1002 IEEE Transactions on Power Delivery, Vol. 10, No.

2, April 1995

NEURAL NETWORK APPROACH TO FAULT CLASSIFICATION

FOR HIGH SPEED PROTECTIVE RELAYING

Thomas Dalstein, Bemd Kulicke

Department of High Voltage and Power Engineering
Technical University of Berlin
Berlin, Germany

Abstract - This paper presents a new approach to fault

I
classification for high speed protective relaying and show its
effectiveness in computer simulations on parallel transmis- power
sion lines. The scheme is based on the use of neural network
architecture and implementation of digital signal processing
concepts. We begin our exploration with classifying several
fault types like 1-phase-to-ground, 2-phase-to-ground and
3-phase-to-ground faults. We proceed with classification of
arcing and non-arcing faults in order to obtain a successful
automatic reclosing. Encouraging results are shown and Figure I Schematic depictmon ofprotective relaymg on aparallel trans-
indicate that this approach can be used for supporting a new mission line system (380kV)
generation of very high speed protective relaying systems.
VRP.U.
A - n - - A

Power Systems, Protection, Relaying, Fault Classification, IRP U

---
Arcing, Automatic reclosing, Artificial Neural Networks.

I. INTRODUCTION
Protecting transmission lines is one important task to safeguard
VSP

'SP
U

U
I
-
n
--U
n -
-
n
U--
A
-
electric power systems. Faults on transmission lines need to be
detected, classified and located accurately and cleared as fast as vTp U A A A A A A
possible. In this framework the most important point make is w "^" w w v v
fast and reliable fault classification. A fundamental part of a
protective relay is a selector module. This module classifies
ITP U
- - -----
h A-- --- ---
time in ms

Zone (in Our simulations 1ookm) and the second time zone (also 3 phase-to-ground faults on phase R &I), s ( ~ 2and
) T(~3),
1OOkm). We present further a phase-to-phase fault occuring
phase-to-phase faults (RS,ST,RT),
after 30ms time mark in Fig.2. 10 types of faults can occur:
3 two-phase-to-ground faults (RSE,STE,RTE),
three-phase faults with or without ground connection
94 SM 488-7 PWRD A paper recommended and approved (RST,RSTE)

0885-8977/95/$04.00 0 1994 IEEE

 1003

usually try an auto-reclosing. This idea is successful if an It should be noted, that FNN processing nodes are connected
arcing-fault occurs and this fault is not permanent. On the other only in forward direction by links of variable weights. The
hand, an auto-reclosing on a non-arcing fault can damage operation of a neural net consists of the presentation of a set of
equipment tremendously. Therefore, we want to classify arcing input signals (input patterns) and subsequent propagation ofthis
and non-arcing faults. Arcing fault detection is a slower decisi- inputs through the net. If activation and output functions are
on, because the relay can estimate signals until the circuit chosen, a neural net is completely described by its weights and
breakers cut the line. Hence, arcing fault detection is a "slow" node thresholds. Finding weights and thresholds for the net-
decision and has to be done in less than 70ms in contrast to the work may be regarded as being equivalent to finding the un-
fault type classification, if the fault is located in the first time known input/output relationship. Thus, neural networks are
zone. appropriate and especially powerful when they are used to find
In this paper we present a design of this new system that is such relationships that are difficult to describe explicitly. The
capable of providing a reliable and fast estimation of fault type next section describes a training process of feedforward neural
and fault class in the basis of real-time measurements of 3 phase networks.
voltages and currents. We used several high end personal com-
puters for training, testing and evaluating our approaches. The B. TRAINING
enabling technologies are real-time signal processing and neu- In order for a neural net to learn certain relationship, data sets
rocomputing. Developed algorithms are implemented in C and describing that relationship must be presented. These data sets
C++, respectively. In the following section, in the interest of consist of input vectors and associated target (output) vectors.
completeness, we give a short description of computing char- A training set describes the full range of expected inputs and
acteristics of a feedfonvard neural network. desired outputs. The neural nets used in this study are trained
by the Back Propagation Learning Algorithm proposed by
2. FEEDFORWARD NEURAL NETWORKS (FNN) Rumelhart [5,6].This learning rule, also known as Generalized
A. DESCRIPTION Delta Rule, exploits gradient information of the error function.
In 1986,Rumelhart, Hinton and Williams [5,6]presented arule Calculating the individual pattern error Em of pattern m
1
capable of adjusting behavior of a feedfonvard neural network
with hidden layers. Generally speaking, a feedforward neural
Em = s3;C(tmromz)
2
(2),
Z
net (FNN) contains an input layer, an output layer and possibly where tmz is the desired output of pattern m and o m is the actual
many hidden layers. Each layer can have one or many proces- net output. We get the error E for all patterns as a sum of all
sing nodes (neurons). A schematic illustration of a processing individual pattern errors as
node is presented in Fig.3.
The node receives its inputs through a set of weighted links. E = C E m =E(w) (3).
These inputs may come from other nodes or li-om outside m
sources. Sum of all weighted inputs represents the node activa- Minimisation of Error E is the task of a gradient search. The
tion. The node output is determined by an output function, weight updating process follows the direction of negative gra-
which responds to this activation. Frequently the so called dient; namely
sigmoidal output function is used: dE
Awij = - T - (4)
1 awlJ
AX) =- (1).
1-e-x or
The node output travels along the link, either to other nodes or AW: = T 6y o r (5 1
to the output of the system. A feedforward neural net is simply where
a layered collection of processing nodes as presented in Fig.5. 6f=(l - 0 J2) OJ2 (OJ2 - 5 2)
(6)

for the node j in the output layer 2 and

x2-&g)Lb
x3 output
61 -
J
1
- (1 - oj ) 0)

for the node j in hidden layer 1.

i
6f &

Using these relationships, we can calculate recursively the

(7)

weight changes "backwards" from the output layer toward the

input layer. Once all weight changes are calculated the weights
are updated. The entire training process is repeated as long as
Figure 3: A n illustration of a processing node (neuron) the error E exceeds a specified threshold.
1004

3. FAULT TYPE CLASSIFICATION USING FNN can be done in a timely fashion. Five consecutive sample points
Many papers are published describing applications of artificial of current and voltage of each line are used as inputs of the fault
neural networks in power systems [7, 8,9, IO, 1 I]. This frame- type neural net. Therefore 30 input nodes build input layer. It
work deals with a new possibility using neural networks in also consists of two hidden layers and an output layer with 1 I
electric power systems. We proceed in two steps. First we nodes. Each output is responsible for one fault type, except the
demonstrate our approach for classification of the fault type. first node that signals the "Normal State". Therefore any time
Second we present the approach for arcing and non-arcing fault one of this 11 outputs is mapped to a value of 0.9 and all other
detection. nodes are mapped to a value of 0. I. Hence, this net should leam
a winner takes all problem. The net is presented in Fig.5.
A. APPROACH TO FAULT TYPE CLASSIFICATION
Fig.4 illustrates functional parts of a protective relay. Sampled B. TRAINING
values of currents and voltages of three lines build inputs of the We trained fault type neural net with more than 45000 training
system. A knowledge control module controls all other parts of pattems simulating all relevant fault types, different fault loca-
the relay and is responsible for sending trip signals. The first tions, loads, fault start times and faultless situations. Training
module, the fault detection, signals that a fault has occurred. pattems are derived from 2268 simulated faults on a 380kV
Different methods are known [ 12, 151, but we use our neural transmission line (Fig. 1). We used the simulation program
network approach to fault detection described in [ 161. A feature NETOMAC [12, 19, 201 to create this large body of training
extraction module that is one part of the knowledge control, data. The network modeling capabilities of the NETOMAC
calculates several features e.g. direction signal produced by a program are equal to the EMTP program [21]. However, NE-
neural net as we published in [ 131, a neural net that is respon- TOMAC is a commercial product of SIEMENS AG and the
sible for time-zone classification [ 141 and common distance power utility companies in Germany have used it for more than
algorithms [ 181. These features e.g. are used by the fault loca- 20 years in network simulations.
tion module. Our main interest module in this paper is the fault It has been found out, that a net with 30 inputs 20 nodes in first
classification module labeled in Fig.4. We now concentrate our hidden layer, 15 nodes in second hidden layer and 1I outputs
explanations on the classification of the fault type which is the (30-20- 15-1 1) is capable to minimise the error E to a final error
first part of the fault classification module (compare Fig.14). of 0.0 167. We used the Backpropagation Training Algorithm
This fault type classification is done by a neural net. It uses with dynamic learning rate. Therefore, the possibility of weight
samples ofnormalised currents (i) and voltages (v) only. A 1kHz changing decreases cycle by cycle until training is stopped. This
sample rate is used to ensure that the fault type classification learning strategy converges quickly. In Fig.6 we demonstrate
the learning error E and a validation error Ev over 100 cycles.
One can see that the leaming error decreases in 100 cycles to
0.0399 (instead of first cycle error 0.356). The entire training
process needs around 500 cycles and 24 hours computing time
on a PC using this 45360 training pattern set. Again, the final
trip signals error was 0.0167. During the first 100 cycles as presented in
Knowledge Control - b Fig.6 the error of the validation set also decreases. This valida-
tion set includes just a few hundred test pattems but indicates
that the net is able to generalise the learning problem accurately.
-
Figure 4 . Important modules of a protective re fay
Learning was done by using a special neural network tool [ 181.
nax: 0.355689
................................................................................
0.3241
Sampled
0.2Y25
Values
01 0.2609

Currents 0.2293
And
0.1977
Voltages
(3Lines) 0.1662

0.1346

0.1030
Hidden Layer 1 output Layer 0.0714
Input Layer Hidden Layer 2

.................................
Figure 5: Fault type classijcalion neural net (30 inputs, 20 nodes first
hidden layer, 15 nodes second hidden layer, I I outputs) Figw e 6 Trmnmng with learning and validation set ( f00 cycles)
 1005

IT p.u.
___ b

// // /I \/ \ sample pmts in ms

in ms
Figure 7. A phase-to-phase fault (Sr) wrthout ground connection Figure 10: Net outputs of aphase-to-phasefault (ST)
trained with 15000patterns

R3E

RTE

STE

RST

sample points in m s sample points in ms

Figure 8: Correct fault type classrfication of 2-phase fault Figure 1I : Net outputs of aphase-to-phasefault (ST)
(RTE) with ground connection (lines R and r ) trained with 30000patterns

sample points in ms sample points in m s

Figure 9: Fast and reliable estimation offault type TE Figure 12: Net oufputs of aphase-to-phasefault (Sr)
(1-phase-to-ground fault line r ) trained with 45000patterns
1006

C. RESULTS 4. ARClNG FAULT DETECTION USING FNN

Trained fault type neural net was tested with a set of inde- Opposite to fault type classification arcing fault detection has
pendent test patterns. In Fig.7, 8 and 9 we illustrate a typical to be done in less than 70ms until the circuit breakers have cut
fault situation and two fault type classifications. Faults occur the lines. On the other hand arcing fault detection is more
after Sms, e.g. a typical 2-phase-to-ground fault on phase S and difficult. Around 80% of all faults are produced by arcing [ 151.
T (Fig.7). In Fig.8 we see the output NS (Normal State = These non-permanent faults can be cleared by trying an auto-
faultless) switched from 0.9 to 0.1 after 8ms. Then some outputs reclosing. The aim of the auto-reclosing is to switch off the
increase for a while because some older net inputs present faulted line, followed by a fast reclosing in 0.5-1.5 seconds,
values of the faultless situation and some younger values alrea- irrespective of the fact whether the arc is extinguished or not. A
dy present a 2-phase-to-ground fault. Therefore, the net outputs reclosing of a permanent fault is very critical for the dynamic
are undefined during that time. We call this effect (at last Smsec) stability of power systems and can damage power equipment
the transient time of the net. After this short transient time the tremendously. Conventional relays usually try an auto-reclo-
RTE-output is 0.9 and other outputs are around 0.1. Classifica- sing. Therefore arcing fault detection is a task of fundamental
tion is accurate, fast and reliable. Another example is Fig.9. The importance.
I-phase-to- ground fault on phase T is also classified very A. APPROACH
quickly and reliably. The NS-output goes down in 7ms and the
TE-output comes up quickly. The reason for the good classifi- Arcing voltage is more or less a rectangular voltage plus noise
cation quality is the extensive body of training pattems. in phase with the short current. In Fig. 13 an arcing voltage in
If we train our net with just 15000 training patterns the net is relation to the short current is illustrated. This arcing voltage is
also able to classify the fault type, but estimation quality decrea- superimposed on the line voltage signal. Therefore we suggest
ses under acceptable bounds. In Fig. 10 we present a phase-pha- for each of the three lines a neural net that classifies arcing and
se (ST) fault type classification. As we can see some disturban- non-arcing faults. These nets are triggered by our fault type
ces of other fault type outputs exist. These disturbances require neural net (see Fig. 14). The arcing voltage has only a very small
a complicate post processing of the net outputs, to estimate the influence on the short current. Therefore, every net uses one
real fault type output but essentially possibly because the ST- period of sampled values of its line voltage. As we use a sample
output is already dominating in Fig.10. If we train our net with rate of lkHz each net has 20 inputs. The outputs of these nets
30000 training patterns, the classification quality is much better are mapped to a value of 1.O if pattems of an arcing fault have
(compare Fig. 1l), because disturbances by other outputs are to be estimated and outputs are mapped to a value of 0.0 if a
reduced but presented. Therefore we trained a net with more non-arcing fault is presented. Also this time, nets require two
than 45000 patterns to make sure that these disturbances are hidden layers.
minimised. As we show in Fig.12 the net is now able to classify
this phase-to-phase fault without any disturbances. The ST-out-
put comes up quickly and is reliable while all other outputs are
around 0.1 any time after the fault has occurred.
To ensure that the fault type decision is correct and fast we
added a post processing unit to the neural net. This post proces-
sing unit locks at three consecutive net outputs. Only if one
Figure 13: Arcing voltage and short current are inphase
output is three consecutive times over a specified threshold
(0.5) the post processing unit signals the resultant fault type to Fault Classification
the system output or other modules of the protective relay e.g.
knowledge control. During the first two sample points after a
fault has occurred, the outputs are locked. Therefore, the mini-
mum estimation time is 5msec. We have tested the net by
presenting 240 different faults with different fault locations,
Faun
loads, arcing and non-arcing faults, etc.. All 240 faults are DeteCtiOl
_.
classified correctly and in short time. In Table I fault type
classification time statistic is presented. The classification
average is less than 6ms. Longest classification time is 7ms.
The results indicate that this approach can be used for very high
speed protective relaying systems. The fault type net is also able
to detect a fault propagation from one line to other lines in a
timely fashion. We go further with an approach to arcing fault
detection. Figure 14: Schematic depiction of fault classrfcation approach
 1007

B. TRAINING Also net T made one mistake (2-phase-to-ground fault RTE).

Training of these nets is done by presenting a set of arcing faults However, detection net of phase R was able to recognise this
and a set of non-arcing faults. Both sets include different fault arcing fault in 3%" Hence, these mistakes can be corrected by
locations, fault start times, loads, noise, etc.. The arcing fault because l-phase-to-gound ar-
set also includes arcing faults with different arcing voltages cing faults are detected correctly, even if one of the arcing fault
(5kV-25kV). We trained only l-phase-to-ground faults in the detection nets Signals an arcing fault of a 2- or 3-Phase fault
matter of the Dresence of one net for one line. Therefore, the situation it makes to try an auto-reclosing.
total number of training patterns decreases to less than 3000
patterns. We found out that a (20-15-10-1) net is able to mini-
mise the error E to a final error less than 0.0 1. We also used the
Backpropagation Training Algorithm with dynamic learning 0

rate and the simulation program NETOMAC to create training

and test data. so
O D

6 0
C. RESULTS I

We tested trained nets with independent test patterns derived 0

from 240 different fault situations. Test pattems include faults

in the fvst and second time zone and some backward faults.
Arcing fault detection nets have a transient time of 20ms sample polnts in ms
because they use one period of sampled values of line voltages.
A post processing unit of 5 time-delay blocks makes sure that
arcing fault detection is reliable. So, minimum detection time Figure 15 3-phase arcingfault located 50kmfrom relay
is 25msec. In Fig. 15,16 and 17 we present three fault situations.
In Fig. 15 we illustrate a 3-phase arcing fault located 50km from
the relay. During the first period (transient time) outputs of the
nets are locked. Therefore, output signals are zero. Then all
outputs signal an arcing fault. Sometimes disturbancesmake the
output signal less "clear". Therefore, a post processing unit is
needed. If non-arcing faults are presented (e.g., Fig.17) outputs
3 06

are near zero. All 120 arcing fault situations are detected accu-
rately by using a post processing unit with 5 delay blocks.
However, estimation quality of arcing faults depends on the
starting time of the faults. For example if a 1-phase fault occurs 5

lms after current zero crossing, detection time increases to at

least 34ms, because it takes time to get pattems with reliable
arcing voltage information. An example is explained in Fig. 17.
A 3-phase arcing fault occurs later than the fault in Fig. 15. Flgure 16 3-phase non-arcingfault located 50kmfrom relay
Arcing fault detection signal of lines R and T are fast (less than
30ms), but output signal of line S needs time to stabilise
(detection time around 53ms). On the other hand, the relay
receives correct data until circuit breakers cut the faulted lines
(70ms). Therefore, arcing fault detection time is measured from 0 1

25ms to 70ms. An arcing fault detection time over 70ms is equal 0

to wrong detection. In table I1 arcing fault detection times of 3

lines of 120 arcing faults (on each line 72 arcing faults are
3 0

estimated from 1-phase-to-ground, 2-phase-to-ground, phase-

to-phase and 3-phase faults) are shown. Most of these arcing 0

faults can be detected very fast (less than 50ms). Less than 10% T

of the faults can be estimated in 50-70ms. 3 arcing faults could

not be detected, one on each line. Arcing detection net of phase sample points in ms
R missed a 3 -phase fault (RST) but nets of lines S and T detected
an arcing fault presenting this fault situation. Detection net S
made one mistake presenting a 2-phase-to-ground fault (STE). Figure 17: 3-phase arcing fiult located 50km from relay occured lOms
later (some disturbances on phase S).
Detection net of phase T estimated this arcing fault in 25ms.
1008

TABLE I1 Arcing fault detection times of3 lrnes (R, S, r ) [4] Heuck, K., Dettmann, K.-D.; Elektrische Energieversor-
gung Vieweg & S o h VerlagsgesellschaftmbH, Braunschweig,
Germany, 1984.
[SIRumelhart, D.E; McClelland, J.L.: Parallel distributed
processing. Explorations in the Micro structure of Cognition.
Volume 1 :Foundations,MIT Press, Cambridge, Massachusetts
1986.
[6]Rumelhart, D.E; McClelland, J.L.: Parallel distributed
processing. Explorations in the Micro structure of Cognition.
Volume 2 : Psychological and Biological Models, MIT Press,
In contrast to conventional methods of fault selector modules Cambridge, Massachusetts 1986.
our approach is able to classify the fault type very fast and can [7]Boehme, K.; Kulicke, B.: Optimierung von Distanzschutzal-
detect whether an arcing fault or anon-arcing fault has occurred. gorithmen mit Hilfe neuronaler Netze; ELEKTRIE, January
Obtained results encourage to use this approach for supporting 1993, Berlin, Germany.
a new generation of very high speed protective relays. It also
might be useful in addition to conventional relays to develop a [SISobajic, D.J., Pao, Y.H.: Artificial neural-net based dynamic
successful automatic reclosing. security assessment for electric power systems; IEEE Tran-
sactions on Power Systems, Vo1.4, No.4, February 1989,
5 . FUTURE pp.220-228.
We want to test our approach by means of other simulation [9]Sobajic, D.J., Pao, Y.H., Dolce, J.: On-line monitoring of
models and transient records of fault situations of real world power systems operating conditions using artificial neural net-
protective relaying systems. Further we will connect fault type works, IEEE International Symposium on Circuit and Systems,
classification approach with our neural fault direction estimati- Portland, OR, May 8-1 1, 1989, pp.2243-2246.
on module [ 131 and our neural network approach to fault time
zone classification [14]. Our goal is a multi-neural network [10]Feser, K.; Braun, U.: Application of neural networks in
based protective relaying system that we will present in a future numerical busbar protection systems (NBPS); First Intematio-
work. nal Forum on Applications of Neural Networks to Power Sy-
6. CONCLUSION stems, Seattle, WA, July 23-26, 1991, pp.117-121.
This paper presented a novel multi-neural network based ap- [ 1lIKhaparde, S.A., Kale, P., Aganval, S.H.: Application of
proach to fault classification of high speed protective relaying neural network in protective relaying oftransmission lines, First
systems. Basically, neural network computing and implemen- International Forum on Applications of Neural Networks to
tation of digital signal processing concepts are used. The new Power Systems, Seattle, WA, July 23-26, pp.139-143.
method is separated into two main tasks. We first suggested an [ 12]Kulicke, B.:Digitalprogramm NETOMAC zur Simulation
approach to fault type classification of 1-phase-, 2-phase- and elektromechanischer und -magnetischer Ausgleichsvorgaenge
3-phase-faults in order to estimate the faulted lines. Classifica- in Drehstromnetzen; Elektrizitaetswirtschaft, Heft 111979,
tion of fault type is fast (5-7ms) and reliable. Second we S. 18-23, Germany.
presented an approach detecting arcing- and non-arcing faults
to support a successful automatic reclosing. Obtained results [13] Dalstein, T. ,Sobajic, D.J, Kulicke, Pao, Y.H.: Neural
were encouraging and indicated that this approach can be used Network Approach To Fault Direction Identification In Electric
to support conventional protective relaying systems. It can also Power Systems, Proceedings North American Power Symposi-
be used as a part of a new generation of high speed protective um, October 11-12, 1993, Washington D.C.
relaying systems. [14] Friedrich, T., Dalstein, T., Kulicke, B.: A Neural Time
7. REFERENCES Zone Classifier For High Speed Protective Relaying. Accepted
for Universities Power Engineering Conference UPEC’94, Sep-
[ 11 Muller, L.; Boog E.: Selektivschutz elektrischer Anlagen tember 14th-l6th, Galway, Ireland, 1994.
VWEW-Verlag, Frankfurt am Main, Germany, 1990.
[ 151 Reck, T.:Vergleich von Signalverarbeitungsmethoden zur
[2] Clemens, H., Rothe, K.; Schutztechnik in Elektroenergiesy-
Bestimmung der Netzimpedanz fir den Distanzschutz, Disser-
stemen; Verlag Technik GmbH, Berlin, Germany, 1991.
tation,, Technische Universitat Berlin, 199 1, FRG.
[3] Ungard, H., Winkler, W., Wiszniewski, A.; Schutztechnik [ 161 Dalstein, T: Fault Detection Using Neural Networks, Pa-
in Elektroenergiesystemen Grundlagen, Stand der Technik, tent Application, SIEMENS AG, 93P8562, Berlin, 21.09.1993,
Neuentwicklungen; Springer Verlag Berlin Heidelberg, Ger-
FRG.
many, 1991.
 1009

[ 171 NEUROTEC: Neuro-Compiler Professional 2.1, Referenz

Manual, Berlin, 1993, FRG.
[ 181Eickmeyer, D.: Simulation, Untersuchung und Beurteilung
der Leistungsfhigkeit verschiedener Distanzschutzalgorith-
men, Master Thesis ,Technische Universitgt Berlin, November
1993, FRG.
[19] Kulicke, B., Bayer, W., Kriiger, K.H., Povh, D.: Studies
for HVDC and SVC using the NETOMAC Digital Program
System. 87JC-32 IEEE P E W S E E Joint Conference on High
Vol-tage Transmission Systems in China, Beijing, October
17-22, 1987.
[20] Lasseter, R.H., Kriiger, K.H.: HVDC Simulation using
NETOMAC. IEEE Montech’86, Conference on HVDC Power
Transmision Sept. 29- Oct. 1, 1986.
[2 13 Dommel, H.W., Meyer, W.S.: Digital Computer Solution
of Electromagnetic Transients in Single- and Multi-Phase Net-
work, IEEE, Vol. PAS-88, p. 388-399, April 1969.
BIOGRAPHIES
was bom in Berlin, Ger-
many, on May 29, 1967. He received the
M.S. degree in electrical engineering
from the Technical University of Berlin,
Berlin, Germany, in October of 1991. He
is pursuing the Doctor degree in electrical
engineering at the Technical University
of Berlin and is with the Institute of High-
Voltage and Power Engineering. He
works in co-operation with the Siemens
AG, Berlin, Munich, Germany. His areas of interest are power
system protection, neural network computing, classification
and clustering. Dip1.-Ing. Thomas Dalstein is a student member
of IEEE Power Engineering Society.

Bemd Kulicke was bom in Wemigero-

de/Germany on November, 1944. He
received the M.S. degree at the Technical
University of Berlin and the Doctor de-
gree in Power Engineering from the Uni-
versity of Darmstadt in 1970, on 1975
respectively. From 1970 to 1983 he was
with the Siemens Company, working in
the High-voltage, Power Engineering De-
partment. He was responsible for the de-
velopment of the NETOMAC program and was mainly invol-
ved in performing system studies including electromechanical
and -magnetical transients and stability problems. In 1984 he
was appointed a Professor and Director of the Department of
High-Voltage and Power Engineering at the Technical Univer-
sity of Berlin. Prof. Dr.-Ing. Kulicke is a member of the IEEE
Power Engineering Society.
1010

Discussion would be appreciated if the authors could provide the methodologyor

V.K. SOOD, (Concordia University, Montreal, Canada) : The authors some details of such a unit.
are congratulated for an excellent publication on the application of neu- [A] N.Kandi1, VKSood, KKhorasani and R.VPate1, “Fault Identifi-
ral networks (NNs) for fault classification in transmission systems. cation in an AC-DC transmission system using Neural Net-
The authors results are in keeping with our findings [A,B,C] for similar works”, 1991IEEE PICA Conference, May 6-10, Baltimore. Also
applications. In [A], we used similar techniques to classify faults, and published in IEEE Trans on Power Systems, May 1992, Vol7, No
in [B,C]we used similar techniques to classify faults and fault location 2, pp 812 - 819.
also. The modelling was done with an electromagnetic transients pro-
[B] T Grigoriu, VK.Sood, KKhorasani and R.VPatel, “Fault Identi-
gram called EMTDC which is similar in its mode of operation to NE- fication in a series compensated AC line using Neural Net-
TOMAC. In our case, we used a hidden layer with only one layer whilst works.”, Canadian Conference on Electrical and Computer Engi-
the authors have used two hidden layers. In Figure Al, the case of a line- neering, Toronto, 13-16 Sept., 1992.
to-line (phases B and C - Channel No.5) at four locations is shown; the
4 locations are at the receivingend (Channel ll), after the series capaci- [C] ‘Grigoriu, V.K.Sood, K.Khorasani and R.V.Pate1, “Fault Identi-
I
tor (Channel l2),before the series capacitor (Channel U) and the send- fication in a series compensated AC‘ line using Neural Net-
ing end (Channel 14). works.”, Canadian Electrical Association Spring Meeting,
Montreal. 29 - 31 March 1993.

Manuscript received August 15, 1994.

Rectiving end BC fault
.... . .. ., _....,..,._
1
3
n
0
T. S. Sidhu (Power Systems Research Group, University of
Saskatchewan, Canada): The paper presents the use of
seconds
Channel No. artificial neural networks for classifying shunt faults on
ARU capacitor
transmission lines. The authors report classification times of
BC fault
5 to 7 ms. Their response LO the following points will be
1 appreciated.
3

0
It is mentioned in the paper that 240 different faults were
seconds used to test the proposed neural network. It is not clear
Channel No.
if these test cases were generated using the same power
Before capacitor BC fault system which was used for generating the training data.
1
The generalization capability of the neural network
3
should be tested by using data taken from different
0 systems because the fault waveforms are affected by
seconds factors, such as, power system configuration, line
Channel No.
configuration and conductor arrangement.
Sending end BC fault
It seems that the authors have trained and tested their
1
Ll
neural network using simulated fault data from 380 kV
0 transmission system. Does this mean that the network
seconds
has to be retrained if it has to be applied to transmission
Channel No. systems of other voltage levels‘?

Would the authors discuss the response of their neural

Figure A1 : Fault locations with common line-to-line (BC) fault network to situations, such as, simultaneous faults,
evolving faults, etc.

I would appreciate the authors comments on the following : Manuscript received August 22, 1994.
1. One of the problems of using NNs is the necessity of pre-proces-
sing data presented to the selected NN architecture. In the case of input
variables of the same type (i.e. ac bus voltages), this problem can be re-
solved by normalization. However, with input variables of different T. Dalstein, B. Kulicke:
types (i.e. ac bus voltages and line currents), problems may arise since We thank Dr. Sood and Dr. Sidhu for their interest in our paper
scaling on a common base may not be feasible. Could the authors dis- and the valuable discussion. The insightful questions and
cuss the, pTs-processing requirements of the input signals to the NNs? thoughtful comments complement the paper and raise several
2. keeping with our findings [A,$C], the authors also experi- interesting marks. In the following we would like to address the
enced Some difficultyin identifyingline-to-fine faults. However, the au- questions. We start with the questions of discusser Sood and
thors propose a post-processing unit (knowledge control module). It proceed with the questions of discusser Sidhu.
 1011

( 1 ) The first point is the pre-processing requirement of our other loads, different fault locations, noise, fault occurring
neural nets. As we pointed out in our paper, we just use norma- times, DC-components, etc. Therefore, it was the first step to
lized, sampled values of the current and voltage signals. These test the generalization capability. Then, we used fault data
data are obtained from current and voltage transformers. The recorded in the Berlin power system and test data analog
inputs range between +I after normalization. Therefore, we simulated. Using these test data we found out the generalization
have no problems with different types of input variables like ac capability is remarkable.
bus voltage and line currents because every input signal is (4) Our NNs have not to be retrained on other voltage levels
normalized individually. because of the normalization property. The fault data of the
(2) The authors thank Dr. Sood for the interesting references. power system used are on a 1 l0kV-level. The results were also
Fault identification in an AC-DC-transmission system is also a fast and reliable. Nevertheless, it might be better to retrain the
task of hndamental importance. Our aim is a very high speed net for other systems but it has to be retrained in series compen-
solution to fault identification. Therefore, we have to use a sated transmission systems. Our system developed is able to
simple post-processing unit. To identify the fault as fast as simulate, train and test a neural net in about 24h on a personal
possible and reliable, each post- processing unit of a neural computer (66MHz).
network output consists ofiwo delay blocks plus an And-ope- (5) We have also tested load jumps, simultaneous faults and
rator with specified threshold. Hence, if three times an output evolving arcing. E.g., two one-phase faults that occur sequently
is above the threshold (e.g., 0.5) the post-processing unit signals in less than 7ms lead to two-phase fault identification. Other-
the obtained fault type. wise, the identification is able to classify the first one in an
(3) The first comment of Dr. Sidhu is right. The 240 test faults accurate manner. Of course, evolving faults can be handled, but
are obtained from the same power system to get comparable the responding time increases up to 8ms after the fault evolved.
results. But the conditions of the test faults were completely
different to the faults trained. We generated these faults with Manuscript received October 31, 1994.