7/27/2019 06497761

1/5

Transmission Line Faults Detection,

Classifcation and Location using Artifcial

Neural Network

is shier M Tyeb Oer AIAziz Alhi

Transm ission lines, among the other electricalpower system components, suffer from unexpected failures dueto various random causes. These failures interrupt the reliability

of the operation of the power system. When unpredicted faults

occur protective systems are required to prevent the propagation

of these faults and safeguard the system against the abnormal

operation resulting from them. The functions of these protective

systems are to detect and classify faults as well as to determine

the location of the faulty line as in the voltage and/or current linemagnitudes. Then aer the protective relay sends a trip signal to

a circuit breaker(s) in order to disconnect (isolate) the faulty line.

The features of neural networks, such as their ability to learn,

generalize and parallel processing, among others, have made

their applications for many systems ideal. The use of neural

networks as pattern classiers is among their most common and

powerful applications.

This paper presents the use of back-propagation (BP neuralnetwork architecture as an alternative method for fault

detection, classication and isolation in a transmission line

system. The main goal is the implementation of complete scheme

for distance protection of a transmission line system. In order to

perform this, the distance protection task is subdivided into

different neural networks for fault detection, fault identication(classication) as well as fault location in different zones.

Three common faults were discussed; single phase to ground

faults, double phase faults and double phase to ground faults.

The result provides a reliable and an attractive alternative

approach for the development of a protection relaying system for

the power transmission systems.

ndx m Transmission lines; fault detection; articialneural network.

I. NTRODUCTION

T

HE greatest threat to the continuit of electricit supply issystem faults. Faults on electric power systems are an

unavoidable problem. Hence, a wellcoordinated protectionsystem must be provided to detect and isolate faults rapidly sothat the damage and disruption caused to the power system isminimized. The clearing of faults is usually accomplished bydevices that can sense the fault and quickly react to disconnectthe fault section. It is therefore an everday fact of life thatdifferent tpes of faults occur on electrical systems, however

Eisa Bashier M Tayeb, Member, IAENG, and Oer ziz Rhim wt t Electrical Engineering Department, College of Engineering, SudanUniversit of Science and Technology.

inequently, and at random locations. Faults can be broadlyclassied into two main areas which have been designated as"Active and "Passive [1].

In the control centers of the electrical power systems alarge number of alarms are received as a result of differenttpes of faults. To protect these systems, the faults must bedetected and isolated accurately. n mainly overhead line

systems, the majorit of shortcircuit faults, tpically 80-90%,tend to occur on overhead lines and the rest on substationequipment and busbars combined [2]. The operators in thecontrol centers have to deal with a large amount of data to getthe required information about the faults.

The information processing via biological neural networksis done by the huge amount of complex interconnections ofthe neuronal cells (neurons) which interact among each otherby exchanging brief electrical pulses or action potential.

Inspired by the biological nervous system, AN operateson the principle of largely interconnected simple elementsoperating as a network nction. In doing so, no previous

knowledge is assumed, but data, records, measurements,observations are considered. AN research stands on the factof leaing from data to mimic the biological capabilit oflinear and nonlinear problem solving [3].

Basically, we can design and train the neural networks forsolving particular problems which are dicult to solve by thehuman beings or the conventional computational algorithms.The computational meaning of the training comes down to theadjustments of certain weights which are the key elements ofthe AN. This is one of the key differences of the neuralnetwork approach to problem solving than conventionalcomputational algorithms which work stepbystep. Thisadjustment of the weights takes place when the neuralnetwork is presented with the input data records and thecorresponding target values.

Due to the possibility of training neural networks with offline data, they are found useful for power system applications.The neural network applications in transmission lineprotection are mainly conceed with improvements inachieving more effective and efcient fault diagnosis anddistance relaying [4].

The goal of this paper is to detect and identi the tpe offault in the line and to determine which zone (segment) of the

Copyright Notice: 978146736008111$31.00 2012 IEEE

7/27/2019 06497761

2/5

line has become fault. Backpropagation neural networkapproach is studied, implemented and modied to performthese three tasks. To identi the existence of faults in thesystem voltage and current signals of a line are observed.These signals are also used to speci the fault tpe andlocation.

The simulation models of the transmission line system areconstructed and the generated information is then channeled

using the soware MATLAB (Version 7) and accompanyingPower System Block Set (Version 2.1). Besides Neuroshell2soware used to provides backpropagation neural networks.

II. ANSMISSION INES ODEL

The commonly model used for AC overhead transmissionlines is called pi model network and is shown in Fig. 1. Whereshunt admittance has been even divided into two shuntelements connecting to both ends of a pi equivalent network.[5], [6].

Z -0 , j. 1' 1 + v, v T

Fig. I -Netork for a Transmission Line Model

A 110 kV transmission line system conects EL FAU withGEDAEF (145 Km) is used to develop and implement theproposed architectures and algorithms for this problem. Fig. 2shows a singleline diagram of the system used to train andtest the neural networks. The system consists of two

Substations and (145 Km) transmission line.

E FAU EARF 0- (5 K):

K 5 K

(]45 K)Fig. 2. Single-line diagram of the system studied.

The threephase voltages and currents, V [Va Vb Vc] Tand I [Ia Ib I] T are measured at substation A in Fig. 2 (4)The simulations presents three categories namely (i) phase toground faults (ii) phase to phase faults and (iii) doublephaseto ground faults.

III. BACK ROPAGATION LGORTHMBack Propagation was created by generalizing the Widrow

Hoff leaing rule to multiple layer neural networks andnonlinear differentiable transfer functions. Input vectors andthe corresponding target vectors are used to train a network

2

until it can approximate a function, associate input vectorswith specic output vectors. The term back propagation refersto the manner in which the gradient is computed for nonlinearmultilayer neural networks [7].

The output values J of a given input patte Xi do notalways correspond to their predetermined value Rj. The errorEj is given by the difference of Rj and J and is to beminimized by the weight changes. The error of a neuron j in

the output layer is:

J = (Rj 02 1The total error E of an output layer is

So, we need to minimize the error E, with respect to the

fweight changes ( J We follow the delta rule toincorporate the leaing rate , along with the gradientdescent algorithm techniques to dee the weight change,

a

aE!wf = awfO

7/27/2019 06497761

3/5

From (6) we get,

Combining (7) to (10), we nally get,

LWkj

=a.(Rj

-0j

).o/I-OJ)Hk 1 1

So, new weights are,

1 2

The eor of the output layer is back propagated to the

weights of the hidden and the input layer. is the change

in weights om the output layer to the hidden layer.

The back propagated eorof the hidden layer is given by:

Ek=E=(R-OY 1 3 J J

Coesponding to (2) is the weight change. Weintroduce a new leaing rate . We nally get the weightchanges in the hidden layer as:

W ='RjOjOHHX 14)

It has been proven that back propagation leaing withsufcient hidden layers can approximate any nonlinearnction to arbitrar accuracy. This makes back propagationleaing neural network a good candidate for signal predictionand system modeling [3].

IV. RTIFICIAL EUL ETWORK ESIGN

Articial neural network is an interconnected group ofarticial neurons that uses a mathematical model orcomputational model for information processing based on aconnectionist approach to the computation [1].

Transfer function in the AN is an imporant key elementto invoke the nonlinear relationships that maps the input(s) tothe output(s). Without the transfer function the wholeoperation is linear and could be solved using linear algebra orsimilar methods. We consider the transfer nction for theweighted sum S (lumped input) of the inputs for a successfulnetwork design. In the process of learning the neuralnetwork presented with pairs of input and output data

then teaches how to produce the output when thecorresponding input is presented. When learning is

complete, the trained neural network, with the updated

optimal weights, should be able to produce the output

within desired accuracy corresponding to an input

pattern. The learning situations in neural networks can

be classied into two distinct sorts. These are

supervised learning, and unsupervised learning [4]. As apreprocessing step the training and the testing data generatedom the transmission line system are collected. The rst stepis that of the detection of a fault situation in the system.

3

Following that, fault classication and fault isolation!Location (zones 1, 2 or 3) tasks are investigated.

A Design a/NeurlNeorks/or Fult DetectionAer extensive simulations of the output eor and network

output, it was decided that the desired selected network wouldhave an input layer with 6 neurons and one hidden layer withthree hidden neurons beside an output layer with one neuron.

The activation nction at input layer is linear (1, 1) functionwhile at hidden layer and output layer is logistic nction. Theselected network is shown in Figure (3) and the perfoanceand eor plots associated with this architecture are given byFigures (4) and (5). [8].

T 1V C P U V U A T S [8]

Case I V P.U Fault TypeNO. V V Vc f f

I .997 9991 9985 9978 9988 .9984 No fault22

0_331.1937 1.1722 3.3345 0.9814 0.9791 A to Ground35

3 17 0.3335 1.1937 0.9814 3.3345 0 9791 B to Ground224 19 1.1722 0.3335 0.9814 0.9791 3.3345 C to Ground375

070.6501 9856 5.3791 5.3791 09834 A to B13

6 0.98 0713 0.6501 0.9834 5.3791 5.3791 B t o C567 07 9856 0.6501 5.3791 0.9834 5.3791 A toC13

800

0045 1.1878 7.1872 7.8554 0 9851 A to B to Ground45

918

0045 0045 0.9851 7.1872 7 8554 B to C to Ground78

10 00 1.1878 0045 7.1872 0.9851 7 8554 A to C to Ground45

La

La

a

a

Fig.3. Back Propagation Neural network used for fault detection.

e

-Tiig Sm

Fig. 4. Output error for the BP neural netork 631

Copyright Notice: 978146736008111$31.00 2012 IEEE

7/27/2019 06497761

4/5

' : Nwrk Oupu

Dr Oupu

Fig. 5 . Network output versus desired output for BP neural network 6-3-1.

A test set was created to analyze the performance of theproposed network. A new set of data will apply to the networkthat's never seen before. A fault cases for each categor offault were utilized in the test set. Appendix includes thevariables considered to form the test set. Error plot of thetesting set is shown in gure (6). Performance of network isshown in gure (7).

1( J\ IT Sm

Fig. 6. Testing samples output error for the BP neural netork 6-3-1.

,

1.2.- .I+++1-

+.Il

0.4 4++ +H +HH HlI-= NetworkOutput

HHFig. 7. Testing samples Netork output versus desired output for the BP

neural network 6-3-1.

. Design a/NeurlNeor/or Fult Clssction

The design and development of the uses BP network as afault classier. The network designed here has six inputs (thethree phase voltages and currents) and four outputs associatedwith the four fault categories. The outputs contain variableswhose values are given as either 0 or I corresponding to thethree phases and the ground (that is, A, B, C and G) and canbe generalized to represent all the practical fault categoriespermutation involving combinations of phases.

The proposed neural networks here should classi thespecic phases involved in the fault scenario. It should be ableto distinguish among nine different categories of faults as

4

illustrated in Table (2).

T 2T BP C N T T

t tt t t tA B C G I I

I I

I I

I I

I I I I

I I I

I I I

A large number of networks were extensively studied.Aer an exhaustive search for the most suitable network size,the one with only one hidden layer and ve hidden neuronswas chosen to carr out the classication task. The activationnction at input layer is linear (-I, 1) function while at hiddenlayer and output layer is logistic nction. The proposednetwork as stated before has six inputs (the three phase

voltages and currents) and four outputs. This network isillustrated in Figure (8).

LaYl

V

V

Vc

Fig. 8. Back Propagation Neural network chosen for fault classication.

The BP neural network for fault classication was testedusing the same test set which was used to test the detectionnetwork. A fault cases for each categor of fault were utilizedin the test. Appendix includes the variables considered to formthe test set. Error plot of the testing set is shown in gure (9).The selected network om the previous section was able torecognize correctly the tpe of the detected fault.

0 G G --

Phs

- Phs

- Phs C

Fig. 9. Testing samples output error for the BP neural netork 6-5-4 .

C Design a/NeurlNeork/or Fult soltionLoctionThe design and development of the detection neural

network is followed in this section in order to choose the mostsuitable BP network as a fault isolator. The network isexpected to identi the location of the fault by classiing the

Copyright Notice: 978146736008111$31.00 2012 IEEE

7/27/2019 06497761

5/5

identied fault into one of the three fault zones, namely ZI,Z2 and Z3.

The proposed neural networks here should isolate thespecic zone involved in the fault. The desired truth table forthe network training is shown in Table (3).

T 3T I N D R.

F N O

L Z Z2 Z3o 1 1 o 2 1 o 3 1

A large number of BP networks with different structureswere studied and analyzed in order to obtain the simpleststrcture. The training includes some of the selected networks,namely structures, 6553, 6663, 6763 and 6543. It isfound experimentally through trial and error that a BPnetwork with two hidden layers provides the best trainingperformance. The rst hidden layer has 5 neurons and thesecond hidden layer has 4 neurons. The activation function at

input layer is linear (1, 1) nction while at hidden layer andoutput layer is logistic nction. This network is illustrated inFigure (10).

n e

Hddn

y,

b

Zne (Zl)

y

Zne (Z)

Zne(Z)

b

Fig. 10. Back Propagation Neural network chosen for fault isolation.

V. COO

This thesis has investigated the use of backpropagation(BP) neural network architecture as an alteative method forfault detection, classication and isolation in a transmissionline system. It uses MS values of phase voltage and phasecurrent as inputs. Simulation models of the transmission linesystem are constructed and the generated information is thenchanneled using MATLAB soware (Version 7) and

accompanying Power System Block Set (Version 2.).Neuroshell 2 soware is also used to provide backpropagation neural networks. Three common faults werediscussed; single phase to ground faults, double phase faultsand double phase to ground faults. Due to the exibility of theneural networks which accept any real values (highlycorrelated or independent) as an input, resistant to errors in thetraining data and fast evaluation. The results obtaineddemonstrate that the performance of the backpropagation(BP) neural network architecture was highly satisfactor.

Neural networks, in general, provide a reliable and an

5

attractive alteative approach for the development of aprotection relaying system for the power transmissionsystems.

VI. EFEENCE

[] Laurene V. Fausett, Fundamentals of Neural Netorks: Architectures,Algorithms, and Applications, Prentice Hall, 1993.

[2] Nasser D. Tleis, Power Systems Modeling and Fault Analysis Theoryand Practice, Elsevier Ltd, 2008.

[3] Hagan MT, Demuth HB, Beale MH, 'eural network desi, PWSPublishing, 1996.

[4] Kevin Gurney, an Introduction to Neural Networks, UCL Press, 1997.[5] Arthur R. Bergen and Vijay Vial, Power Systems Analysis - 2nd

Edition, Prentice Hall, 2000.[6] C. R. Bayliss, B. J. Hardy, Transmission and Distribution Electrical

Engineering- Third Edition, Newnes, 2007.[7] R. Kamyab Moghadas and S. Gholizadeh "A New Wavelet Back

Propagation Neural Networks for Structural Dynamic Analysis AEGEngineering Letters, February 2008.

[8] 0 A A Elmubark, Fault Detection, Classication and Location in PowerTransmission Line System Using Articial Neural Network, MSc.dissertation, Dept. Elec. Eng., Sudan Univ. of Sci. & Tech. Khartoum,March 2011

VII.IOGAPHY

Eisa Bashier M Tayeb; was born in Wadbunda in thenorthern Kurdofan State, Sudan; on Jan 1964. Hegraduated om the Khartoum Polytechnic, and studiedMSc in Instrumentation & Control at Aligarh MuslimUniversity; India. His PhD in ElectronicsInstrumentation was also om AMU-India. His totalemployment experience is 20 years at SUST.

Tayeb received honors, awd and grants some of themare ICCR Scholship (Cultural Exchange Progr) India-1996, UGC ofndiaas junior Research fellowship 1999 and Senior Research fellowship in 2002.Works as a review specialist and team member in the peer review 2008;sponsored by UDP. Currently is the depu De for Academic Affairs,College of Engineering. His special elds of interest included ControlEngineering and Instrumentations.

Copyright Notice: 97814673600811$3.00 2012 IEEE