Académique Documents
Professionnel Documents
Culture Documents
`` |
Figure 3: Single line to ground fault simulated at bus 1.
esistaoce-eactaoceoiagac|o|au|tiozooeT o|toepotecteo|ioe
4
_icpeoaocetajecto,
---
|
----
|
----
'
-
ce|oeaooa|te|au|t
,::
' ' '
l
' '
| | ' ' '
.::
I
I
l
_
_
::
:: 1 I l I l
|
.:: - - - - - - I- - - -l- - - - - - - -I - - - - - - - - -
|
,
. ..
I
I , ,
... ::..
-
Figure 4: Fault trajectory on R-X plane.
III. FILTERING (PRE-PROCESSING) :
The pre-processing stage can signifcantly reduce
the size of the neural network based distance relay, which in
tu improves the performance and speed of the training
process. The fault voltage and current signals are ofen
noisy. In addition, when a fault occurs on a transmission
line, voltage and current signals develop a decaying DC
offset component whose magnitudes depends on many
factors that are random in nature. Thus, the input data
should be pre-processed before being fed to the network.
The block diagram of a typical numerical relay
fltering scheme and it's realization in MATLAB is shown
in fg. S and fg. 6 [12].
Input signal --Low pass flter --AID Converter
Discrete Fourier <Digital flter <
Figure 5: Block diagram of a typical numerical relay fltering scheme
Figure 6: Implementation of fltering scheme in MA TLAB
The signal is passed through low pass flter to remove
the effects, on the voltage and current signals, of the
travelling waves instigated by the fault. The input fltered
signals then passed through AID convertor. The output
signal becomes ready to be used by the Discrete Fourier
Transform. Here complete cycle discrete fourier transform
is used.
The discrete Fourier transform (DFT) is a digital fltering
algorithm that computes the magnitude and phase at discrete
fequencies of a discrete time sequence. Fast Fourier
transforms are computationally efcient algorithms for
computing DFTs. FFTs are usefl if we need to know the
magnitude andlor phase of a number individual or band of
fequencies. The DFT is ideal method of detecting the
fndamental fequency component in a fault signal.
IV. ANN BASED PROCESSING
In this section work done in implementation of ANN
method in the feld of distance protection is discussed.
Once trained, a network response can be, to a degree,
insensitive to minor variations in its input. This ability to
see through noise and distortion to the patter that lies
within is vital to patter recognition in a real world
environment [13]. Fig. 7 shows a simple model of a neuron
characterized by a number of inputs Pl,P2, . . . , Pn, the
weights WI, W2, .... Wn, the bias adjust b and an output a.
The neuron uses the input, as well as the information on its
current activation state to determine the output a, given as in
equation (1),
o__
W
k
P
k
+b
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (1)
ISBN: 978-81-909042-2-3 2012 IEEE
IEEE - International Conference On Advances In Engineering, Science And Management (ICAESM -2012) March 30, 31, 2012 592
Figure 7: Perceptron representation
The neurons are normally connected to each other in a
specifed fashion to form the ANNs. These arrangements of
interconnections could form a network which is composed
of a single layer or several layers. The ANN models must be
trained to work properly. During training each input vector
is assigned a particular target value. The algorithm adjusts
weights so that the output response to the input patters will
be as close as possible to the respective desired response. In
other words, the ANNs must have a mechanism for learing.
Learing alters the weights associated with the various
interconnections and thus leads to a modifcation in their
strength.
Literature shows that most of the ANN based schemes
use 2 hidden layers structure with backpropagation
algorithm and Levenberg-Marquardt (ML) algorithm for
training purpose. However, in this paper only 1 hidden layer
is used and fom the results it can be seen that desired neural
network response is obtained. The backpropagation
algorithm applied is discussed in brief in this paper.
.. 10.ha.k0aa||00 m.|00J.
The backpropagation algorithm is central to much
current work on learing in neural networks. The
backpropagation method works very well by adjusting the
weights which are connected in successive layers of multi
layer perceptrons. The algorithm gives a prescription for
changing the weights in any feed-forward network to lear a
training set of input-output pairs. The use of the bias adjust
in the ANNs is optional, but the results may be enhanced by
it. A multilayer network with one hidden layer is shown in
fg. 8.
X
Figure 8: Multilayer Perceptron representation
This network consists of a set of N input units (Xi, i
= 1,,,. N), a set of p output units ,Y _, p= 1,,,.,P) and a set of
J hidden units (Vj, j = 1 ,... J). Thus, the hidden unit Vj
receives a net input and produces the output:
where j=I, ...,J
Final output is then produced
YPF{ I
,
`pm`m }
(3)
where p=l, . . . . . p.
F[. ] is a non-linear transfer fnction which can be
of various forms. Backpropagation networks ofen use the
logistic sigmoid as the activation transfer function. The
logistic sigmoid transfer function maps the neuron input
fom the interval (-0,+0) into the interval (0,+1). The
logistic sigmoid, shown in (4), is applied to each element of
the proposed ANN.
Ij.|ljij|n.)1|1c|-|n)) )
Where, n - summation output& -bias adjust
The usual error measure or cost fnction for the process
is:
22)Y
i)
and now becomes,
This is clearly a continuous differentiable function
of every weight, so we can use a gradient descent algorithm
to lear appropriate weights. Since the weight errors are
successflly back-propagated fom the output layer, this
specifc training algorithm 1b known as error
backpropagation.
V. DISTANCE PROTECTION TPLEMENT A TlON USING ANN
This section describes how a neural network can classif
faults in a particular zone, once it is tained properly. The
results show how an ANN based relay distinguishes itself
fom electomechanical relay. The ANN relay is supposed to
identif the zone in which fault has occured correctly even
if the fault involves resistance and resultant tajectory settles
outside the coresponding zone, having the magnitudes of the
resistance and reactance corresponding to the postfault
fndamental fequency as inputs.
Concering the ANN architecture, parameters such as the
number of inputs to the network as well as the number of
neurons in the input and hidden layers were decided
empirically. This was done by observing network response to
various confgurations.
To create a feed-forward network suitable for back
propagation algorithm 'newff instruction fom MA TLAB
ISBN: 978-81-909042-2-3 2012 IEEE
IEEE - International Conference On Advances In Engineering, Science And Management (ICAESM -2012) March 30, 31, 2012 593
[14] is used. It provides fexibility to vary the nwnber of
hidden layers and neurons in particular hidden layers. It also
allows user to change the taining algorithm easily.
In the algorithm, only 1 hidden layer is used. It has 30
neurons with Logsig as activation fnction. A purelin
activation fnction is used for outut layer. The network is
trained by Levenberg-Marquardt algorithm.
The instruction newff is used as shown below;
net newff (p,t,[30], { , logsig , , ' purelin } , ' tainlm ' ,
'leargdm')
VI. RESULTS
While taining the network, patters coresponding to
various conditions such as fault resistance, fault initiation
time, series compensation, etc. are used (Table I). Target
vector is assigned value 1 or 0 according to the network
condition. Threshold is set at 0.5, i. e. , values above 0.5 are
treated as 1 and values below 0. 5 are teated as O. Once
performance goals are met, an unknown patter is applied to
verif whether the network is trained properly or not. It is as
tabulated below:
TALE!. TRIIG PATTENkS
^clWorkcondlon
npulpallcrnnamc Assigncd largclvaIuc
[I;
No fault with diferent
C_NFI, C_NF2
0;
percentage of Series
,
0;
compensation
C_NF3, C_NF4
0]
C_SLGJI,
Single line to ground fault
C_SLGJ2,
[0;
in zone I with diferent
C_SLGJ3,
I;
fault initiation timings
C_SLGJ4,
0;
and fault resistance
C_SLGJ5,
0]
C_SLGJ6,
C SLG F7
C_SLGJ21,
Single line to ground fault
C_SLGJ22,
[0;
in zone 2 with diferent
C_SLGJ23,
0;
fault initiation timings
C_SLGJ24,
I;
and fault resistace
C_SLGJ25,
0]
C_SLGJ26,
C SLG F27
C_SLGJ31,
Single line to ground fault
C_SLGJ32,
[0;
in zone 3 with diferent
C_SLGJ33,
0;
fault initiation timings
C_SLGJ34,
0;
and fault resistance
C_SLGJ35,
I]
C_SLGJ36,
C SLG F37
Out of these patters, some patters are used for training
the network using Levenberg-Marquardt algorithm. The
unknown patters are used for testing purpose. The training
and testing results are as follows:
.. 1a|0|0 J.|a||\.
The training QUI invoked which shows the number of
epochs required and time of training is as shown in fg.9.
Fig. 1 0 shows the performance plot during training process.
Training stops when desire goals related to error are
achieved.
to|WIwem
Tan|q. lrvnbq-Ma:zudt |uainl::
rdencz meansquated MO |n.e;
0n 0wiom 8dem |c u1
Hu
j Pmrmsncc | tp|uqo\
1tnin9te | tpmui :.|+te
kgmtior tplottie. DN?
Iot1--t: J
t~h.
Figure 9: Multilayer Perceptron representation
-s _
1
I
mT
uu
.
_
Figure 1 0: Performance plot
D. 1.\||0 .\u||\.
While testing the network, patters not used during
training are used. Results are tabulated below (Table II):
TALE IT. 1ESTINGKESLLTS
^clWorktondlon npulpallcrn RcsuIlanl0ulpul
namc vaIuc
No fault with diferent
1.0731
percentage of Series C NF2
0.4791
-0.1963
compensation
0.0413]
Single line to ground fault [0.0144
in zone 1 with diferent
C_SLGJ2
0.9813
fault initiation timings 0.0008
and fault resistance -0.0383]
ISBN: 978-81-909042-2-3 2012 IEEE
IEEE - International Conference On Advances In Engineering, Science And Management (ICAESM -2012) March 30, 31, 2012 594
^clWorktondlon npulpallcrn RcsuIlanl0ulpul
namc vaIuc
Single line to ground fault [0.0136
in zone 2 with diferent
C_SLGJ25
0.0079
fault initiation timings 0.9922
and fault resistance -0.0002]
Single line to ground fault [-0.0017
in zone 3 with diferent
C SLG F36
-0.0027
fault initiation timings 0.0002
and fault resistance 0.9947]
It can be seen that, depending on input patter conditions,
the corresponding target neuron has a value close to 0 or l.
VII. CONCLUSION
The use of an 1^ as a patter classifer to improve the
performance of distance relay is discussed in this paper.
Results show that LM method requires less epochs but
longer training period. The developed neural network is able
to detect whether the patter corresponds to fault or no fault
condition. In addition to this, if there is a fault condition,
network is also able to deterine the zone of fault. It also
has the advantage that it can adapt itself with the changing
network conditions with some additional training.
REFERENCES
[I] A P. Apostolov, D. Tholomier, S. H. Richards, "Distance
Protection and Dynamic Loading of Transmission Lines"
Power Engineering Society General Meeting, 2004, Vol. I pp.
100-105.
[2] Erezzaghi M. E. , Crossley P. A., "The Effect Of High
Resistance Faults On A distance relay" Power Engineering
Society General Meeting 2003, Vol. 4,pp. 2128-2133.
[3] H. Khorashadi-Zadeh, M. R. Agha Ebrahimi, "An ANN based
Approach to Improve the Distance Relaying Algorithm"
Proceedings of the 2004 IEEE Conference on Cybernetics and
Intelligent Systems Singapore, 1-3 December, 2004, pp.13 74-
1379.
[4] Sahoo. S.: Rav. P.: Panigrahi. B.K.: Senrov. N.: " A
Computational Intelligence Approach for Fault Location in
Transmission Lines"
Power Electronics. Drives and Energy Systems (PEDES)
20 I0 Power India, 20 I0, pp I-6.
[5] Chawla. G.: Sachdev. M.S.: Ramakrishna. G.: "An Improved
/ Based Admittance Relav Using Pre-Processed Inputs" .
India Conference. 2008. INDICON 2008. Annual IEEE
Vol. I , pp 234 -239.
[6] Janison R. de Carvalho, Denis V. Coury, Carlos A Duque,
David C. Jorge, "6 Development of Detection and
Classifcation Stages for a New Distance Protection Approach
Based on Cumulants and Neural Networks"
Power and Energy Society General Meeting, 2011 , pp 1-7.
[7] Jain. A: Thoke. A.S.: Patel. R.N.: "Double Circuit
Transmission Line Fault Distance Location using Artifcial
Neural Network". Nature Biologicallv Inspired Computing,
2009. NaBIC 2009. World Congress on 2009 , pp13 -18.
[8] Anamika Jain, AS.Thoke, R.N. Patel ,Ebha Koley,
"Intercircuit and Cross-country Fault Detection and
Classifcation Using Artifcial Neural Network " India
Conference (INDICON), 2010 Annual IEEE , 2010 , pp I-4.
[9] Ahad Esmaeilian, Sajjad Astinfeshan, "A Novel Power Swing
Detection Algorithm Using Adaptive Neuro Fuzzy
Technique", 2011 International Conference on Electrical
Engineering and Informatics, 17-19 July 2011, Bandung,
Indonesia, pp 1-6.
[10] He Jiali 1um Yuqian Li Yongli Wang Oun_ Luo Shanshan,
Distance Relay Protection Based on Artifcial Neural
Network Proceedings of the 4th Interational Conference on
Advances in Power System Control, Operation and
Management, APSCOM-97, Hong Kong, November 1997, pp
515-520.
[11] Fitiwi, D.Z.; Rao, K.S.R.,"Assessment of / based Auto
Reclosing scheme Developed on Single Machine-Infnite Bus
Model with IEEE 14-Bus System Model Data", TENCON
Publication Year: 2009 , pp 1 - 6
[12] Abdlmnam A Abdlrahem Dr. Hamid H Sherwali "Modelling
of Numerical Distance Relays Using MATLAB" 2009 IEEE
Symposium on Industrial Electronics and Applications
(ISIEA 2009), October 4-6, 2009, Kuala Lumpur, Malaysia,
pp 389-393.
[13] D. V. Coury, D. C. Jorge, "Artifcial Neural Network
Approach to Distance Protection of Transmission Lines"
IEEE Transactions on Power Delivery, Vol. 13, No. 1,
January 1998, pp 102-108.
[14] Neural Network Toolbox 6, Howard Demuth, Mark Beale,
Martin Hagan
ISBN: 978-81-909042-2-3 20l2 IEEE