IEEE Transactions on Power Delivery, Vol. 11, No.

2, April 1996

748

and Implementation of an Adaptive Single Pole Autoreclosure Technique for

Transmission Lines using Artificial Neural Networks
D. S. Fitton R. W. Dunn
R. K. Aggarwal (SM) A. T. Johns (SM)
School of Electronic and Electrical Engineering, University of Bath,
Claverton Down, Bath, BA2 7AY, UK.

Keywords: transmission lines, single-pole tripping, adaptive

autoreclosure, artificial neural networks

Abstract Adaptive Single Pole AutoReclosure (SPAR) offers

many advantages over conventional techniques. In the case
of transient faults, the secondary arc extinction time can be
accurately determined and in the case of a permanent fault,
breaker reclosure can be avoided.This paper describes, in
some detail, the design and implementation of a SPAR
technique using Artificial Neural Networks ( A " s ) . The
design described includes special methods for extracting
features from post-circuit breaker opening fault data, which
is a prerequisite for setting up training data sets. The
technique is then implemented in hardware based on a high
performance T800 transputer system and some results
obtained from laboratory tests of this equipment are
presented.

1. INTRODUCTION
The most common faults on EHV transmission lines are single
phase to ground type and for such faults SPAR provides an
improvement in the overall operation of the transmission
system [l]. Furthermore, SPAR may become imperative in
applications where construction of additional circuits may not
be possible due to environmental pressures and/or costs.
However, with conventional SPAR, unsuccessful reclosure
using a fixed dead time in the case of a transient fault, or
reclosure onto a permanent fault, may aggravate the potential
damage to the system and equipment. In this respect, adaptive
SPAR offers many advantages such as increased rate of
successful reclosure,improved system stability and a reduction
in system and equipment shock under a permanent fault.
Previous studies [2] have shown certain characteristic voltage
waveforms that develop on the faulted phase during the
secondary arc period, following the initial circuit breaker
opening. In this respect there are many factors that affect
A,paper recommended and approved
95 SM 432-5 PWRD
by the IEEE Power System Relaying Committee of the
IEEE Power Engineering Society for presentation at
the 1995 IEEE/PES Summer Meeting, J u l y 23-27, 1995,
Portland, OR. Manuscript submitted November 15,
1994; made available for printing April 27, 1995.

A. Bennett
Reyrolle Protection, PO Box 8, Hebbum
Tyne and Wear, NE31 lTZ, UK.

these waveforms such as, line configuration, fault point on

wave, fault location, pre-fault loading, source parameters and
atmospheric conditions. The complexity of the functional
relationship between these various factors precludes a simple
algorithmic approach to the design of adaptive SPAR using
traditional methods.
There has been an upsurge in the application of A " s in
power systems in recent years which has clearly demonstrated
their ability in solving some long standing problems where
conventional techniques have had difficulty or have been
unable to meet functional requirements. However, their
application in the development of new improved adaptive
SPAR techniques is heavily dependent upon accurate models
depicting the behaviour of power systems under faults. In
this respect, the simulation models based on the well proven
and widely accepted Electro-Magnetics Transient Program
(EMTP) provides the necessary accuracy and realism. This
software also has a facility whereby a realistic nonlinear
secondary arc [3] can be embedded into the simulation. This
paper describes developments in the analysis, design and
implementationof an adaptive SPAR technique using A " s .
The technique, initially designed using Computer-Aided
Design (CAD) studies, includes a methodology adopted for
identifying and extracting the characteristic features from the
fault waveforms using Discrete Fourier Transform (DFT)
techniques.
The faulted waveforms are based on the
previously mentioned EMTP software with realistic modelling
of all the line components, including nonlinear arc
phenomenon and the effect of practical Capacitor Voltage
Transformers (CVTs). Characteristic features form the
training patterns for the chosen ANN architecture. The CAD
technique is then accomplished in hardware based on the
T800 transputer and the paper concludes by presenting some
interesting results to illustrate its performance. For the latter,
analogue voltage waveforms from a Programmable
Transmission Line (PTL) for practically encountered faults
on a typical 400kV application are considered.
2. ADAPTIVESPARCONCEPT

To clarify the objectives of using adaptive SPAR it is worth

considering all the possible eventualities. For clarity only
steady state comparisons are made here; some of the inherent
advantages which result from the reduced dead times of an
adaptive system are thus ignored. Quantification of potential
results is essential to enable a real assessment and discussion
of performance requirements and improvements. There are
three possible conditions to be considered:1 Permanent fault
(perm)
2 Transient fault
(tran)
0885-8977/96/$05.00 0 1995 IEEE

749

Table 1. THE OPTIONS AVAILABLE TO AN ADAPTIVE SPAR COMPARED TO THE CONVENTIONAL CASE.
(Shading denotes the ideal choice for the adaptive SPAR given the fault type.)

3 Incorrect trip
(inc.tr)
The third category of incorrect trip includes any trip which
has occurred for no legitimate reason, such as a sympathy
trip.
Experience of EHV transmission [4] shows that typically 85 %
of faults on EHV systems are transient in nature, and the rest
are permanent. The possibility of incorrect tripping must be
considered, and is often quantified in terms of dependability
and security [5]. A typical value of incorrect tripping as a
percentage of all trips might be 1 %.
The possible outcomes of conventional versus adaptive
protection, described in Table 1, are discussed using the
conditional probability notation:-

P(Action of circuit breaker IType of fault)

P(Action I Event)

(1)

P(event) = probability of event occurring

Pc
= probability in the conventional case
= probability in the adaptive case
Pa
R
= reclose

R"
NR
qC
TA

The result of this ideal case is to prevent shocks to the

system occurring in response to a permanent fault. However,
in the worst possible case, the adaptive technique will always
reclose onto permanent faults. As such, in this worst case,
the performance of the adaptive system is only equal to that
of the conventional system. The performance efficiency of an
adaptive technique for permanent faults can be described as:-

In summary, on encountering a permanent fault, an adaptive

SPAR can only provide an improvement or no change in the
performance of a conventional system because:-

(7)

B. Transientfault conditions
In the case of a transient fault, a conventional protection
system can either reclose successfully or reclose too soon,
resulting in unsuccessful reclosure. Thus: -

Pc(Rs I trM)+Pc(Ru I trun)=l

= reclose successful
= reclose unsuccessful
=

(5)

PJR berm) =O

(2)

Equation 1 is "the probability of 'Action of circuit breaker'

occurring given that 'Type of fault' has occurred". Using the
symbols:-

Rs

reclose onto a permanent fault, ie:-

The choice of dead time affects the efficiency of the SPAR

relay. The efficiency of the conventional system on
encountering a transient fault can be described as:-

no reclose

= conventional SPAR efficiency

= adaptive SPAR efficiency

(9)

A. Permanent Fault Conditions

First consider the case of a permanent fault. In the case of a
conventional SPAR system, the probability of a reclose
happening is equal to one for all classes of fault (excluding
the possibility of maloperation of the SPAR):-

P,(R Ipenn) = 1

(8)

(3)

In the adaptive case the dead time is adjusted in real time so

as to attempt to minimise the probability of an unsuccessful
reclosure. The adaptive circuit should, however, also
minimise the possibility of not reclosing onto a permanent
fault. The efficiency of an adaptive system can be described
~

as:-

In the adaptive case there is a choice of possible outcomes:-

Pa(R I penn)+Pa(NR I penn)=l

(4)

The ideal situation is that the adaptive SPAR system does not

The adaptive SPAR will be more efficient in response to a

transient fault than a conventional system i f -

750

Or: Ill

Transient fault (fig la)

Time

C. Incorrect Trip Conditions

If protection equipment causes an incorrect trip, such as a
sympathy trip, the circuit breakers will open on a healthy
line. In a conventional SPAR circuit breaker reclosure will
normally occur after a fixed dead time and the line is returned
into service. An adaptive SPAR introduces, in theory, ihe
possibility of choosing not to reclose onto the line. The
conventional circuit always performs the correct action thus:-

Permanent fault (fig I 6)

-90
FiguFe 1TYPICALTRANSIENTAND PERMANENTFAULT VOLTAGE

Where the efficiency of the adaptive system is given by:-

However the incorrect trip condition is mitigated by the fact

that incorrect trips are usually the easiest to identify and, as
such, adaptive SPAR can if necessary restore tripped lines
more quickly than the conventional fixed dead time case.

D.All Fault Conditions

In summary for all types of fault, two performance criteria
can be applied. First, and considered more important, an
adaptive SPAR scheme should increase the possibility of
successful reclosure. This i s true if the following inequality
from equations 6, 10 and 12 holds for all faults:-

fi@@.pa(Rs Itran)
+P(inc.tr).P,(R 1inc.tr)

.p(tran).P,(R, [Pun)
+ P(inc.tr)

Another criteria is that the adaptive technique should reduce

the possibility of shocking the power system by preventing
unsuccessful reclosure.
This occurs if the following
inequality is true for all fault types:-

WAVEFORMS

spacing, configuration, etc.

Transient faults are simulated using a realistic arc model, in
particular the secondary arc model which develops once the
faulted phase line breakers have opened [3].
In the
simulation of permanent faults, fault arc path impedances,
typically up to about 200 ohms, are used.
In any fault study voltages and currents seen at the end of a
faulted h e depend on a number of different system
parameters. These range from source and load parameters to
fault location, fault inception angle, pre-fault loading, breaker
opening time, etc. Thus in the generation of fault examples
all of these factors are varied.
The results presented here are for faults on a typical 400kV
transmission system of the type encountered on the UK
supergrid system [3]. The simulated voltage waveforms at
the ends of the faulted line are modified by the CVT,
analogue anti-aliasing filters and the quantisation process. It
is vital that these band limiting and other effects are included
in the CAD because of their significant influence on the
primary system voltage.
Fig. 1 typifies the simulated voltage waveforms measured on
a faulted phase via a CVT for both transient and permanent

fadts.

3. SYSTEM
SKMULATION
Due to a limited amount of practical fault data available, it is
necessary to generate examples of fault waveforms using
simulation. Simulations were carried out using the EMTP [6]
software package, which can model the behaviour of a
transmission line. The physical line construction must be
described to the software, together with conductor type,

For the particular transient case, shown in Fig. la, at the

point marked A on the waveform a fault develops on the line
and there is a subsequent reduction in the voltage. The
protection system detects the fault and opens the circuit
breakers at p i n t B. A secondary arc is then established and
this can be seen extinguishing and re-striking, by the
characteristic high frequency components in the waveform.
Finally the arc extinguishes completely at point C . There
remains a small system frequency voltage sinusoid component
on the line after point C , which is due to electrostatic
coupling between the faulted phase and the two healthy

75 1
phases. After point C there is a DC offset on the line which
is due to the stored charge at the point on wave at which the
arc finally extinguished. In practice this primary voltage DC
offset is attenuated by the effect of the CVT, but a lower than
fundamental frequency surge can be seen.
A permanent fault voltage waveform is shown in Fig. lb.
The fault occurs at point A, and the protection trips the circuit
breakers at point B. After the circuit breakers have tripped
there is a small system frequency voltage induced onto the
tripped phase. The magnitude of this voltage depends on how
well the other two healthy phases are coupled to the faulted
phase, and the fault impedance.

4. FEATURE
EXTRACTION
METHODS
ANN solutions often employ a pre-processing stage of feature
extraction. The aim of feature extraction is to extract
information from the data, in this case the fault simulation
voltage waveforms, which can be used to identify the data as
required. The data should be rendered into a form which
makes the ANN more effective at making decisions and easier
and faster to train. Typically this means that the amount of
data is reduced and the data undergoes some form of a
transformation.
The feature extraction process takes
information from the time domain voltage data which allows
the correct decision to be made as when, if ever, to reclose.
To examine the voltage waveforms a frequency decomposition
approach was taken. This approach was adopted because the
frequency spectra of the voltage waveforms v a + s with time,
and this time varying spectra is indicative of the conditions on
the line. The frequency decomposition method transforms the
voltage magnitude against time, x(t), data into the
voltage/frequency/time domain, X E t). There are many
different methods for performing this transformation, many of
which are discussed in [7].
A method called the Short Time Fourier Transform (STFT)
[8] was employed in this research to examine the waveforms
and aid the choice of a feature extraction scheme. This
process uses a DFT on a window of data of length N points,
then moves the window of data on by a number of points M,
and then repeats the DFT. The output of this process is a
series of frequency spectra from consecutive windows in the
time domain.
Using the STFT, the waveform is analyzed by taking a series
of N point DFTs at intervals of MT,sec. The first window is
calculated using an N point DFT; then the reference point is
moved along M points and the process is repeated. If r
represents the window number then the rth discrete frequency
spectra is given by:-

Figure 2 3D PLOT

OF VOLTAGE MAGNlTUDE vs. FREQUENCY vs

TIME

was examined, certain characteristic behaviour was apparent.

There is more high frequency energy while a secondary arc
exists than when it has extinguished. The constant impedance
permanent fault waveforms contain a small system frequency
component only. Post arc transient fault waveforms contain
a system frequency component usually larger than that of the
permanent fault. The points A, B and C as described in
section 3 are indicated in Fig. 2.
A feature extraction method which measures the energy in
five different frequency bands was used to implement the
adaptive SPAR. The frequency bands chosen were selected
by an iterative process of empirical and theoretical selection
and subsequent testing with the ANN [9]. The chosen
frequency bands were:
1 DC-20 HZ
2
30-70 HZ
3
80-120Hz
4
130-170Hz
5 170-220 HZ
Each frequency domain snapshot of the system, r, can be
associated with a particular desired outcome; "safe to
reclose", or "do not reclose". Where a state transition
occurs within a window, there is an ambiguity about which
class the window is a member of, and as such these cases are
omitted from the training set. Every other frequency domain
feature set is combined with its desired outcome to produce
a training set suitable for training the ANN.

5. ARTIFICIALNEURALNETWORKS
The result, XJk,r), is a function of both frequency, ( k Q ,
and time, (rMTJ, and renders a 3D spectral portrait,
examples of which can be seen in Fig. 2.
When the frequency spectra for different waveforms of data

The problem addressed here is essentially one of pattern

recognition. Given a section of a waveform from a faulted
transmission line, a decision must be made as to when, if
ever, the faulted phase can be reclosed. One of the most

752

Hidden Layer

D.C.

__

50Hz

__

100Hz150Hz-

200Hz-

Figure 3 NEURAL NETWORK ARCHITECTURE

effective types of ANN for pattern recognition is the feed

forward multi-layer type network, trained using the error back
propagation method (also known as the generalised delta
rule). This type of ANN was chosen in this study because
feed forward networks are fast at making decisions and can
be trained 'off line' by presenting examples.
The ANN is made up from a number of simple processing
elements or nodes, which are connected to other nodes by
weights. For a node with N inputs, xi,and weights, wi , a
bias 8 , and a firing rule F(z), the output Y is given by:N

Y=F( ~ x i w i + O)
i=o

Y is a value which lies in a range determined by the transfer

function, in this case the hyperbolic tangent tanh(x). The bias
allows the transfer function to be shifted and can be
considered as a weighted signal from an output which is
always one for training purposes.
The connections between the nodes have weights associated

with them which amplify the signal travelling along theh. By

adjusting these weights, the network learns to provide the
required output for a given input.
In the feed forward architecture, the nodes are grouped in
layers with the outputs from nodes in one layer being
connected to the inputs of nodes in the next. The connections
are unidirectional, so that a signal is presented to the input
layer, and propagates through the hidden layer(s) to the
output layer. The nodes in the input layer have a linear
transfer function, not a nonlinear firing rule, and act as a
scaling stage to the inputs data. Scaling also occurs at the
output node to make the output to lie in the range of 0 to 1.
The ideal network architecture depends on the problem being
addressed. The number of inputs to the network is detennined
by the features chosen. Having one or more hidden layers in
the ANN topology allows the network to make more complex
associations between input and output [lo]. The number of
nodes required in the hidden layer(s) depends on the
complexity of the relationship between the inputs and outputs.
There is no defrnite way of pre-determining the optimum
network size and architecture, without trying and testing
different configurations.
A neural network with five inputs, corresponding to the
energy in five frequency bands was used. The network had
one hidden layer and an output layer with one node. The
number of nodes in the hidden layer was varied and good
performance on the test set was obtained with between five
and nine nodes. The results here use a network with seven
nodes in the hidden layer, and the layers are fully connected,
as shown in Fig 3.
In summary ANNs are good at performing pattern
recognition, can cope with previously unencountered
situations and are robust in the presence of noise. Also,
because the processing information of the network is stored
in a distributed manner among the weights, neural networks
can still function, albeit with some degradation, if parts of the
network are malfunctioning.

5. PRACTICAL
IMPLEMENTATION
A prototype hardware has been built by Bath University in
conjunction with Reyrolle Protection Ltd (RP), illustrated in
Fig. 4. The prototype is based on an existing RP modular
protection hardware design, with the addition of a new

Faulted phase
Eaable and Phase Selection
I

...._____.....
.........................

.. ....___.____.....
.. ..

.. ....

N&
Network
Figure 4 BLOCK DIAGRAM OF ADAPTIVE SPAR CIRCUIT

Trip and
lock out

753

transputer based processor card. The neural network and

feature extraction software processes and assorted control
logic run on the transputer card.
The adaptive SPAR prototype samples three voltage phase
waveforms at 600Hz and digitally selects the faulted phase,
from programmed status inputs. The most recent N samples
one full cycle at the power frequency are passed to the feature
extraction process, which measures the energy contained in
the five frequency bands. These features are passed to the
ANN input layer which scales the values to fall within a
suitable range for the activation function used. The ANN
calculates a scaled output value in a range of 0 to 1, that
indicates the class in which the network considers the input to
be. The window of data from which the features are
calculated is then moved on by M points, as in equation 18,
in this case 1 point.
In a practical system the output of the network would be
compared with a threshold value over some period of time,
ie. some number of consecutive outputs of the network, and
then a decision would be made whether to reclose, or trip the
two healthy phases and lock out. In this prototype, an input
is presented to the network and an output generated every
twelfth of a cycle (1/600sec). The values of N and M
(equation 18) used were 12 and 1 respectively.
The 12 bit analogue to digital converter uses filters to prevent
aliasing around the sampling frequency with a cut off
frequency of around 230Hz. Included in the prototype is a
timer which causes the relay to trip the two healthy phases
and lock out if reclosure is not initiated within a prescribed
time.

6. RESULTS
OF ANALOGUEAND DIGITAL TESTS

In the engineering of an ANN solution to a problem, it is

essential to test the network on previously unseen data. The

output from the ANN software was measured over a test set
of 40,000training cases.
Analogue tests were carried out from voltage signals
generated by a PTL. These voltage signals were at nominally
63.5V rms, the substation voltage which the circuit would
expect to meet in practice under full system voltage. The
PTL was driven by simulated waveforms from the EMTP
software package. The transputer processor has hardware
links which provide a channel through which results were
taken from the adaptive SPAR in real time operation.
The CAD modelling of the input transformers, filters and the
quantisation process were proved by comparing theoretical
and actual input circuit responses.
The output of the network was trained to be 0.1 to indicate
do not reclose and 0.9 to indicate safe to reclose. The actual
output of the network varies between zero and one.
The output of the ANN in the relay in response to a typical
transient fault voltage is shown in Fig. 5. It can be seen that
the NN output switches from low to high when a full cycle of
post secondary arc waveform is encountered.
A typical permanent fault is shown in Fig. 6 with the
corresponding output from the neural network.
The
permanent fault illustrated here is a 100 ohm bolted fault and
it is seen that, as required, the ANN output remains at the
desired low level.
A graph of the normalised density of the output of the
network over the test set is shown in Fig. 7. This displays
the number of outputs of the ANN over the range of all the
possible values, normalised to unity area. It can be seen
from Fig. 7 that there is a wider spread of values around the
high condition than around the low condition.
In general the errors that occurred are grouped around
transitions of state in the behaviour of the tripped phase. For

Volts (v.)

Volts (V.)

Network Output
/&nu\

Time

r"\

Figure5 OUTPUT RESPONSE OF NEURAL NETWORK TO A

TRANSIENT FAULT VOLTAGE WAVEFORM

Figure6 OUTPUT RESPONSE OF NEURAL NETWORK TO A

PERMANENT FAULT VOLTAGE WAVEFORM

,754
well, with the only significant errors appearing during the
final extinction of secondary arcing. With the addition of the
constraint that a contiguous number of outputs of the neural
network would be examined before a decision was made, the
technique correctly classified all of the simulated examples.
The simulations consisted of various configurations of line
and system. The introduction of low levels of noise into the
voltage waveforms did not significantly degrade the
performance of the neural network.
The next step is to obtain field data to test the adaptive SPAR
technique with real world data. Another reason to collect
data is to measure the noise characteristics on the system and
to include these in the CAD studies.

g 0.3
3

0.2

.
3

0.1
0

8 ACKNOWLEDGEMENT

0.2

0.4

0.6

0.8

output value
Figure 7 NEURAL NETWORK OUTPUT DENSITY OVER TEST SET

example, in Fig. 5 it can be seen that the A " ' s output

change from the low state into the high state is not immediate.
The band limiting effect of the CVT causes a phase delay in
frequency components which compounds these transitional
errors. For this reason the output of the ANN is compared
to a threshold over a number of contiguous outputs, P,before
a decision is made. The constraint that the network output
must be above a threshold for P contiguous samples makes
the decision to reclose more robust.
With P set to the period of the fundamental frequency and a
threshold of 0.5, the adaptive SPAR always correctly
classified the fault type and instigated the correct action; thus
from equations 6, 10 and 15:-

P(NR bm)= 1
P(R, Itran)= 1
p(R I inc.tr)= 1

(20)

This was measured digitally over the whole of the test set of
simulated faults. Due to the time that would be required, not
all of the 40,000 cases of the test set were generated as
analogue voltages. The waveforms which were generated
gave results that were always within 3% of the digital tests,
and performed equally well when using the threshold over
multiple cycles criteria.
The above results were taken in the absence of noise. The
effect of small amounts of gausian noise is to increase the
spread of the output values, without degrading the
performance of the adaptive SPAR circuit.

7. CONCLUSIONS
In this paper the results of performing analogue tests on an
adaptive SPAR prototype are presented. The adaptive SPAR
prototype was tested on sifnulations of previously
unencountered faults and the technique proposed performed

The authors would like to acknowledge the financial and

other support for this project provided by the Engi.neering and
Physical Science Research Council (EPSRC) and Reyrolle
Protection Ltd. and the facilities at the University of Bath.

9 REFERENCES
[l] IEEE Committee report "Single phase tripping and auto
reclosing of transmission lines "IEEE Trans. PWRD-1,
pp. 182-192, 1992.
[2] Y.H.Song,
R.K.Agganva1, A.T.Johns,
"Digital
simulation of the effects of fault and circuit breaker arcs
on power system fault transients" Proc. 27th UPEC
(UK) 189-192, 1992
[3] A.T.Johns, R.K.Agganva1, Y.H.Song, "Improved
technique for modelling fault arcs on faulted EHV
transmission systems", IEE Proc. Gener. Trans. Distrb.
1994, pp 148-121
[4] A. R. van C. Warrington. "Protective Relays: Theory
and application", 1977, J. Wiley and sons, ISBN 0 412
15380 7
IS] J,Lewis Blackburn, "Protective Relaying: Principles and
applications pp 21-22 ". 1987 Marcel Dekker, ISBN. 08247-7445-0
[6j H. W. Dommel, "Electro magnetics transients program",
August 1986 Boneville Power Administration, USA.
[7] Riou1,A. Vetterli,M. "Wavelets and signal
processing"JEEE Signal Processing magazine,pp 14-38
October 1991.
[81 Hlawatsch,F. Boudreaux-Bartels,G. F. "Linear and
quadratic time-frequency signal representations" IEEE
Signal Processing magazine,pp 21-67 April 1992.
[9] Fitton,D.S. Dunn,R.W. Agganva1,R.K. Johns,A.T. Li,
H.Y. "Feature extraction from voltage and current
waveforms." Proc 29th UPEC (UK) 1994
[lo] J. A.Anderson, E,Rosenfeld, (Editors) "Neurocomputing:
Foundations of research", pp 673-696 1989, MIT Press,
ISBN 0-262-01097-6

David Fitton received the degree of BEng in 1991 from the

University of Bath. He is currently working for his Phd at

755
the University of Bath in the Power and Energy Systems
Group. His main areas of interest are Artificial Intelligence
and Neural Networks applied to control classification and
prediction.

Rod Dunn received his BSc and PhD in Electrical

Engineering from the University of Bath in 1981 and 1984
respectively. He became a lecturer in computing and control
at the University of Bath and is now a member of the Power
and Energy Systems Group. His research areas include
parallel and real time computing, power system modelling and
control using AI methods. He has published over 30
technical papers and is involved in the IEE UK.

Rqj Agganval obtained the degrees of BEng and PhD from

the University of Liverpool, UK, in 1970 and 1973
respectively. He then joined the Power Systems Group at the
University of Bath,where he is now a Reader. His main areas
of interest are power system modelling and the application of
digital techniques and artificial intelligence to protection and
control. He has published over 130 technical papers and is an
active member of the IEE UK.

Allan Johns received the degrees of BSc and PhD from the
University of Bath and in 1982 was awarded the degree of
DSc for an original and substantial contribution to knowledge
of Electrical Engineering. He is currently Professor of
Electrical Engineering and Head of School of Electronic and
Electrical Engineering at the University of Bath. He is the
author of over 200 publications in the area of Electrical
Power Systems and is a fellow of the IEE UK.

Alan Bennett is currently the research manager of Reyrolle

Protection and has attained considerable experience of
protection design during his 30 years association with the
company. He is the UK representative on CIGRE 34.02
Adaptive Protection working group.

Discussion

correctly right at the outset, then it generalises for all situations;

this effectively means that it is then capable of always recognising
the key features in the waveforms. The technique described in this
paper
has been tested for a very large number of practically
S. H. Horowitz (3143 Griggsview Court, Columbus, OH
encountered different system and fault conditions, including using
43221-4612):
real data gathered from a practical transmission system of a totally
The authors have described a unique and potentially powerful diagnosticto
different line configuration (including a different voltage level)
prevent autoreclosureinto a single phaseto-ground fault. I would like to mgg& that the.
authors have restricted the value of this technique. The tool is more powerhl than
from that for which the NN was originally trained; this technique
described. Sigle phase faults, either initially or upon reclosing, do not present a v e q
gave correct performance in all cases tested, thereby re-enforcing
serious threat to system stability or to equipment integrity. I appreciate that the authors
the previous point that this technique is very robust and is virtually
are concerned with single phase tripping which is a common practice.in Europe but
insensitive to changes in system and/or fault conditions.
suggest that ifthe authors extended the technique to considerthrespbase tripping they
would add another dimension to the value of their concept. This discusserCO-authored a
paper [l]which describes three phase trips, (a more common practice in theU.S.) and
single phase reclose which would test the system to determine S a permanaa fault exis&.
The technique descnied in reference 1 used voltage measurements made after one phase
was energized. We recognized several limitations at that time; transposed lines offered
some di5culties, calculations had to be made for each situationto establishthe voltages to
be measured. Is the authors techniques more general and could it be applied in every
situation without individual studies? If so, then the advantage of never reclosing into any
fault more severe than single phase-to-ground or phase-to-phase would offer very
significant advantages to system planning and operation In the US.,with the associated
long lines, single phase tripping usually requires shunt reactor compensation to extinguish
the secondary arc. Hive the authors examined the waveforms that would result with the
use of such shunt reactors?
Reference 1-Adaptive Transmission System Relaying, S.H. Horowitz, A.G.Phadke,
J.S.Thorp, EEE Trans Power Delively,vol3, no.4, pp1436-1445, Oct,1988.

Manusdript received August 2, 1995.

D.S. FITTON, R.W. DUNN, R.K. AGGARWAL, A.T. JOHNS,

A. BENNETT: The authors would like to thank the discusser for

his interesting comments and questions.

Whilst it is true that single-pole tripping associated with singlephase-earth faults is less onerous than three-phase tripping from a
system stability point of view, however, the main attribute of the
neural network-based technique described in this paper is in its
ability to inhibit breaker reclosure altogether when the fault is
permanent, thereby preventing a second shock to the system and
expensive equipment; in practice of course, for faults other than
single-phase-earth, all three phases are tripped. Although not
discussed in this paper, an extensive series of studies have revealed
that if that were to be the case ie, if all three phases were to be
tripped for faults other than single-phase-&, then this technique
would still be able to correctly distinguish between transient and
permanent faults. Moreover, in the case of a transient fault, the
secondary arc extinction time (albeit very short in comparison to
that associated with single-pole tripping) would also be accurately
identified.
With regard to the question of whether the technique is generic
enough in terms of its application to different situations without
individual studies, a neural network OUN) is a pattern classifier and
if the topology and the training pattems have been chosen

We fully agree with the discusser that the technique described in

this paper is potentially very powerful if it could also be applied to
transmission systems employing three-phase autoreclosure for all
types of faults, including single-phase-earth (this is the standard
practice adopted in many countries in order to essentially simplify
circuit breaker mechanism). We at Bath had already identified this
potential benefit quite some time ago and there has been ongoing
work by the authors in this very area; the investigations carried
out so far look very promising and it is hoped that the results of
this research will be published in another paper at a later.
Concerning the application of the -based adaptive autoreclosure
technique to long distance transmission lines employing shuntreactor compensation and single-pole autoreclosure, it is well
known that the voltage waveform patterns associated with such
systems under faults (in particular single-phase-earth faults) are
very distinctly different from those encountered on short plain
feeders. In the case of the former, characteristic low frequency
bearing phenomenon is apparent and is due to the fact that in shunt
compensated lines, the recovery voltage (after the faulted phase
breakex opening) contains a sinusoidal power frequency component
which is modulated by a much lower frequency voltage component
associated with the natural frequency of the line capacitance/reactor
inductance combination; the NN topology and the training patterns
are therefore somewhat different from those employed in plain
feeder applications; these have been discussed in some detail by the
authors in reference [l] which also clearly illustrates that a NN can
be used as an effective strategy in the development of adaptive
autoreclosure schemes for long distance transmission lines with
shunt reactor compensation.
Reference 1 YH Song, RK Aggarwal, AT Johns, RW Dunn, DS
Fitton: Adaptive autoreclosure technique for longdistance compensated transmission systems using a
neural network approach, Proceedings 28th
Universities Power Engineering Conference, Vol. 1
September 1993, pp 146-149.
Manuscript received October 17, 1995.