH et al.: POWER QUALITY DISTURBANCE DATA COMPRESSION USING INTEGRATED SPLINE WAVELET 599

Fig. 5. (a)–(e) Spline decomposition and reconstruction for frequency

distortion and (f) S -transform output.
Fig. 6. (a)–(e) Spline decomposition and reconstruction for voltage
a power quality problem. From -transform contours, the im- interruption and (f) S -transform output.
pulsive transient at 22 ms is clearly detected and localized. Fur-
ther, the time-frequency contours show the nature of disturbance Fig. 7(b)–(e) shows a three-scale decomposition and recon-
waveform (in this case an impulse) through visual inspection. struction of the distorted signal mentioned before. Using the
The nature of this disturbance can be classified by obtaining a same threshold value, we obtained a compression ratio of 6.45.
few simple features such as standard deviation and amplitude Fig. 7(f) shows -transform plot for the reconstructed signal.
factors (described in Section VIII-B). From Fig. 7(f), it is found that -transform provides clear
Case 2: As a second example, we use a variable frequency information on the appearance of the disturbance and its type.
signal as shown in Fig. 5(a), where, at 20 ms, time frequency The reduction in contour magnitude is clearly analogous to the
changes from 48 to 50 Hz and then it changes to 52 Hz at description of voltage sag (reduction of RMS or peak value)
60.1 ms. This distorted signal is compressed, and then recon- magnitude plot.
structed using spline for power quality analysis purpose. Here
we obtained a compression ratio of 6.8. -transform is applied
B. Classification of Power Quality Disturbances
on this reconstructed data for classification purposes. The fre-
quency variation is clearly detected and localized from -trans- From the results presented in Figs. 4–7, it can be seen that the
form contours. -transform provides visual classification of the power quality
Case 3: A momentary interruption can be seen as a mo- disturbance waveforms. However, a few simple features can be
mentary loss of voltage on a power system. These disturbances derived from the -transform output to provide an automated
describe a drop of 90 to 100% of the rated system voltage lasting power quality recognition system from the compressed wave-
for 0.5 cycle to 1 min. Fig. 6(a) shows a signal having interrup- form data. As the computation of -transform involves the use
tion from 20 to 60 ms. Using the same threshold as before, we of FFT routines, the output matrix contains informations re-
find a compression ratio of 6.58. A three-scale decomposition garding amplitude-time, and amplitude-frequency variations of
and reconstruction using spline is shown in Fig. 6(b)–(e). From the disturbance signal. These informations can be used to obtain
-transform plot, the interruption mentioned before in supply the following features.
is clearly detected and identified. Further, the -transform con- 1) standard deviation of the lowest -transform con-
tours show a collapse in magnitude similar to the RMS (or peak) tour (level 1) that is above the normalized fundamental
value plot of voltage magnitude during a voltage interruption. frequency.
Case 4: To enhance the compression capability of spline 2) standard deviation of the amplitude versus time
and classification capability of -transform, harmonic distorted curve obtained from the rows of the -matrix that is above
signal with 50% sag from 20 to 60 ms is shown in Fig. 7(a). the normalized fundamental frequency.
600 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 18, NO. 2, APRIL 2003

30 dB). However, in the presence of more noise and harmonics,

this accuracy deteriorated to less than 90%. Some typical cases
of compressed waveforms presented in this paper in Figs. 4–7
were classified accurately using the simple rule-based approach.

IX. CONCLUSION
The paper has proposed a new approach for the compression
of large-quantity data obtained for analysis of power quality
violations in distribution networks. The results presented
in the paper clearly reveal that the signal can be perfectly
reconstructed from the compressed data in order to save energy
using splines. Due to the superior time-frequency resolution
property, the -transform contours display significantly im-
proved patterns to detect, localize, and visually classify the
types of disturbances. However, for automated classification
of power quality disturbances, the -transform output matrix
is searched to provide a few simple features which when used
in a rule-based system yields the disturbance class. Hence,
sources of such disturbances can be identified and controlled to
improve electric power quality.

REFERENCES
[1] R. P. Binghom, D. Kreises, and S. Santoso, “Advances in data reduction
techniques for power quality instrumentation,” in Proc. Third European
Power Quality Conf., Power Quality, Bremon, Germany, Nov. 1995.
[2] S. Santoso, E. J. Powers, and W. M. Grady, “Power quality disturbance
data compression using wavelet transform methods,” IEEE Trans. Power
Delivery, vol. 12, pp. 1250–1257, July 1997.
[3] I. J. Schoenberg, “Contribution to the problem of approximation of
Fig. 7. (a)–(e) Spline outputs for voltage sag and (f) S -transform output. equidistant data by analytic function,” Quart. Appl. Math., vol. 4, pp.
45–99, 112–141, 1946.
[4] M. Unser, A. Aldroubi, and M. Eden, “Fast B-spline transforms for con-
3) tinuous image representation and interpolation,” IEEE Trans. Patterns
where denotes the rows of the -matrix output with Anal. Machine Intell., vol. 13, pp. 277–285, Mar. 1991.
[5] R. G. Stockwell, L. Monsinha, and R. P. Lowe, “Localization of the
disturbanc, and denotes the corresponding output without complex spectrum: The S -transform,” IEEE Trans. Signal Processing,
disturbance. vol. 44, pp. 998–1001, Apr. 1996.
For disturbances such as voltage sag, voltage swell, and in- [6] S. Santoso, E. J. Powers, W. M. Grady, and A. C. Parsons, “Power quality
disturbance waveform recognition using wavelet based neural classi-
terruption, it is observed that and are less than 0.1 (in- fier—Part 1: Theoretical foundation,” IEEE Trans. Power Delivery, vol.
cluding the cases with harmonics and noise) whereas for tran- 15, pp. 222–228, Jan. 2000.
sients, there are oscillatory, impulsive types, and notches [7] S. Santoso, W. M. Grady, E. J. Powers, J. Lamoree, and S. C. Bhatt,
“Characterization of distribution power quality events with Fourier and
0.1. However, is greater than 0.05 for all oscillatory transients wavelet transforms,” IEEE Trans. Power Delivery, vol. 15, pp. 247–253,
and is less than 0.05 for notches and impulsive transients. The Jan. 2000.
other important feature clearly distinguishes low- or power- [8] M. Unser, “Approximation power of bio-orthogonal wavelet expan-
sions,” IEEE Trans. Signal Processing, vol. 44, pp. 519–527, Mar.
frequency disturbances such as voltage sag, swell, and interrup- 1996.
tions from higher-frequency transients. For example, for 50%
voltage sag, 0.5 and for 20%, 0.2. In a similar
way, it is found from tests that for 50% voltage swell, 0.5. P. K. Dash is currently Professor in the faculty of Engineering, Multimedia
In case of voltage interruption, the factor 0.9. For tran- University, Malaysia. He was previously a Chairman Centre for intelligence
systems, National Institute of Technology, Rourkela, India. His interests are
sients, the factor is small and is less than 0.1. However, for in fuzzy logic and ANN applications toward power system and power quality
single- or multiple-type notches, is small and is less than 0.1. studies.
In the previouse examples, the factor quantifies the magnitude
of voltage sag, swell, or interruption. From these examples, it is B. K. Panigrahi is pursuing the Ph.D. degree in power quality studies.
He is currently a Lecturer in electrical engineering at the University College
quite clear that a simple rule-based system will be adequate to of Engineering, Burla, India.
provide a power quality waveform classification system. How-
ever, due to some overlapping of features for steady-state and D. K. Sahoo received the M.Engg. degree from R.E.C. Rourkela, India.
transient disturbances, intelligent systems, such as fuzzy logic Currently, he is a Research Engineer.
and neural networks, can be used for providing more accurate
and robust classification. As a test case, out of 100 disturbances G. Panda is a Professor of the Department of Applied Electronics and Instru-
mentation Engineering, National Institute of Technology, Rourkela, India. He
of steady and transient types, more than 97% were classified ac- has published a large number of papers in signal and image processing as well
curately in the presence of noise levels of less than 3% (SNR as in power quality studies.