Documente Academic
Documente Profesional
Documente Cultură
Abstract—This paper presents an expert system that is able to system are described. Special attention is given to the segmenta-
classify different types of power system events to the underlying tion scheme. The expert system is tested using recordings from
causes (i.e., events) and offer useful information in terms of a medium voltage network (33, 11, and 6.6 kV) obtained over a
power quality. The expert system uses the voltage waveforms and
distinguishes the different types of voltage dips (fault-induced, two-month period. The results are presented and discussed.
transformer saturation, induction motor starting) as well as
interruptions (nonfault, fault-induced). A method for event-based
classification is used, where a segmentation algorithm is first II. BACKGROUND
applied to divide waveforms into several possible events. The
expert system is tested using real measurements and the results The voltage events that are considered by the expert system
show that the system enables fast and accurate analysis of data can be divided into the following classes: fault-induced events,
from power quality monitors. transformer events, induction motor events, interruptions, and
Index Terms—Expert systems, Kalman filtering, power quality, step-change events. The first three classes and their character-
power system monitoring, voltage dips (sags). istics are presented in [3]. Their common feature is that they
cause a temporary decrease in voltage magnitude. Classifica-
tion is possible based on the following characteristics of voltage
I. INTRODUCTION
magnitude.
(a) (b)
(c) (d)
(e) (f)
(g) (h)
(i)
Fig. 1. Fundamental voltage magnitude and segmentation results (shadowed parts) for (a) energizing, (b) nonfault interruption, (c) fault interruption,
(d) transformer saturation, (e) induction motor starting, (f) step change, (g) transformer saturation followed by protection, (h) single stage dip due to fault, and
(i) multistage dip due to fault.
4) transformer saturation; for the purpose of segmentation but also for obtaining the fun-
5) induction motor starting; damental frequency magnitude and the harmonics of the voltage
6) step change; waveforms. The order of the filters is set to 20 to accommodate
7) transformer saturation followed by protection; the harmonics that are caused by events like transformer satura-
8) single stage dip due to fault; tion or arcing.
9) multistage dip due to fault. It is common that there are more than one of the above-men-
tioned events in one single recording. Therefore, the parts of the
signal associated with different events must be found and ana-
III. SEGMENTATION OF VOLTAGE WAVEFORMS lyzed separately. For this purpose, the expert system separates
The classification scheme presented in this paper is based the recording into groups of segments that start and end with
on identifying and characterizing the different stages of voltage segments for which all the phases have normal voltage (above
during an event. All the above mentioned events are associated 0.95 pu) or zero voltage (interruption).
with at least one sudden change in the voltage magnitude and For events that present two changes close in time (for ex-
these changes define the stages in the recording. To find the time ample, fuse-cleared faults), the detection index increases due
indices where the fundamental magnitude changes a change de- to the first change but does not decay adequately before the in-
tection scheme is employed. The method is described in detail crease due to the second change. Therefore, the two changes
in [5]. It can be summarized as it is in the following. cannot be resolved. To overcome this problem, a search is done
The change detection scheme is based on monitoring how to find parts of the voltage signals where the estimated mag-
well the measurement data fit a model that consists of the fun- nitude presents low standard deviation for at least a half-cycle.
damental frequency component and a certain number of har- The magnitude of a Kalman filter of order 1 is used for this pur-
monics. A detection index is obtained using the residuals of pose in order to improve the time properties of the magnitude
three Kalman filters (one for each phase). The recordings are estimation [6]. The search is done for transition segments that
split into event segments (where the detection index is low) and present a voltage dip of magnitude lower than 0.90 pu.
transition segments (where the detection index is high). The Fig. 1 shows examples of the fundamental voltage magni-
Kalman filters are not only used to obtain the detection index tude and the segmentation results for all the cases considered
STYVAKTAKIS et al.: EXPERT SYSTEM FOR CLASSIFICATION AND ANALYSIS OF POWER SYSTEM EVENTS 425
dure terminates when all the events within the recording are
classified. If no segments are found initially and the measure-
ment does not present a dip, then the segmentation algorithm is
applied again with a lower threshold. Only the step change is
checked in this case.
B. Classification Module
The possible options for an event are checked according
to whether the recording contains zero values or according
to the number of event segments. The procedure ends when
the event matches with the characteristics of one of the events
in the knowledge base of the expert system. The modules
corresponding to the events are called in the following order.
a) If the recording contains zero voltage values: energizing,
fault, interruption, and nonfault interruption.
b) If the number of event segments is two: transformer satu-
ration, induction motor starting, and step change.
c) If the number of event segments is three: transformer satu-
ration and protection, single stage voltage dip due to fault.
d) If the number of event segments is higher than three: mul-
tistage voltage dip due to fault.
B. Thresholds
Two groups of thresholds are used by the expert system. The
first group of thresholds is utilized by the segmentation scheme
(see [5] for more details). The selection of these thresholds is a
trade-off between sensitivity and false alarms. The setting was
done after analyzing a large number of data. The tests that were
presented above showed that the performance of the system is
limited in a few cases where step changes (lower than 0.05 pu)
in the voltage magnitude were not captured by the segmenta-
tion algorithm. These small steps are usually due to protection
operation. Therefore, multistage events might be classified as
single stage and protection operation after transformer satura-
tion might not be detected.
Additionally, changes in the voltage magnitude which are
close in time might not be detected as mentioned already in Sec-
tion III. This might be the case for faults that are cleared by fuse
blowing. In these cases, classification of the recordings is not
transient: in the case of transformer saturation, the dominant possible.
harmonic is the second or the fourth. The second group of thresholds are used by the expert system
3) Transformer Saturation: Overvoltage might occur during for the characterization of the segments of the recordings. The
saturation [Fig. 3(e)] [10]. Due to the nature of the phenomenon setting of these thresholds is important for the identification of
(half-cycle variation within one cycle), the estimated magnitude transformer events. In this case, the effect of the thresholds is re-
is significantly lower than the maximum value of the voltage duced by using four different features for classification. The cor-
waveform. responding thresholds were set after analyzing a large number
4) Energizing: Travelling waves are initiated by the switching of data. Other events might present some of the characteristics
of a line that could cause transient overvoltages [7]. of transformer saturation as defined by these thresholds. How-
5) Step Change: Overvoltages that are related with a stepwise ever, the probability of all four characteristics being present in
change in the voltage are due to load, capacitor, reactor, or line an event other than transformer saturation is low. This was con-
switching [9]. A typical case of such an overvoltage is due to firmed by the tests.
capacitor energizing. In this case, oscillatory transients are pro- Finally, it is important to realize that the classification proce-
duced of significant magnitude [Fig. 3(f)]. dure requires that all the different stages of the event are cap-
tured by the monitor. If, for example, the monitor does not cap-
ture the voltage recovery after fault clearing, then classification
VII. PERFORMANCE EVALUATION is not possible.
A. Application of the Expert System
VIII. CONCLUSION
The system has been used to analyze measurements from
a medium voltage network [11]. The measurements were ob- This paper presents an expert system for automatic clas-
tained in a two-month period and the monitors are triggered sification of power quality recordings. The main objective
by current or voltage disturbances. As mentioned already, only of the system is to distinguish between the different types of
measurements which present a voltage dip lower than 0.95 pu or power system events according to their causes. The expert
an overvoltage higher than 1.05 pu are considered by the expert system considers events that present a considerable change in
system. voltage. These events are mainly voltage dips (fault-induced,
The classification and the analysis results are given in Table I. transformer saturation, and induction motor starting dips) and
Only 3% of the recordings could not be classified by the system. interruptions (fault and nonfault).
428 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 17, NO. 2, APRIL 2002
The classification procedure is based on segmenting the [7] M. A. Chapman, A. Martinez, E. Sabir, K. Wang, and Y. Liu, “Switching
voltage waveforms in points of sudden changes in the fun- and fault caused transients in electric power systems,” in Proc. IEEE
Power Eng. Soc. Winter Meet., vol. 2, 1999, pp. 1015–1021.
damental magnitude. Based on the segmentation results, [8] L. A. Kojovic, S. P. Hassler, K. L. Leix, C. W. Williams, and E. E.
a set of classification modules is utilized to classify the Baker, “Comparative analysis of expulsion and current-limiting fuse op-
event. Classification is based on features extracted from the eration in distribution systems for improved power quality and protec-
tion,” IEEE Trans. Power Delivery, vol. 13, pp. 863–869, July 1998.
voltage waveforms. The magnitude, its characteristics, and [9] A. Greenwood, Electrical Transients in Power Systems. New York:
the harmonic contents of the voltage waveforms are the most Wiley, 1991.
important features. [10] G. Sybille, M. M. Gavrilovic, J. Belanger, and V. Q. Do, “Transformer
saturation effects on EHV system overvoltages,” IEEE Trans. Power
The system is used to analyze measurements from a medium App. Syst., vol. 104, pp. 671–680, Mar. 1985.
voltage network and the results are presented. The system suc- [11] I. Hunter, “Power quality issues: A distribution company perspective,”
cessfully classifies the largest part of the recordings. The only Power Eng. J., vol. 15, no. 2, pp. 75–80, Apr. 2001.
problems that are found are related with either the failure in
detecting very small changes in the voltage magnitude or the
time resolution problems of the magnitude estimation and the
Emmanouil Styvaktakis received the B.Sc. degree in electrical engineering
detection. from the National Technical University of Athens, Athens, Greece, and the
The expert system enables fast and accurate analysis of large M.Sc. degree in electrical power engineering from the University of Manchester
databases and classification of the recordings in terms of the Institute of Science and Technology (UMIST), Manchester, U.K., in 1995 and
1996, respectively. Currently, he is pursuing the Ph.D. degree in signals and
origin. Event classification (instead of disturbance classifica- systems at Chalmers University of Technology (CUT), Gothenburg, Sweden.
tion) offers the means for better understanding and description He is currently a Research Assistant in the Department of Signals and Systems
of the operation of the system in terms of power quality. The at CUT. His research interests include signal processing applications in power
systems and power quality.
part of the expert system for the analysis of overvoltages shows
how the system can be used to obtain more detailed statistics
about the recordings.
Math H. J. Bollen (M’93–SM’96) received the M.Sc. and Ph.D. degrees from
ACKNOWLEDGMENT Eindhoven University of Technology (EUT), Eindhoven, The Netherlands, in
1985 and 1989, respectively.
The authors wish to thank Scottish Power for kindly offering Currently, he is a Professor in electric power systems in the Department
of Electric Power Engineering at Chalmers University of Technology (CUT),
the measurements that were presented in this paper. Gothenburg, Sweden. Before joining CUT in 1996, he was a Research
Associate at EUT from 1989 to 1993, and a Lecturer at the University of
REFERENCES Manchester Institute of Science and Technology (UMIST), Manchester, U.K.,
between 1993 and 1996. His research interests cover various aspects of power
[1] P. K. Dashand, M. M. A. Salama, and S. Mishra, “Classification of quality and reliability. He has published a number of fundamental papers on
power system disturbances using a fuzzy expert system and a Fourier voltage dip analysis and a textbook on power quality.
linear combiner,” IEEE Trans. Power Delivery, vol. 15, pp. 472–477, Dr. Bollen is active in several IEEE working groups on power quality.
Apr. 2000.
[2] S. Santoso, J. Lamoree, W. M. Grady, E. J. Powers, and S. C. Bhatt, “A
scalable PQ event identification system,” IEEE Trans. Power Delivery,
vol. 15, pp. 738–742, Apr. 2000.
[3] E. Styvaktakis, M. H. J. Bollen, and I. Y. H. Gu, “Classification of power Irene Y. H. Gu (M’94) received the Ph.D. degree in electrical engineering from
system events: Voltage dips,” in Proc. 9th Int. IEEE Conf. Harmonics Eindhoven University of Technology (EUT), Eindhoven, The Netherlands, in
Quality Power, vol. 2, Orlando, FL, Oct. 1–4, 2000, pp. 745–750. 1992.
[4] M. H. J. Bollen and E. Styvaktakis, “Characterization of three-phase She is an Associate Professor in digital signal and image processing at
unbalanced dips (as easy as one-two-three?),” in Proc. 9th Int. IEEE Chalmers University of Technology (CUT), Gothenburg, Sweden. She was
Conf. Harmonics Quality Power, vol. 1, Orlando, FL, Oct. 1–4, 2000, a Research Fellow at Philips Research Institute IPO (NL), Eindhoven, The
pp. 81–86. Netherlands, and Staffordshire University, Staffordshire, U.K., and a Lecturer
[5] E. Styvaktakis, M. H. J. Bollen, and I. Y. H. Gu, “Expert system for at the University of Birmingham, Birmingham, U.K., from 1992 to 1996.
voltage dip classification and analysis,” in Proc. IEEE Power Eng. Soc. Her current research interests include time-frequency signal analysis, image
Summer Meet., 2001. processing, and object recognition.
[6] , “Voltage dip detection and power system transients,” in Proc. Dr. Gu is an Associate Editor of IEEE TRANSACTIONS ON SYSTEMS, MAN,
IEEE Power Eng. Soc. Summer Meet., 2001. AND CYBERNETICS.