coury2002
coury2002
Taylor & Francis makes every effort to ensure the accuracy of all
the information (the “Content”) contained in the publications on our
platform. However, Taylor & Francis, our agents, and our licensors
make no representations or warranties whatsoever as to the accuracy,
completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of
the authors, and are not the views of or endorsed by Taylor & Francis.
The accuracy of the Content should not be relied upon and should be
independently verified with primary sources of information. Taylor and
Francis shall not be liable for any losses, actions, claims, proceedings,
demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in
relation to or arising out of the use of the Content.
This article may be used for research, teaching, and private study
purposes. Any substantial or systematic reproduction, redistribution,
reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access
and use can be found at http://www.tandfonline.com/page/terms-and-
conditions
Downloaded by [Moskow State Univ Bibliote] at 11:37 31 January 2014
Elec tric P o we r C ompo ne nts and Syste ms, 30:1137–1149, 2002
C opyright ® c 2002 Taylor & Francis
1532-5008/ 02 $12.00 + .00
DO I: 10.1080/ 15325000290085433
D. V. COURY
M. OLESKOVICZ
Department of Electrical Engineering
Downloaded by [Moskow State Univ Bibliote] at 11:37 31 January 2014
R. K. AGGARWAL
School of Electronic and Electrical Engineering
University of Bath
Bath, England
1. Introduction
Protective relaying aims to safeguard expensive equipment such as generators,
transformers, and cables, and also to maintain the system integrity that is necessary
for a continuous and economical supply of electric power. The role of protective
1137
1138 D. V. Coury et al.
of the network can then continue to work normally or at least under near-to-normal
conditions.
With the recent advances of Articial Neural Network (ANN) techniques, diŒer-
ent architectures have been suggested in the literature to solve problems related to
power systems. Neural networks are useful for power system applications by virtue
of the fact that they can be trained with o‚ ine data. An overview of the application
of ANNs to power systems can be found in [1]. Concerning the application in the pro-
tection eld, some pattern recognition approaches to fault identication should be
mentioned such as those in [2{5]. Design and testing of an ANN applied as a fault di-
rection discriminator for transmission lines is presented in [6]. Moreover, [7{12] show
some approaches concerning fault classication using ANNs. References [13{15]
show the ANN approach proposed to distance protection. It has been shown that
these schemes improve the protection system selectivity and they adapt to the
actual power system condition [14, 15].
The work presented in this article is an ANN approach to simulate a complete
scheme for distance protection of a transmission line. The distance protection
technique is based on a modular approach whereby diŒerent neural network modules
have been adopted for fault detection, fault classication, and fault location. A
complete integration among these diŒerent modules is therefore essential for the
correct behavior of the proposed technique. The three-phase voltages and currents
sampled at 1 kHz, in pre- and post-fault conditions, were utilized as inputs for
the proposed scheme. The Alternative Transients Program (ATP) [16] software
was used to generate data for a 400 kV transmission line under diŒerent faults.
The NeuralWorks software [17] was used to set up the ANN topology, train it,
and obtain the synoptic weights as an output. This software provides a exible
environment for research and the application of techniques involving ANNs. The
supervised error backpropagation (EBP) algorithm was utilized during the training
process.
198.9 km) , where incorrect results from the ANN distance relay were also expected.
Finally, the third zone was set at 150% of the line length EF; this corresponds to
240 km from bus E. Faults next to the region where trip/no trip condition exchanges
(95% and 130% for the lines used) had special treatment. In such cases, a lower
degree of sparsity was chosen when considering fault locations used for training.
As diŒerent operation conditions were used in the training process, an adaptive
behavior was attributed to the relay concerning fault location in the protection
zones.
In order to evaluate the applicability of the proposed scheme, a simulation of
the transmission line in a faulted condition was utilized. This work makes use of
a digital simulator of faulted Extra High Voltage (EHV) transmission lines known
as Alternative Transients Program (ATP) , as mentioned earlier. It should be men-
tioned that although the technique described is based on Computer Aided Design
(CAD) , practical considerations such as the Capacitor Voltage Transformer (CVT)
and Current Transformer (CT) transients, anti-aliasing lters (cutoŒfrequency of
300 Hz) and quantization (assuming a 16-bit ADC) on system fault data were also
included in the simulation. The simulated data obtained was thus very close to that
found in practice. The technique also considers the physical arrangement and the
characteristics of the conductors, mutual coupling, and the eŒect of earth return
path. Perfect line transposition was assumed. The 400 kV transmission line used to
train and test the proposed ANNs is shown in Figure 1.
This study takes into account phase-to-earth faults, phase-to-phase faults,
phase-phase-to-earth and three-phase faults. Voltage and current postfault values,
measured at bus E and sampled at 1 kHz, were utilized as data inputs to the ANN-
based technique. The data set was composed of various fault situations considering
diŒerent fault locations, fault resistances, and fault inception angles. It should be
pointed out that for the training process, approximately 333, 138, and 1,000 diŒerent
fault cases (for each type of fault) were utilized for the detection, classication, and
location schemes, respectively.
The NeuralWorks software was used to set up the ANN topologies, train them,
and obtain the appropriate weights. The data used by the neural networks learning
algorithms were divided into two parts: one used for training (called the total
training set) , and another one used for tests (called the test set) . In the experiments
carried out in this work, the total training set is further subdivided into the training
and the validation sets. The sizes of the training and validation sets are 80% and
20% from the total training set, respectively. Concerning the training/validation
data set, the fault conditions utilized in the ANN scheme were used in a random
Downloaded by [Moskow State Univ Bibliote] at 11:37 31 January 2014
fashion.
The data involved in the ANN training and test processes used data windows.
Four samples each of the three-phase voltages and currents from bus E were utilized
as inputs. In the data windows, when a new sample becomes available, the oldest
value is discarded and the new sample value is included.
Several diŒerent topologies for the MLP (multilayer perceptron) neural net-
works were tried for all the modules described. The supervised algorithm known
as the Norm-Cum-Delta Learning Rule [17] was used. It is a variation of the cu-
mulative Back-Propagation learning rule, which is immune to changes in the epoch
length. As an alternative to the sigmoid transfer function, the hyperbolic tangent
transfer function was utilized in this work. During the training process, diŒerent
learning rates in the range of 0.01 to 0.4 were used in the hidden and output layers.
Likewise, the momentum value was varied between 0.001{0.2 interval.
After training, the ANN architectures were converted to a C code and a
complete algorithm was developed. In the proposed algorithm, all the modules
were connected in a sequential form as shown in Figure 2. The process starts
with the fault detection module: (24 nodes in the input layer, 9 units in the
hidden layer, and one output layer with 2 units, i.e., architecture ANN1 24-9-2
(Table 1) . This module is the directional detector; i.e., it should recognize whether
a reverse/forward fault or a normal network situation occurs. The expected outputs
from ANN1 (D1 and D2 ) are represented in Table 1. When a fault is detected
(reverse or forward fault) the latest voltage and current samples are transferred to
the classication module (ANN2) .
As shown in Figure 2, ANN2 (architecture ANN2 24-16-4) is designed to classify
the phases involved in a reverse or forward fault situation. The ANN2 distinguishes
10 types of faults as illustrated in Table 1. The ANN2 expected outputs as sig-
nied by C1 , C2 , C3 , and C4 are also shown in the table for the diŒerent types of
faults.
After the nal decision concerning the classication module, the sampled data
are directed to one of three ANN architectures in the location module (ANN3 24-48-
44-3, ANN4 24-42-40-3 or ANN5 24-24-20-3) . ANN3 is responsible for phase-earth
fault location (phase A , B, or C to earth) . ANN4 for phase-phase, and phase-phase
to earth fault location (faulted phases AB, AC, BC, and faulted phases AB, AC,
or BC to earth) . The location of the three-phase fault is eŒected by ANN5. For the
location modules, the outputs L1 , L2 , and L3 are associated with the protection of
zones 1, 2, or 3 as shown in Table 1.
ANN Routine for Fault Detection 1141
Downloaded by [Moskow State Univ Bibliote] at 11:37 31 January 2014
Figure 2. A complete scheme for fault detection, classication, and location in transmis-
sion lines using neural networks.
Downloaded by [Moskow State Univ Bibliote] at 11:37 31 January 2014
Table 1
The ANNs expected answers
Classication scheme
Detection scheme ANN2 (24-16-4) Location scheme
1142
AB 1 1 0 0
AC 1 0 1 0
BC 0 1 1 0 ANN3 (24-48-44-3)
ANN1 (24-9-2) AB-E 1 1 0 1 ANN4 (24-42-40-3)
AC-E 1 0 1 1 ANN5 (24-24-20-3)
BC-E 0 1 1 1
ABC 1 1 1 0
ANN Routine for Fault Detection 1143
Downloaded by [Moskow State Univ Bibliote] at 11:37 31 January 2014
3. Simulation Results
A test set was created to evaluate the performance of the proposed algorithm.
A total of 405 diŒerent cases for each type of fault were utilized in the test set. The
test set was comprised of the following variables: (1) fault locations in the range
10{320 km in approximately 10 km intervals; (2) fault resistances ( «) (phase-earth,
30, 60, 90, and phase-phase, 0.3, 0.6, and 0.9) ; and (3) fault inception angles, 20¯ ,
50¯ , and 80¯ .
As in the training process, a window of four samples per phase for voltages and
currents (pre- and post-fault values) from bus E, sampled at 1 kHz, were utilized
as inputs to the ANNs. Figures 3 to 11 depict the results attained in relation to the
performance of the complete scheme. As can be seen from the gures, the detection,
classication, and location modules give a very high percentage of accurate answers.
The processing time for each module is seen in the gures.
The results presented in Figures 3 to 11 typify the e ciency of the scheme.
In all the tests, the detection module (ANN1) was able to recognize correctly a
normal condition as well as a fault condition in 100% of the cases. In almost all of
the 4,050 cases tested, the processing time for fault detection was 2 ms after fault
occurrence (Figures 3, 4, and 5) .
For the 4,050 cases which were tested (405 cases for each type of fault) ,
the second architecture (ANN2{classication module) was able to classify faults
accurately in 99.44% of the test cases, with classication times between 4 to 12 ms
after fault occurrence (Figures 6, 7 and 8) . It should also be mentioned that the
ANN2 is able to classify the phases involved in both reverse and forward fault
situations.
Figures 9, 10 and 11 illustrate the performance of the last three ANN archi-
tectures (ANN3, ANN4, and ANN5) associated with fault location. It can be seen
that the majority of the 3,240 cases tested on protection zones 1, 2, and 3 were cor-
rectly located; these gures also include the performance associated with transition
zones 1 and 2. One observes a high percentage of correct answers when considering
the protection zones. However, the transition zones, as expected, present a lower
percentage of correct answers.
With regard to the overall processing time, it is estimated that the complete
scheme (detection, classication, and location) can operate in an average time of
about 13 ms after the fault occurrence.
It should be pointed out that for all the modules, an ANN output µ 0.3 was
considered as equal to zero and an output ¶ 0.7 as equal to unity. Any other
outputs were considered as uncertain cases. It should also be mentioned that the
nal decision from the detection network is based on a consistent output for two
windows (not necessarily consecutive) . Likewise, the classication and location
networks outputs are taken as nal only if the outputs are consistent for four
windows (again, not necessarily consecutive) . This is done essentially to increase
the stability of the ANNs spurious noise.
5. Conclusion
The use of an ANN as an alternative method for fault detection, classication,
and location tasks for transmission line protection has been investigated in this
work. The three-phase voltages and currents sampled at 1 kHz, in pre- and post-
fault conditions, were utilized as inputs for the proposed scheme. This study took
into account phase-to-earth faults, phase-to-phase faults, phase-phase-to-earth and
three-phase faults.
A total number of 4,050 diŒerent faulted cases were applied to the transmission
system analyzed. The results obtained show that the global performance of the
ANN architectures was highly satisfactory in both accuracy and speed of response
for all the modules described. Considering the test cases for the detection module,
the ANN outputs converged to the requisite levels very quickly within up to
3 ms of the fault occurrence. This module also works as a direction discriminator
recognizing the diŒerence between reverse and forward faults on a transmission line.
1148 D. V. Coury et al.
The classication module was able to discriminate correctly the phases involved in
the fault in between 4 to 10 ms after the fault occurrence. Finally, considering
the three protection zones, the results clearly show that this approach leads to a
reliable location for all types of faults (with times between 8 to 15 ms after fault
occurrence) . It should be pointed out that the ANN relay estimated the expected
response in approximately 98% of the 4,050 patterns tested (including all types of
faults and protection and transition zones) considering changes in the operational
conditions of the system. An extension of the relay primary protection zone to 95%
of the line length was implemented, enhancing system security and improving its
performance; this is a signicant advantage over conventional distance relays which
have a reach setting of only 80%. It should be noted also that in almost all the test
cases, the operating time of the global scheme is 13 ms after fault occurrence.
Although the training and testing of diŒerent ANN-based modules involved
Downloaded by [Moskow State Univ Bibliote] at 11:37 31 January 2014
References
[1] Dillon, T. S., and Niebur, D., 1996, Neural Networks Applications in Power Systems,
CRL Publishing, London, England.
[2] Kandil, N., Sood, V. K., Khorasani, K., and Patel, R. V., 1992, “ Fault identication in
AC-DC transmission system using neural networks,” Transactions on Power Systems,
Vol. 7, No. 2, pp. 812–819.
[3] Berniere, A., D ’apuzzo, M., Sansone, L., and Savastano, M., 1994, “ A neural network
approach for identication and fault diagnosis on dynamic systems,” IEEE Transac-
tions on Instrumentation and Measurement, Vol. 43, No. 6.
[4] Al-Hasawi, W. M., Abbasy, N. H., and Mansour, M. M., 1997, “ Using fault induced
transients and neural networks for T. L. ultra high-speed fault detection and classi-
cation,” IPST ’97–International Conference on Power Systems Transients, Seattle,
pp. 412–417.
[5] Su, W . B., Mao, P. L., Bo, Z. Q., Aggarwal, R. K., and Johns, A. T., 1997, “ A
fault detection technique for compositive transmission circuits using articial neu-
ral networks,” UPEC ’97–32nd Universities Power Engineering Conference, Vol. 1,
pp. 291–294.
[6] Sidhu, T. S., Singh, H., and Sachdev, M. S., 1995, “ Design, implementation and
testing of an articial neural network based fault direction discriminator for protecting
transmission lines,” IEEE Transactions on Power Delivery, Vol. 10, No. 2, pp. 697–
706.
[7] Ghosh, A. K., and Lubkeman, D. L., 1995, “ The classication of power systems
disturbance waveforms using a neural network approach,” IEEE Transactions on
Power Delivery, Vol. 10, No. 1, pp. 109–115.
[8] Dalstein, T., and Kuliche, B., 1995, “ Neural network approach to fault classication
for high speed protective relaying,” IEEE Transmission on Power Delivery, Vol. 10,
No. 2, pp. 1002–1011.
[9] Chowdhury, B. H., and Wang, K., 1996, “ Fault classication in power systems using
articial neural networks,” International Journal of Engineering Intelligent Systems
for Electrical Engineering and Communications, Vol. 4, No. 2, pp. 101–112.
[10] Kezunovic, M., and Rikato, I., 1996, “ Detect and classify faults using neural nets,”
IEEE Computer Applications in Power, pp. 42–47.
ANN Routine for Fault Detection 1149
[11] Keerthipala, W . W. L., Wai, C. T., and Huisheng, W., 1997, “ Neural network based
classier for power system protection,” Electric Power Systems Research, Vol. 42,
pp. 109–114.
[12] Aggarwal, R. K., Xuan, Q. Y., Johns, A. T., and Bennett, A., 1999, “ Fault classica-
tion technique for double-circuit lines based on a combined unsupervised/ supervised
neural network,” IEEE Transactions on Power Delivery, Vol. 14, No. 4, pp. 1250–
1256.
[13] Khaparde, S. A., Warke, N., and Agarwal, S. H., “ An adaptive approach in distance
protection using an articial neural network,” Electric Power Systems Research,
Vol. 37, pp. 39–44.
[14] Jongepier, A. G., and Sluis, L., 1997, “ Adaptive distance protection for high speed
protective relaying,” IEEE Transactions on Power Delivery, Vol. 12, No. 1, pp. 97–
103.
[15] Coury, D. V., and Oleskovicz, M., 1998, “ Multi-layer neural networks applied to
Downloaded by [Moskow State Univ Bibliote] at 11:37 31 January 2014