Electrical Power and Energy Systems 25 (2003) 747–758

www.elsevier.com/locate/ijepes

A novel algorithm for fault classification in transmission lines using

a combined adaptive network and fuzzy inference system
S.M. Yeoa, C.H. Kima,*, K.S. Honga, Y.B. Limb, R.K. Aggarwalc, A.T. Johnsc, M.S. Choid
a
School of Information & Computer Engineering, Sungkyunkwan University, 300, Cheoncheon-dong Jangan-gu, Suwŏn 440 746, South Korea
b
LG Electronics CDMA Handsets Laboratory, Seoul, South Korea
c
Department of Electronic & Electrical Engineering, University of Bath, Bath BA2 7AY, UK
d
Division of Electrical and Information control Engineering, Myongji University, Yongin, South Korea/NPTC
Received 31 January 2002; revised 10 February 2003; accepted 25 February 2003

Abstract
Accurate detection and classification of faults on transmission lines is vitally important. In this respect, many different types of faults occur,
inter alia low impedance faults (LIF) and high impedance faults (HIF). The latter in particular pose difficulties for the commonly employed
conventional overcurrent and distance relays, and if not detected, can cause damage to expensive equipment, threaten life and cause fire
hazards. Although HIFs are far less common than LIFs, it is imperative that any protection device should be able to satisfactorily deal with
both HIFs and LIFs.
Because of the randomness and asymmetric characteristics of HIFs, the modelling of HIF is difficult and many papers relating to various
HIF models have been published. In this paper, the model of HIFs in transmission lines is accomplished using the characteristics of a ZnO
arrester, which is then implemented within the overall transmission system model based on the electromagnetic transients programme.
This paper proposes an algorithm for fault detection and classification for both LIFs and HIFs using Adaptive Network-based Fuzzy
Inference System (ANFIS). The inputs into ANFIS are current signals only based on Root-Mean-Square values of three-phase currents and
zero sequence current.
The performance of the proposed algorithm is tested on a typical 154 kV Korean transmission line system under various fault conditions.
Test results show that the ANFIS can detect and classify faults including (LIFs and HIFs) accurately within half a cycle.
q 2003 Elsevier Science Ltd. All rights reserved.
Keywords: Adaptive network; Fuzzy inference system; Fault classification; Transmission line

1. Introduction mission system and isolate the faulted part from the rest of
the system thereby preventing the propagation of the fault
Since the complexity of modern power systems is into other parts [1].
increasing (longer lines, increased power transfer over In this respect, there is now ongoing work to further
existing lines due to the limitations imposed by environ- improve the performance of digital protection relays.
mental pressures, etc.), traditional analogue relaying is no Protecting transmission lines is one important task to
longer able to cope with the performance requirements, and safeguard electric power systems. Faults on transmission
hence the advantage of employing digital protection relays lines need to be detected, classified and located accurately
which are much better suited to cope with the modern-day and cleared as fast as possible. Fast and reliable fault
protection problems, particularly in terms of speed and classification is thus paramount in the overall protection
accuracy. The purpose of a protective relaying system is to strategy [1,2].
detect the abnormal signals indicating faults on a trans- Many researchers have studied the application of neural
networks to overcome many of the aforementioned
problems [3 – 11]. Hitherto, the algorithms developed
* Corresponding author. Tel.: þ 82-31-290-7124; fax: þ82-31-290-7955. include the high frequency voltage signal method [5], a
E-mail address: [email protected] (C.H. Kim). statistical method [6], a numerical algorithm [7,8], wavelet
0142-0615/03/$ - see front matter q 2003 Elsevier Science Ltd. All rights reserved.
doi:10.1016/S0142-0615(03)00029-2
748 S.M. Yeo et al. / Electrical Power and Energy Systems 25 (2003) 747–758

transform [9], the neural network and neuro-fuzzy network fn ¼ pn x þ qn y þ rn ð1Þ

[10,11].
The fuzzy set theory is also used to solve uncertainty 2.1. ANFIS architecture [12,13]
problems. However, all these methods make use of either
neural networks or fuzzy sets in isolation. The use of The ANFIS architecture consists of two trainable
neural nets in applications is rather sparse due to its parameter sets.
implicit knowledge representation, the prohibitive com-
putational effort, and so on. In principle, the key benefit 1. The antecedent membership function parameters
of fuzzy logic is that its knowledge representation is {a; b; c}; also called the premise parameters.
explicit, using simple ‘IF-THEN’ type of relations. 2. The polynomial parameters {p; q; r}; also called the
However, this ‘attribute’ is at the same time its major consequent parameters.
limitation, as the power system operation in the transient
period, which is effected by many unknown parameters, The ANFIS training paradigm uses a gradient descent
cannot be easily described by artificial explicit knowl- algorithm to optimise the antecedent parameters and a
edge. The integration of neural network into the fuzzy least square algorithm to solve for the consequent
logic system, however, makes it possible to learn from parameters. Because it uses two very different algorithms
data a priori [3,4]. to reduce the error, the training rule is designated as being
Jang and Sun introduced the Adaptive Network-based hybrid. The consequent parameters are updated first using
Fuzzy Inference System (ANFIS). This system makes use of a least squares algorithm and the antecedent parameters
a hybrid learning rule to optimise the fuzzy system are then updated by back propagating the errors that still
parameters of first-order Sugeno system [12,13]. It is exist.
composed of five layers. Each layer is the component of The ANFIS architecture consists of five layers with the
fuzzy inference system and performs different actions. output of the nodes in each respective layer represented by
Using back propagation method of training a neural an output variable, as detailed in Ref. [13]; these five layers
network, fuzzy premise and consequent parameters are are categorised as: (i) fuzzification (here all nodes are
tuned properly. adaptive); (ii) conjunction (here all nodes are fixed); (iii)
This paper develops an algorithm of fault classification normalization (here all nodes are also fixed); (iv) defuzzi-
using ANFIS. The inputs for ANFIS are Root-Mean-Square fication (here all nodes are adaptive), and (v) summation
(RMS) values of three-phase currents and zero sequence (this comprises of a single fixed node). It is in the last layer
current. The performance of the proposed algorithm is that the consequent parameters can be solved for, using a
least square algorithm.
evaluated for a typical 154 kV Korean transmission line
system under various fault conditions.
2.2. Hybrid learning algorithm

From the ANFIS architecture shown in Fig. 1(b), it can

2. ANFIS be seen that when the values of the premise parameters are
fixed, the overall output can be expressed as a linear
Jang and Sun introduced the ANFIS. This system makes combination of the consequent parameters. In symbols, the
use of a hybrid learning rule to optimise the fuzzy system output f in Fig. 1(b) can be rewritten as
parameters of first-order Sugeno system, which can be
f ¼w

1 f1 þ w

2 f2
graphically represented by Fig. 1.
For a first-order Sugeno fuzzy model, a typical rule set ¼ ðw
1 xÞp1 þ ðw

1 yÞq1 þ ðw

1 Þr1 þ ðw

2 xÞp2 þ ðw

2 yÞq2
with two fuzzy if-then rules can be expressed as:
þ ðw

2 Þr2 ð2Þ
Rule 1
which is linear in the consequent parameters p1 ; q1 ; r1 ; p2 ;
If x is A1 and y is B1 ; then f1 ¼ p1 x þ q1 y þ r1 ;
q2 ; r2 : Therefore, the hybrid learning algorithm developed in
Rule 2
Section 2.1 can be applied directly. More specifically, in the
If x is A2 and y is B2 ; then f2 ¼ p2 x þ q2 y þ r2 :
forward pass of the hybrid learning algorithm, node outputs
go forward until layer 4 and the consequent parameters are
Fig. 1(a) gives the reasoning mechanism for this Sugeno identified by the least squares method. In the backward pass,
fuzzy model. The corresponding equivalent ANFIS archi- the error signals propagate backward and the premise
tecture is shown in Fig. 1(b), where nodes of the same layer parameters are updated by gradient descent. Table 1
have similar functions. summarizes the activities in each pass.
The consequent parameters (p; q and r) of the nth rule As mentioned earlier, the consequent parameters thus
contribute through a first-order polynomial of the form: identified are optimal under the condition that the premise
 S.M. Yeo et al. / Electrical Power and Energy Systems 25 (2003) 747–758 749

Fig. 1. ANFIS architecture for a two-input, two rule first-order Sugeno model. (a) A two-input first-order Sugeno fuzzy model with two rules. (b) Equivalent
ANFIS architecture.

parameters are fixed. Accordingly, the hybrid approach impedance fault (HIF) simulation is based on a ZnO arrester
converges much faster since it reduces the dimension of the model [14 – 17] which is embedded into the overall
search space of the original back propagation method. transmission line model, and the impedance of HIF is
approximately 200 V.
Table 2 shows source and line parameters of the model
3. Fault classification algorithm system considered and Table 3 shows the various fault
conditions.
3.1. System model studied
Table 1
In the work presented herein, the model system studied is
Two passes in the hybrid learning procedure for ANFIS
the Korean 154 kV system as shown in Fig. 2.
It comprises of a 26 km line length terminated in two Forward pass Backward pass
sources of 240 and 180 MVA each at both ends of the line
and fault data is generated for fault conditions such as Premise parameters Fixed Gradient descent
Consequent parameters Least-squares estimate Fixed
variation in fault distance, fault inception angle and fault
Signals Node outputs Error signals
types, at a sampling rate of 64 samples per cycle. The high
750 S.M. Yeo et al. / Electrical Power and Energy Systems 25 (2003) 747–758

Fig. 2. Simulation model of 154 kV system.

3.2. The characteristics of fault currents In this paper, this problem is overcome by developing a
fault classification algorithm using RMS values of phase
Generally, when faults occur on transmission lines, currents combined with the zero sequence current as shown
magnitudes of faulted phase currents are increased and are below.
often manifested with either harmonic or dc-offset com- Zero sequence current is calculated by the equation
ponents. However, the increase in fault phase current(s) in 1
the case of HIF is much smaller in comparison to the low I0 ¼ ðI þ Ib þ Ic Þ ð3Þ
3 a
impedance faults (LIF).
Fig. 3 typifies the RMS values of currents for each fault But zero sequence current in the case of HIF is increased at a
type. As expected, the increase in fault currents under LIF is much smaller rate in comparison with LIF and thus in the
much higher than under HIF. Thus, in principle, fault technique developed herein, Eq. (3) is modified into the
classification between LIF and HIF can be achieved through following equation:
a knowledge of RMS values of currents. However, deducing I0 ¼ ðIa þ Ib þ Ic Þ ð4Þ
only change of RMS values is not sufficient to classify fault
type between double line-to-ground fault and line-to-line As can be seen, although the magnitude of modified zero
fault because of similarities of change in the two types of sequence current is increased, it still retains the character-
fault and other means of accurately classifying these types istics of the zero sequence current. Fig. 4 shows the RMS
of fault have to be found. values of zero sequence current for each fault type. As
expected, magnitudes of the zero sequence currents are near
zero for fault types such as line-to-line fault, three-phase
Table 2 fault, and normal state, but a fault involving ground has a
Source and line parameters non-zero value.
Fig. 5 shows the distribution region of RMS values of
Zero sequence Pos.
sequence zero sequence currents for each fault type on condition that
the magnitude of RMS value of phase current at normal
Line constants R (V/km) 0.2293 0.0419
L (V/km) 1.0050 0.3316
C (m9/km) 1.6260 4.8309 Table 3
Fault conditions
YongIn substation Capacity (MV A) 240
Power factor 0.91 Test condition
Zsource (V) Z0 ¼ 1:128 þ j0:0155
Z1 ¼ 0:820 þ j6:7482
Type SLG
Zload (V) ZY ¼ 89:93 þ j40:96
DLG
Ansung substation Capacity (MV A) 180 DLL
Power factor 0.91 3fG
Zsource (V) Z0 ¼ 1:5050 þ j7:5775 HIF
Z1 ¼ 0:9400 þ j8:6595 Inception angle (8) 0, 30, 60, 90
Zload (V) ZY ¼ 119:90 þ j54:62 Location (km) 5.2(20%), 13(50%), 20.8(80%)
 S.M. Yeo et al. / Electrical Power and Energy Systems 25 (2003) 747–758 751

Fig. 3. RMS values of current for each fault type.

steady state is set to be 100%. Results of Fig. 5 are herein has only one input, layer 2 can be combined with
summarised in Table 4. layer 1.
The following three rules are used:
3.3. Fault classification algorithm Rule 1
If x is A1 then f1 ¼ p1 x þ r1 :
As shown above, the fault classification algorithm is Rule 2
developed using ANFIS based on the characteristics of If x is A2 then f2 ¼ p2 x þ r2 :
faults. Fig. 6 shows the diagram of the proposed algorithm. Rule 3
The proposed algorithm comprises of three basic steps. If x is A3 then f3 ¼ p3 x þ r3 :

1. Acquire three-phase currents The various layers can be described as follows.

2. Calculate zero sequence current Layer 1. Fuzzification (membership): Here the trapezoid
3. Deduce fault type using ANFIS function is used as membership function and is specified by
four parameters {ai ; bi ; ci ; di } as follows

If inference results are zero, then this signifies an trapezoidðx : ai ; bi ; ci ; di Þ

unfaulted situation. However, if inference results are non-    
x 2 ai d 2x
zero, then the technique indicates both the presence of a ¼ max min ; 1; i ;0 ð5Þ
bi 2 ai di 2 c i
fault and fault type.
Fig. 7 shows the ANFIS architecture adopted herein and where the input x represents RMS values of phase currents
is somewhat similar to that shown in Fig. 1. It consists of and zero sequence current for the proposed algorithm, and,
five layers, but since the ANIFS based algorithm employed the membership functions used in ANFIS are depicted in
752 S.M. Yeo et al. / Electrical Power and Energy Systems 25 (2003) 747–758

Fig. 4. RMS values of zero sequence current for each fault type.

Figs. 8 and 9. It is apparent that the two types of membership of incoming signals is avoided
functions are very different from each other; {ai ; bi ; ci ; di } is
the parameter set and the parameters in this layer are O2;i ¼ O1;i ; i ¼ 1; 2; 3 ð6Þ
referred to as premise parameters.
Layer 2. Conjunction: Every node in this layer multiplies Layer 3. Normalization: This architecture has three rules.
incoming signals for multi-input. But since only one input is Each node in this layer calculates the ratio of the ith rule’s
adopted in the proposed algorithm (Eq. (6)), multiplication firing strength to the sum of all the rules’ firing strengths:

Fig. 5. Distribution for RMS values of zero sequence currents.

 S.M. Yeo et al. / Electrical Power and Energy Systems 25 (2003) 747–758 753

Table 4
Fault characteristics for each fault type wi
O3;i ¼ w

i ¼ ; i ¼ 1; 2; 3 ð7Þ
Fault type Phase voltage Phase current Zero sequence w1 þ w2 þ w3
current compare
with normal Layer 4. Defuzzification: Each node i in this layer is an
state current adaptive node with a node function
O4;i ¼ w

i fi ¼ w

i ðpi x þ ri Þ ð8Þ
SLG a Decrease Much increase Large
SLG b Decrease Much increase Large
is the output of layer 3 and {pi ; ri } is the parameter
where w
SLG c Decrease Much increase Large
set.
DLG ab Decrease Much increase Large Layer 5. Summation: The single node in this layer
DLG bc Decrease Much increase Large computes the overall output as the summation of all
DLG ca Decrease Much increase Large
incoming signals
DLL ab Decrease Much increase 0
DLL bc Decrease Much increase 0 O5;1 ¼ f ¼ ðw

1 xÞp1 þ ðw

1 Þr1 þ ðw

2 xÞp2 þ ðw

2 Þr2
DLL ca Decrease Much increase 0
3fG abc Decrease Much increase 0
þ ðw
3 xÞp3 þ ðw
3 Þr3 ð9Þ

HIF a Very little variation Very little variation Small The proposed algorithm developed herein consists of four
HIF b Very little variation Very little variation Small ANFIS for ‘a’, ‘b’, ‘c’-phase currents and zero sequence
HIF c Very little variation Very little variation Small current, respectively. The inputs to the ANFIS are RMS
values of three-phase currents and zero sequence current.

Fig. 6. Diagram of fault classification algorithm.

Fig. 7. ANFIS architecture.

754 S.M. Yeo et al. / Electrical Power and Energy Systems 25 (2003) 747–758

Table 5
Definition of outputs

Fault type ANFIS

Ia Ib Ic I0

SLG a 2 0 0 2
SLG b 0 2 0 2
SLG c 0 0 2 2
DLG ab 2 2 0 2
DLG bc 0 2 2 2
DLG ca 2 0 2 2
LL ab 2 2 0 0
LL bc 0 2 2 0
LL ca 2 0 2 0
3fG abc 2 2 2 0
HIF a 1 0 0 1
Fig. 8. Membership function for phase current. HIF b 0 1 0 1
HIF c 0 0 1 1
Unfaulted 0 0 0 0
The values of output signify four categories associated with
the ‘a’, ‘b’, ‘c’-phase and zero sequence currents. If any of
the outputs from the first three ANFIS (i.e. Ia ; Ib ; Ic ) is ‘1’ various fault conditions such as fault inception angle, fault
then this indicates an HIF, in the case of ‘2’, it is LIF, and in distance and fault types.
the case of ‘0’, there is no fault. For the fourth ANFIS (i.e.
I0 ) considered, ‘2’, ‘1’, ‘0’ signifies a ground fault for LIF,
HIF or a fault clear of ground, respectively, this criterion is 4.1. Results of the simulation under LIF
summarised in Table 5.
Figs. 10 –13 typify the outputs of the various ANFIS for
fault classification for single line-to-ground fault, double
4. Simulation and results line-to-ground fault, line-to-line fault, three-phase fault,
respectively. As expected, all the results show accurate
In this section, results illustrating the performance of the outputs for each of the fault conditions considered.
proposed algorithm are presented. Importantly, the whole process of fault detection and
Fault cases studied are described in Table 3. As classification is achieved in approximately 5 ms from the
mentioned before, the performance of the proposed time of the fault. A more comprehensive performance of the
algorithm is tested for the Korean model system under technique under LIF is summarised in Table 6.

Fig. 10. Result for single line-to-ground fault. (a phase, at 20% of line and
Fig. 9. Membership function for zero sequence current. at a fault inception angle of 908).
 S.M. Yeo et al. / Electrical Power and Energy Systems 25 (2003) 747–758 755

Fig. 11. Result for double line-to-ground fault (b; c phases, at 80% of line Fig. 13. Result for three-phase fault (at 80% of line and at a fault inception
and at a fault inception angle of 608). angle of 908).

4.2. Results of the simulation under HIF 4.3. Summary

As mentioned before, because of the randomness, Table 6 summarises the performance of the proposed
unsymmetry and low levels of fault currents, HIFs pose algorithm for all the faulted and unfaulted case studied. The
classification time is calculated as the interval time from
particular difficulty in detection and classification when
fault inception time to the time of classifying fault type, and
employing traditional protection techniques. But Fig. 14
error rate is calculated as:
graphically depicts the satisfactory performance of the
technique for a high impedance single line-to-ground fault error samples
under HIF. Again a more comprehensive performance is error rate ð%Þ ¼ £ 100 ð10Þ
total samples
summarised in Table 6.
As expected, the ANFIS technique also gives the correct As shown in Table 6, the proposed algorithm can classify
output (i.e. all 0s) under normal steady-state conditions, and fault types under LIF and HIF in less than half a cycle, and
Fig. 15 shows the result of the proposed algorithm for an has low error rate.
unfaulted situation. The output of the proposed algorithm carries two
meanings; one is detection of faults and the other is
classification of faults. In other words, if all of the outputs
are not zero, this means that a fault has occurred and also
implicit is the type of fault. The classification error means
that the output indicates an incorrect fault type.
The cause of classification error is based on the
transient phenomena. The fault current has been divided
into two components, a steady-state component and a
transient component. The steady-state component has the
frequency of the applied voltage, but shifted in phase by
the angle and the constant angle of the system
impedance, and with a magnitude that is determined by
the magnitude of the applied voltage and of the system
impedance. The transient component has two parts, one
that depends on the angle on the voltage wave at which
the fault is applied. The other component is a function of
the prefault current that is flowing at the instant the fault
is applied.
In the case of a multi-phase fault such as double line-
Fig. 12. Result for line-to-line fault (b; c phases, at 50% of line and at a fault to-ground fault and line-to-line fault, the prefault current
inception angle of 308). of each phase has different magnitude and phase angle
756 S.M. Yeo et al. / Electrical Power and Energy Systems 25 (2003) 747–758

Table 6
Summary table

Fault type Fault inception Classification time (ms) Error rate (%)
time (ms)
5.2 km (20%) 13 km (50%) 20.8 km (80%) 5.2 km (20%) 13 km (50%) 20.8 km (80%)

SLG (‘a’, 08) 45.8 3.9 4.4 4.7 0 0 0

SLG (‘a’, 908) 50.0 2.6 2.9 3.1 0 0 0
SLG (‘b’, 608) 45.8 2.6 2.9 3.4 0 0 0
SLG (‘b’, 308) 50.0 3.1 4.4 6.5 0 1.41 2.66
SLG (‘c’, 308) 45.8 2.6 3.1 3.4 0 0 0
SLG (‘c’, 608) 50.0 2.9 3.4 3.7 0 0 0
DLG (‘ab’, 08) 45.8 2.9 3.1 3.7 0 0 0
DLG (‘ab’, 908) 50.0 4.2 5.0 5.5 0.63 0.94 0.94
DLG (‘bc’, 608) 45.8 3.9 4.4 5.0 0 0 0
DLG (‘bc’, 308) 50.0 3.9 4.2 4.4 0.78 0.78 0.63
DLG (‘ca’, 308) 45.8 2.9 3.4 3.9 0.16 0.16 0.31
DLG (‘ca’, 608) 50.0 3.4 6.0 6.8 0 2.34 2.81
DLL (‘ab’, 08) 45.8 1.8 2.1 2.3 0 0 0
DLL (‘ab’, 908) 50.0 1.6 1.8 2.6 0 0 0
DLL (‘bc’, 608) 45.8 1.0 1.3 1.6 0 0 0
DLL (‘bc’, 308) 50.0 2.9 3.4 3.7 0 0 0
DLL (‘ca’, 308) 45.8 4.4 4.7 5.2 0 0 0
DLL (‘ca’, 608) 50.0 1.3 1.6 1.8 0 0 0
3fG (08) 45.8 1.6 1.8 1.8 0.31 0.16 0
3fG (908) 50.0 0.5 0.8 0.8 0 0 0
HIF (‘a’, 08) 45.8 5.0 5.2 5.7 0 0 0
HIF (‘a’, 908) 50.0 2.1 2.3 2.6 0 0 0
HIF (‘b’, 608) 45.8 2.3 2.6 2.9 0 0 0
HIF (‘b’, 308) 50.0 6.5 6.8 7.0 0 0 0
HIF (‘c’, 308) 45.8 6.5 7.8 8.3 0 0 0
HIF (‘c’, 608) 50.0 3.7 3.9 4.2 0 0 0
Normal – 0 0 0 0 0 0
Average – 3.1 3.6 4.0 0.07 0.21 0.27
Maximum – 6.5 7.8 8.3 0.78 2.34 2.81
Minimum – 0.5 0.8 0.8 0 0 0

from each other at fault initiation instant. Accordingly, still indicates that a fault has occurred and due to the
the transient current of each phase is different from each behaviour of the signals with time, the classification error
other. Such a difference of the transient current causes is only for a very short period and hence the impact of
the incorrect classification results. Importantly, the latter the classification error is not very significant. For

Fig. 14. Result for high impedance fault (a phase, at 20% of line and at a
fault inception angle of 08). Fig. 15. Result for an unfaulted system.
 S.M. Yeo et al. / Electrical Power and Energy Systems 25 (2003) 747–758 757

example, as shown in Table 6, a DLG fault (distance: [3] Wang H, Keerthipala WWL. Fuzzy-neuro approach to fault
80%) has the maximum error of 2.81% and this is classification for transmission line protection. IEEE Trans Power
Deliv 1998;13(4):1093 –104.
equivalent to just 18 samples i.e. about a quarter cycle of
[4] Keerthipala WWL, Wang H, Wai CT. Experimental validation of a
the fundamental frequency. After this initial period of fuzzy-neuro based fault classifier for transmission line protection.
error, the output of the algorithm starts to indicate International Conference Power Systems Transients, IPST ’99,
correct fault classification. Budapest, Hungary; June 1999. p. 549 –54.
[5] Agrawal P. An investigation into a method of detecting the fault
induced high frequency voltage signals of EHV transmission lines
for protection applications. IEEE Trans Power Deliv 1991;6(1):
5. Conclusions 119–26.
[6] Gilbert DM, Morrison IF. A statistical method for the detection of
Any fault detection and classification technique should power system faults. Electr Power Energy Syst 1997;19(4):
be capable of giving satisfactory performance under a 269–75.
wide variety of system and fault conditions, including [7] Djuric MB, Radojevic ZM, Terzija VV. Numerical algorithm for
arcing faults detection and fault distance calculation on overhead
contingencies such as HIFs; conventional techniques,
lines. Electr Mach Power Syst 1997;939–53.
however, have difficulties in dealing with such faults [8] Djuric MB, Radojevic ZM, Terzija VV. Arcing faults detection on
(they either fail to detect the fault or fail to discern transmission lines using least error squares technique. ETEP 1998;
between double line-to-ground and line-to-line faults) 8(6):437– 43.
principally due to the limitations imposed by the low [9] Fitton DS, Dunn RW, Aggarwal RK, Johns AT, Bennett A. Design
and implementation of an adaptive single pole autoreclosure
levels of fault currents.
technique for transmission line using artificial neural networks.
In this paper, a novel algorithm for fault detection and IEEE Trans Power Deliv 1996;11(2):748–55.
classification of LIF and HIF has been developed using [10] Aggarwal RK, Song YH, Johns AT. Adaptive three-phase auto-
ANFIS, which overcomes the aforementioned problem. reclosure for double-circuit transmission system using neural net-
Equally importantly, this algorithm can detect and classify works. IEE Second International Conference on Advances Power
fault type in a transmission line based on RMS value of System Control, Operation and Management, Hong Kong; Dec. 1993.
p. 389 –92.
phase currents and zero sequence current in one single
[11] Kim CH, Byun SH, Kim H, Kim ID, Aggarwal RK, Johns AT. A
algorithm. novel approach to detecting arcing faults in transmission lines using
The performance of the proposed algorithm is tested on a wavelet transforms. International Conference on Electrical Engineer-
typical 154 kV Korean transmission line system under ing, ICEE ’98, Kyongju, Korea; July 1998. p. 775–8.
various fault conditions and the results presented clearly [12] Hines JW. MATLAB supplement to fuzzy and neural approaches in
demonstrate that the proposed algorithm can detect faults engineering. New York: Wiley; 1997.
[13] Jang JSR, Sun CT. Neuro-fuzzy modeling and control. Proc IEEE
and classify fault type accurately, in less than about half 1995;83(3):378–406.
cycle. The proposed algorithm has been developed for real [14] Lee JB, Lee SJ, Kim CH, Lim HY. A study on the development and
time implementation and can be coded into existing digital relaying scheme under high resistance earth faults on HV, EHV line.
relay hardware. Electr Engng Sci Res Inst 1997;.
[15] Kim CH, Aggarwal RK, Johns AT. Digital simulation of the fault
transient phenomena on EHV transmission lines under non-linear
high impedance arcing faults. International Conference on Power
Acknowledgements Systems Transients, IPST ’99, Budapest, Hungary; June 1999. p.
164–8.
The authors would like to acknowledge the financial [16] Kim CH, Ahn SP. A study on the arc modeling in transmission lines
and other support for this project provided by using EMTP. International Power Engineering Conference, IPEC ’99,
the Electrical Engineering and Science Research Institute Mandarin Hotel, Singapore; May 1999. p. 52– 7.
[17] Kim CH, Choi HS, Kang SH, Aggarwal RK, Johns AT. A
(EESRI), and this work was sponsored by the neural network approach to the detection of high impedance
NPTC (Next-Generation Power Technology Center), faults in transmission networks. International Power Engineering
supported by MOST (Ministry of Science and Technol- Conference, IPEC ’99, Mandarin Hotel, Singapore; May 1999.
ogy) and KOSEF (Korea Science and Engineering p. 798 –803.
Foundation).

References
Sang-Min Yeo was born in Korea on July 21, 1976. He received his BS
degree in Electrical Engineering and MS degree in Electrical and
[1] Kim CH, Lee MH, Aggarwal RK, Johns AT. Educational use of
EMTP MODELS for the study of a distance relaying algorithm for Computer Engineering from Sungkyunkwan University, Korea, in 1999
protecting transmission lines. IEEE Trans Power Syst 2000;15(1): and 2001, respectively. He is now in the course of PhD at the
9–15. Sungkyunkwan University. His current research interests include
[2] Dalstein T, Kulicke B. Neural network approach to fault classification power system protection and computer applications using EMTP
for high speed protective relaying. IEEE Trans Power Deliv 1995; software and signal processing.
10(2):1002–11.
758 S.M. Yeo et al. / Electrical Power and Energy Systems 25 (2003) 747–758

Chul-Hwan Kim was born in Korea on January 10, 1961. He received Raj Aggarwal obtained the degrees of BEng and PhD in Electrical
his BS and MS degrees in Electrical Engineering from Sungkyunkwan Engineering from the University of Liverpool, UK, in 1970 and 1973,
University, Korea, in 1982 and 1984, respectively. He received PhD in respectively. He then joined the Power System Group at the University
Electrical Engineering from Sungkyunkwan University in 1990. In of Bath, where he is now a professor and Head of the Electrical power
1990, he joined Cheju National University, Cheju, Korea, as a full-time and Energy Systems Group of Electrical Power Systems. His main
lecturer. He has been a visiting research professor at University of Bath, areas of research interests are power system modelling and the
UK, in 1998, 1999 and 2001. Since March 1992, he has been a application of digital techniques and artificial intelligence to protection
professor in School of Electrical and Computer Engineering, and control. He has published over 250 technical papers and is a fellow
Sungkyunkwan University, Suwon, Korea. His research interests of the IEE UK.
include power system protection, neural network applications, the
modelling/protection of underground cable and the EMTP software.

Allan Johns received the degree of BSc and PhD from the University
Kwang-Seok Hong was born in Seoul, Korea, on February 1959. He
of Bath and in 1982 was awarded the degree of DSc for an original and
received the BS, MS and PhD degrees in Electronics from the
substantial contribution to knowledge of Electrical Engineering. He is
Sungkyunkwan University, Korea, in 1985, 1988, and 1992, respect-
currently emeritus professor of Electrical Engineering at the University
ively. In March 1990, he joined the Seoul Health College, Seoul, Korea,
of Bath. He is the author of over 300 publications in the area of
where he was a full-time lecturer of Computer Engineering. From
Electrical Power Systems and is a fellow of the IEE UK.
March 1993 to February 1995, he was a full time lecturer at Cheju
University, Cheju, Korea. Since March 1995, he has been an associate
professor at Sungkyunkwan University, Suwon, Korea. His current
research focuses on statistical signal processing, speech recognition,
speaker identification and power system application.

Myeon-Song Choi received his BS, MS, and PhD degrees in Electrical
Engineering from Seoul National University Seoul Korea in 1989, 1991
and 1996, respectively. He is currently an assistant professor of
Young-Bherm Lim was born in Korea on January 7, 1969. He received Electrical Engineering in Myongji University, Yongin, Korea. He is
his BS degree and MS degree in Electrical Engineering from currently a visiting professor at the Pennsylvania State University. His
Sungkyunkwan University, Korea, in 1993 and 1995, respectively. current research interests include power system protection, artificial
He works at LG Electronics CDMA Handsets Lab. His current research neural networks, expert systems in power system protection and
interests include power system protection and communication. control.