www.elsevier.

com/locate/ijepes

Power quality monitoring using an integrated Fourier linear combiner and

fuzzy expert system
P.K. Dash a,*, R.K. Jena a, M.M.A. Salama b
a
Centre for Intelligent Systems, Regional Engineering College, Rourkela 769008, India
b
Department of Electrical Engineering, University of Waterloo, Waterloo, Ontario, Canada

Abstract
The paper presents a new approach for the classification of transient disturbance waveforms in a power system by using a Fourier linear
combiner and a fuzzy expert system. The measured voltage or current waveforms at a distribution bus are passed through a Fourier linear
combiner block to provide peak or root mean square (RMS) amplitude and phase of the fundamental component at every sampling instant.
The peak or RMS amplitude and computed slope of the waveforms are then passed on to a diagnostic module that computes the truth value of
the signal combination and determines the class to which the waveform belongs. Computer simulated tests are carried out using emtp
programs to obtain the disturbance waveform classification with the help of a new hybrid approach which is much simpler than the recently
postulated neural network and wavelet based techniques. The classification is found to be robust and yields accurate results in most cases with
the least amount of computational burden. q 1999 Elsevier Science Ltd. All rights reserved.
Keywords: Fourier linear combiner; Fuzzy expert system; Disturbance waveform classification

1. Introduction from the measured voltage or current samples and the

classification of the waveforms. In addition, the power
Power supply quality issues and the resulting problems quality monitoring device displays the captured waveforms
are the consequences of the increasing use of solid state when certain thresholds are exceeded. These waveforms
switching devices, non-linear and power electronically exhibit certain distinguishing characteristics and can be
switched loads, unbalanced power systems, lighting classified to belong to a certain waveform class like voltage
controls, computer and data processing equipments as sag, voltage swell, outage, voltage impulse, or normal etc.
well as industrial plant rectifiers and inverters. These elec- The most important characteristic of the classification
tronic type loads cause quasi-static harmonic dynamic system is the accuracy and robustness in the presence of
voltage distortions, inrush, pulse type current phenomena noise and harmonics in the data collected from the trans-
with excessive harmonics and high distortion. Further cyclic mission and distribution networks. Artificial neural
and acyclic loads with temporal variations, or sudden start- networks (ANN) have attracted a great deal of attention
ing of large induction motors can cause low frequency because of their pattern recognition capability, parallel
voltage waveform modulation producing a major power computational architecture, associate memory, and the
quality problem termed as voltage flicker. ability to handle noisy data. However, the convergence
Power quality problems can cause system equipment speed, robustness and accuracy of ANN base methods
malfunction, voltage flickers, computer data loss and depend heavily on the choice of the architecture of the
memory malfunction of sensitive loads such as computer, network, the weight adaptation algorithm and the amount
PLC (Programmable Logic Controller) controls, protection of noise in the data.
and relaying equipment as well as erratic operation of The application of ANN for the classification of low and
electronic controls. high impedance faults [1–4], magnetizing inrush [5] and
Power supply quality monitoring involves the estimation power quality assessment [6] has resulted in significant
of voltage or current peak or root mean square (RMS) values success. Both multilayered feed-forward and time delay
neural networks [7] architectures have been used in classi-
fying transient power quality disturbances like voltage sag,
* Corresponding author. Tel.: 1 91-661-576056; fax: 1 91-661-
572926. voltage swell, outage, voltage surge etc. The success rate in
E-mail address: [email protected] (P.K. Dash) these classifications varies from 72 to 93%, and the number
498 P.K. Dash et al. / Electrical Power and Energy Systems 21 (1999) 497–506

Fig. 1. Block diagram for power quality monitoring.

of training cycles and computational overheads are very combiner is in the form of an adaline [9], which has an
large. Another powerful technique presented in the recent input sequence, and a desired response–signal sequence,
time to assess the power quality disturbance type is the and weight parameters. It uses an adaptive algorithm
wavelet transform [8]. The wavelet transform results in a based on a least mean p-power error criterion to produce
robust classification at the cost of a large computational fast convergence and noise rejection unlike the earlier algo-
overhead as several levels of wavelets in time and frequency rithm [9]. The fuzzy expert system [10] uses a rule base to
frame have to be considered to yield the proper class of the classify the disturbance waveform from the estimated
disturbance pattern. values of amplitude, slope, and distortion factors if any.
This paper, therefore, presents a new approach using a Several waveforms having swell, sag impulse, outage,
Fourier linear combiner and a fuzzy rule-based expert harmonic distortion, or frequency excursion etc. embedded
system to classify the disturbances. The Fourier linear in random noise are tested using this new hybrid estimator
combiner is used to estimate the phase, amplitude, and and fuzzy expert system. The robust and accurate classifica-
rate of change of the voltage waveform of the power system tion of disturbance waveforms validate the efficacy of this
under various operating conditions. The Fourier linear new approach.
 P.K. Dash et al. / Electrical Power and Energy Systems 21 (1999) 497–506 499

Fig. 2. Membership function for amplitude and slope.

2. Monitoring and waveform classification strategy The approach chosen for this purpose consists of a Four-
ier linear combiner, a preprocessor and a fuzzy diagnostic
This section describes the development of a hybrid expert module. Fig. 1 shows how these modules are interrelated.
neural system designed to improve the knowledge of the The raw data in an actual system is to be captured by using a
power systems engineer in pursuit of an accurate diagnosis signal conditioner, a data acquisition interface and an
of power system operating problems such as voltage sag, analog-to-digital (A/D) conversion kit installed in a PC.
voltage swell, faults, harmonics etc. described below: This data comprises voltage and current waveforms of a
disturbed power system. A Fourier linear combiner module
• Impulse: The waveform in this class are described as
with an adaptive LMS algorithm is then used to estimate the
high-frequency transients and occur due to capacitor
amplitude, phase and THD of the captured waveforms.
switching, load start-up and lighting etc.
• Voltage sag: The waveform in this class are character-
ized by short-term decrease of the voltage magnitude and
are measured on a cycle-to-cycle basis. System faults and 3. Fourier linear combiner
motor start-ups etc. cause this kind of voltage problem.
The majority of voltage sags have a magnitude of around The voltage or current signal of a power network is
80% and a duration of 4–10 cycles. expressed in the discrete form such as
• Harmonic distortion: Voltage and current signals
become distorted due to the harmonic penetration into X
N
the power network and voltage and current THDs exceed y…k† ˆ s…k† 1 v…k† ˆ …ai cos vi k 1 bi sin vi k† 1 v…k† …1†
iˆ1
5%, which is quite significant. This class of distortion is
due to the non-linear loads and adjustable speed drives
where v i is the frequency of the ith component and in this
etc.
case v i (the fundamental frequency of the power system
• Voltage swell: These waveforms are characterized by
signal) is known a priori and N is the order of the harmonics
short-term increase in line voltage caused primarily by
in the signal. In the above formulation, v(k) is the additive
over-excitation, load unbalancing and capacitive loads.
white Gaussian noise with zero mean and variance s2v which
• Outage: An outage is an absence of usable power at some
has no correlation with the signal s(k). A decaying dc
point of the power network, and the waveforms that can
component can also be added to the signal model given in
be characterized as such fall into this class. These are
(1).
caused by system faults and opening of circuit breakers.
The coefficients ai and bi of the above signal, corrupted by
After choosing the transient disturbance waveforms to be noise v(k) are obtained by minimizing the error e(k) between
taken up for classification, the next step is to select suitable ^
the desired signal y(k) and the estimated signal y…k†
input features and the method to extract them. Any success-
ful classification would depend on the ability to accurately ^
e…k† ˆ y…k† 2 y…k†
extract the relevant features from the captured waveform
samples in the presence of noise and harmonics. Also the where
diagnostic module should be easy to implement in real-time
X
N
without much computational overhead and should yield ^ ˆ
y…k† …a^ i cos vi k 1 b^ i sin vi k† …2†
accurate classifications. iˆ1
500 P.K. Dash et al. / Electrical Power and Energy Systems 21 (1999) 497–506

where

xi …k† ˆ ‰cos vi …k† sin vi …k†ŠT …6†

W^ i …k† ˆ ‰a^ i …k†b^ i …k†ŠT

and a i is the step size of parameter for the ith frequency

component and sgn(.) is the sign function.
For the value of p greater than 5, the simulations exhibit
performance deterioration of the algorithm for tracking
power system sinusoids corrupted with noise. Also for p ˆ
1 and 2, the performance of the standard LMS algorithm is
obtained. For p taking on values 3 or 4, the algorithm gener-
ates significantly better discrete Fourier transform estimates
and hence a value of p ˆ 3 is chosen in this paper. Thus for
Fig. 3. Membership function for THD. p ˆ 3, the parameter W is updated as

performance index J(k) W^ i …k 1 1† ˆ Wi …k† 1 ai e2 …k† sign{e…k†…xi …k††}=‰li

J…k† ˆ E‰ue…k†up Š: …3† 1 xTi …k†xi …k†Š: …7†

The step size a i will be selected to be large when there is

and E is an expectation operator.
significant divergence between the actual signal and the
Using the steepest descent algorithm, the parameter W is
computed signal. For small divergence, however, the step
calculated as
size a i is selected to be small. A suitable value of a i lies
between 0.2 and 1.6. The adaptive variation of the step size
a ep21 …k†xi …k†
W^ i …k 1 1† ˆ W^ i 1 i a i is given by
l 1 xTi …k†xi …4†
ai …k† ˆ ai …k 2 1† 1 gEi2 …k†
i ˆ 1; 2; …; N for even p and …8†
Ei …k† ˆ bEi …k 2 1† 1 …1 2 b†ei …k†·ei …k 2 1†:

a sign…e…k††ep21 …k†xi …k† In general, the power system disturbance waveforms

W^ i …k 1 1† ˆ W^ i 1 i
l 1 xTi …k†xi …5†
contain odd and even harmonics, and a decaying dc offset
during outages. The presence of the decaying dc offset
i ˆ 1; 2; …; N for odd p; makes it very difficult to track the fundamental component.

Fig. 4. Comparison of performance of conventional and modified LMS algorithm.

 P.K. Dash et al. / Electrical Power and Energy Systems 21 (1999) 497–506 501

Fig. 5. Amplitude and phase tracking for sudden changes in amplitude and phase of fundamental component in the presence of decaying dc.

The signal embedded with decaying dc can be easily tracked The decrement ratio of I0 in two consecutive samples is
after extracting the dc component as given below. b ˆ I0 …m 1 1†=I0 …m† ˆ e2Ts =ldc …10†
Let the time varying signal y(t) be represented as
where Ts is the sampling time.
F…t† ˆ I0 e2t=ldc 1 Im1 sin…wt 1 u1 † 1 Im3 sin…wt 1 u3 † 1 … On the other hand, a small mathematical manipulation
…9† shows that
F…m† 1 F…m1N† ˆ I0 bm
dc …1 1 bdc †
N
…11†
where I0 is the amplitude of dc component and l dc the decay
factor. where N is half of the number of samples per cycle of

Fig. 6. Comparison of tracking of a fault fundamental current using conventional and new modified LMS algorithm.
502 P.K. Dash et al. / Electrical Power and Energy Systems 21 (1999) 497–506

Fig. 7. (i) Variation of alfa in conventional LMS algorithm; and (ii) variation of alfa in modified LMS algorithm.

fundamental waveform and m the present sample b dc and computed using Eq. (14). Fig. 2 shows the membership
l dc are calculated as function for amplitude A and slope S of the disturbance
Fm11 1 Fm1N11 ts waveform. The normalized slope is calculated as S(K) ˆ
bdc ˆ ldc ˆ …12† [A(K) 2 A(K 2 1)] /Smax, Smax is the maximum slope and k
Fm 1 Fm1N ln…bdc †
the sampling instant. In computing the membership func-
tions, the values of a1, a2, b1, b2 and c for the fuzzy set ALN
I0 ˆ ‰F…m† 1 F…m1N† Š=‰bm
dc …1 1 bdc †Š:
N
…13† are chosen as a1 ˆ 0.0, a2 ˆ 0.1, b1 ˆ 2.0, b2 ˆ 3.0, c ˆ 0.3.
From the above calculation, it is clear that with a minimum For the fuzzy set SN, the above constants are: as a1 ˆ 0.25,
of (N 1 4)/2 samples the dc component can be filtered out. a2 ˆ 0.9, b1 ˆ 4.0, b2 ˆ 8.0, c ˆ 0.1. If the transient
disturbance contains harmonics, the total harmonic distor-
tion (THD) can be fuzzified using a membership function of
4. Fuzzy expert system for classification the form

After obtaining the peak or RMS value of the disturbed 1

m…THD† ˆ   …15†
power system signal, a fuzzy logic based expert systems THD 2 5 0:2
11
used to classify the type of disturbance. As the power 50
system data is uncertain and the disturbance classification
is a pattern recognition problem, the fuzzy expert system is where THD is expressed in percentage. From the equation,
well suited to this kind of problem. The fuzzy expert system it is quite evident that upto 5% of THD in a power network
is designed using the knowledge about the disturbances is tolerated and beyond 5%, the membership grade of THD
occurring on a distribution bus. is 1. Fig. 3 shows the membership function of the total
For classifying the power system transient disturbances, harmonic distortion.
three fuzzy sets designated as SN, SZ and SP are chosen for The fuzzified inputs are inferred to a fuzzy rule base,
the absolute value of the slope (normalized) or change in the which is used to characterize the relations between fuzzy
amplitude of the peak value of the waveform. The member- inputs and fuzzy outputs. In this study, the fuzzy rule base of
ship function for computing the absolute value of the slope S the fuzzy expert system is fixed as shown below.
is obtained using a bell shaped function as
Rule 1 If A(K) ˆ ASP and S(K) ˆ SP then W ˆ Swell
1 Rule 2 If A(K) ˆ ALP and S(K) ˆ SP then W ˆ Surge
m…x† ˆ   for x , a1
x 2 a1 b Rule 3 If A(K) ˆ ASN and S(K) ˆ SN then W ˆ Sag
11 …14†
c Rule 4 If A(K) ˆ AZ and S(K) ˆ SZ then W ˆ Normal
Rule 5 If A(K) ˆ ALN and S(K) ˆ SN then W ˆ Surge
m…x† ˆ 1 for a1 , x , a2 Rule 6 If A(K) ˆ ASP and S(K) ˆ SZ then W ˆ Swell
and Rule 7 If A(K) ˆ ALN and S(K) ˆ SN then W ˆ Surge
Rule 8 If A(K) ˆ ALP and S(K) ˆ SZ then W ˆ Sag
m…x† ˆ 
1 Rule 9 If A(K) ˆ ALN and S(K) ˆ SZ then W ˆ Outage
 for x . a2 :
2a2 b2 Rule 10 If A(K) ˆ ALN and S(K) ˆ SP then W ˆ Outage
x
c
In the above, W stands for the disturbance waveform.
In a similar way, five fuzzy sets designated as ALN, ASN, Although the power system disturbance belongs to five
AZ, ASP and ALP are chosen for the peak absolute ampli- categories like voltage sag, swell, outage, surge and normal,
tude A of the waveform. The membership functions are the harmonic distortion can be present in each of them.
 P.K. Dash et al. / Electrical Power and Energy Systems 21 (1999) 497–506 503

Fig. 8. Output of the fuzzy-module for the composite signal (noise-free case).
504 P.K. Dash et al. / Electrical Power and Energy Systems 21 (1999) 497–506

Fig. 9. System configuration of the model used for testing the fuzzy disturbance classifier.

Hence, two more rules are added to the fuzzy rule base to transformer switching and starting of large induction
classify the distortion as motors.
Rule 11 If m(THD) , 1 then W ˆ Normal
Rule 12 If m(THD) . 1 then W ˆ Distorted 5. Simulation results
Unlike conventional fuzzy inferencing procedures, the
Computer simulated waveforms for various transient
inferencing is done using the maximum product rule of
disturbances of a power system are generated using matlab.
fuzzy inference. Assuming the above five categories of the
The sampling rate of 16 based on a 50 Hz waveform is used
disturbance outlined above, the output (m0) from the fuzzy
for testing the effectiveness of the new algorithm in classi-
expert system module is obtained as
fying disturbance waveforms. A SGN function is used for
m0 …k† ˆ m1 V m2 V m3 V m4 V m5 ˆ max…m1 ; m2 ; m3 ; m4 ; m5 † updating the weight vector of the neural estimator which is
…16† initialised using a set of random weights. The value of initial
parameter is chosen and the limits a max and a min are main-
where V stands for OR operator and m1, m2, m3, m4, m5 are the tained as 1.2 and 0.6, respectively. The value of b and g are
firing strengths of the rules for each category of distur- chosen as b ˆ 0.8 and g ˆ .01.
bances.
An uncertainty index l is incorporated to the computa- 5.1. Monitoring of disturbance waveforms
tion process to get the final value of the output m0, and which
is designated as m0F. m0F is related to m0 as For testing the efficacy of the new algorithm, the follow-
ing power system signal corrupted with noise and decaying
m0F …k† ˆ lm0 …k† 1 …1 2 l†m0 …k 2 1†: …17† dc component is used for simulation:
The uncertainty index l is used to compensate the discre- y…t† ˆ 1:5 sin…wt 1 p=6† 1 0:3 sin…3wt 1 p=10†
pancy between the actual value and the observed value. This
is due to the time lag introduced by the transducer used for 1:02 sin5 wt 1 0:3 exp…212t† 1 :03 rand…t†
measurements of power system disturbance waveforms.
Further, the severity problem is realised when the magni- for t , :05312 s
tude of the voltage phaser changes is accompanied by a and
change in the phase angle as is observed in the case of
y…t† ˆ 3:5 sin…wt 1 p=3† 1 :3 sin…3wt 1 p=10† 1 :0 sin…5wt

1 p=4† 1 :03 rand…t†

for t . :05312 s:

From the above description of signals it can be observed that

the signal has a sudden jump at time t greater than.05 s
approximately. Also the decaying dc component is present
in the signal at time t , .0535. The signal is corrupted by a
random white noise of variance s ˆ .03 and zero mean. The
tracking efficiencies of the standard LMS algorithm and the
least mean p-power error criterion for the above mentioned
signal are shown in Fig. 4. In case of the conventional LMS
algorithm, the decaying dc component is filtered out using
the principle described in Section 4. From the results, it can
Fig. 10. Converter bus voltage (phase-a). be seen that the p-power error criterion yields a better
 P.K. Dash et al. / Electrical Power and Energy Systems 21 (1999) 497–506 505

Fig. 11. Output of fuzzy-expert module for the test waveform.

tracking capability of the fundamental component of the track the signal component in the presence of decaying dc
power system disturbance signal when dc offset is present. and random noise and the result is not as good as the one
The tracking time is less than a cycle based on the funda- obtained using the new approach. The results for the
mental frequency of the signal. This new algorithm does not Kalman filter are not presented in the paper.
require any filtering technique as needed for the conven- Another interesting case of practical value is the tracking
tional LMS approach in the presence of random noise. of the fundamental component of the power system distur-
Further, the new algorithm performs extremely well (Fig. bance waveform during a fault.
5) during sudden changes in the amplitude and phase angle An emtp software is used to simulate a synchronous
of the fundamental component, which is the one required for generator feeding a transmission line subjected to a
monitoring and classification. Kalman filter is also used to single-line-to-ground fault at a certain distance from the
506 P.K. Dash et al. / Electrical Power and Energy Systems 21 (1999) 497–506

sending end. The fault current waveform is shown in Fig. 6. 6. Conclusion

This figure also presents the amplitude and phase angle of
the fundamental component of the faulted current wave- The paper presents a new approach to the assessment of
form. The generator frequency also varied during the power quality using an integrated Fourier linear combiner
fault, which was estimated by another frequency estimation and fuzzy decision support system. The tracking capability
algorithm presented in Ref. [10]. of the Fourier linear combiner is enhanced with an adaptive
From the results presented in the figure, it can be seen that LMS algorithm using a new p-power error criterion. A fuzzy
the mean p-power error criterion produces an excellent logic based reasoning process is used to classify the power
convergence to the true signal value in the presence of quality disturbance waveforms from the computed ampli-
noise, decaying dc components, harmonics and frequency tude and slope of the waveforms obtained from the Fourier
excursion. If the frequency excursion is very small, the linear combiner. The newly modified LMS algorithm tracks
tracking time remains within a cycle. the fundamental component of the signal corrupted with
The conventional LMS algorithm shows large oscilla- noise within a time period less than one cycle. The paper
tions due to frequency changes and the presence of dc offset also presents the variation of the learning rate using the new
and noise. The variations of m for both the conventional p-power error criterion. The fuzzy expert system used for
LMS and the new approach with mean p-power error criter- the classification power network disturbances like voltage
ion are shown in Fig. 7. sag, voltage swell, outage, voltage surge etc. is very simple
to implement in a real time in comparision to the recently
5.2. Classification of disturbance waveforms proposed ANN and wavelet based approaches, which are
computationally very involved. Several computational test
Fig. 8 shows the category of the simulated waveforms
results are also presented in the paper to show the improved
like sag or swell and the corresponding output from the
performance of the integrated Fourier linear combiner and
Fourier linear combiner and the integrated fuzzy and Fourier
fuzzy logic approach.
linear combiner diagnostic systems. The training of the
monitoring system for different class of waveform is essen-
tial for tuning the parameters of the various modules. From References
the figure, it is observed that each category of waveform is
successfully classified as the output from the model, shows a [1] Lubkeman DL, Fallon CD, Girgis AA. Unsupervised learning strate-
100% truth value of the particular class that suddenly rises gies for the detection and classification of transient phenomena on
from 0 to 100% in most cases in comparison to the normal electric power distribution systems, Proceedings of First International
Forum on Applications of Neural Networks to Power Systems, IEEE
waveform. 91 THO 374–9, 1991, pp. 107–111.
After ascertaining the efficacy of the proposed fuzzy [2] Kim CJ, Russell BD. Classification of faults and switching methodol-
expert system in classifying the simulated disturbances, ogy. IEEE Transactions on Power Delivery 1989;4(3):1631–1637.
a practical case is taken for a detailed study. Fig. 9 [3] Ebron S, Lubkeman D, White M. Neural network power distribution
shows a generator supplying a power network which feeders. IEEE Transactions on Power Delivery 1990;5(2):905–914.
[4] Sultan AF, Swift GW, Fedirchu DJ. Detection of high impedance
comprises a short-transmission line section and resistive arcing faults using a multi-layer perceptron. IEEE Transactions on
and constant impedance loads. The resistive load Power Delivery 1992;7(4):1871–1877.
consumes 3.3 MW of power and is supplied from a [5] Perez LG, Flechsig AJ, Meador JL, Obradovic Z. Training an artificial
rectifier connected to the load bus as shown in Fig. 9. neural network to discriminate between magnetizing inrush and inter-
The initial starting current of the rectifier is 18 A and is nal faults, IEEE Winter Power Meeting, 1993, paper no.WM037-2,
PWRD.
started at t ˆ 0.043 s. An outage at the generator end is [6] Daniels R. Power quality monitoring using neural networks, Proceed-
initiated at t ˆ 0.49 s and persists for 2 cycles based on ings of the First International Forum on Application of Neural
50 Hz waveform. Variable static capacitors are installed Networks to Power Systems, IEEE 91 THO 374–9, 1991, pp. 195–197.
at the load bus to improve the load power factor as [7] Ghosh AK, Lubkeman DL. The classification of power system distur-
shown in the figure. The converter voltage waveform bance waveforms using a neural network approach. IEEE Transac-
tions on Power Delivery 1995;10(1):109–115.
(one phase only) is presented in Fig. 10 using an [8] Dash PK, Swain DP, Liew AC, Rahman S. An adaptive linear combi-
emtp software package. Fig. 11 presents the classifica- ner for on-line tracking of power system harmonics, 1996, papers
tion results showing clearly the type of disturbance IEEE 96 WM 181-8 PWRS.
class (voltage sag or voltage swell etc.) from changes [9] Dash PK, Liew AC, Rahman S. Fuzzy neural network and fuzzy
in membership grades. For a particular type of distur- expert system for load forecasting. Proceedings IEE (London),
Generation, Transmission and Distribution 1996;143(1):106–115.
bance, the membership grade will become unity and for [10] Dash PK, Swain DP, Routray A, Liew AC. An adaptive neural
other non-occuring types, membership grades will be very network approach for the estimation of power system frequency.
small. Electric Power Research 1997;41:203–210.