Vous êtes sur la page 1sur 7

102

IEEE Transactions on Power Delivery, Vol. 13, No. 1, January 1998

Artificial Neural Network Approach to Distance Protection


D. V. Coury, Member IEEE
Electrical Engineering Department University of S%oPaulo, C.P. 359 13560-250 S%oCarlos - SP - Brazil
Abstract: A distance relay for the protection of transmission lines is usually designed on the basis of fixed settings The reach of such relays is therefore affected by the changmg network conditions The implementation of a pattem recognizer or power system diagnosis can provide great advances m the protection field Thls paper demonstrates the use of an Artificial Neural Network as a pattem classifier for a distance relay operation. The scheme utilizes the magnitudes of three phase voltage and current phasors as inputs An improved performance with the use o f an Artificial Neural Network approach is experienced once the relay can operate correctly, keeping the reach when faced wth different fault conditions as well as network configuration changes

D. C. Jorge, Member BEE

Keywords: Distance Protection, Relaymg, Power Systems, Artificial Neural Networks

I. INTRODUCTION Distance relaying techtuques have attracted considerable attention for the protection of transmission lines. The principle of these techniques measures the impedance at a fundamental frequency between the relay location and the fault point, thus determining if a fault is internal or external to a protection zone. Voltage and current data are used for this purpose and they generally contain the fundamental frequency signal added with harmonics and the DC offset. With digital technology being ever increasingly adopted in power substations, more particularly in the protection field, distance relays have experienced some improvements, mainly related to efficient filtering methods (such as Fourier, Kalman, etc.). As a consequence, shorter decision time has been achieved [1],[2],[3]. The trip/no trip decision has been improved, compared to electromechanical/solid state relays.
PE-772-PWRD-0-04-1997 A paper recommended and approved by the I Power System Relaying Committee of the IEEE Power
Engineering Society for publication in the IEEE Transactions on Power Delivery. Manuscript submitted December 16, 1996; made available for printing March 26, 1997.

However, the digital distance protection is usually designed on the basis of fixed relay settings. The reach accuracy of a distance relay can therefore be affected by the different fault conditions @articularly in the presence of the DC offset in the current waveforms) as well as network configuration changes. In order to face such a problem in conventional settings, a safety margin is necessary so as to avoid overreaching. This paper demonstrates that the theory of Artificial Neural Networks (ANNs) can be used as an alternative computational concept to the conventional approach, based on a programmed instruction sequence. A N N s attempt to reproduce the learning and generalization processes of the human brain. The potential of ANNs has attracted power system researchers to look at it recently as a way to solve problems related to different fields such as load forecasting, security assessment, economic dispatch, etc. [4], [51. Concerning the application in the protection field, the pattern recognition approach to distance relaying [6],[7] as well as a fault direction discriminator for transmission lines [SI should be mentioned. Indeed, very encouraging results are presented in such schemes. This work describes the application o f ANNs for the distance protection of transmission lines. The Artificial Neural Network (ANN) approach works as a pattern classifier, being able to recognize the changing power system conditions and consequently improve the performance of ordinary relays using the digital principle. The capability of the relay based on ANN theory to keep the reach accuracy when subjected to dfferent fault conditions as well as network confguration changes is shown.
I1 . ARTIFICIAL NEURAL NETWORKS
Artificial Neural Networks (ANNs) are inspired by biologcal nervous systems and they were first introduced as early as 1960, Nowadays, studies of ANNs are growing rapidly for many reasons:
e e

ANNs work with pattern recognition at large. ANNs have a high degree of robustness and ability to learn. ANNs are prepared to work with incomplete and unforeseen input data.

0885-8977/98/$10.00 0 1997 IEEE

103

"+T

output u t s

Fig. 1- Perceptron representation 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 pattern that lies within is vital to pattern recognition in a real world environment [9]. The neuron is the nervous cell and is represented in the ANN universe as a perceptron. Fig. 1 shows a simple model of a neuron characterized by a number of inputs P,, P2,..., PN, the weights WI, W2,...., W,, the bias adjust b_ and an output 3. The neuron uses the input, as well as the information on its current activation state to determine the output _a, given as in (1).

Fig. 2 - Three Layer Artificial Neural Network in any feed-fonvard network to learn a training set of inputoutput pairs. The use of the bias adjust in the A N N s is optional, but the results may be enhanced by it. A multilayer network with one hidden layer is shown in Fig. 2. Ths network consists of a set of N input units (XI, i = 1,...N), a set of n output units (Y, i = 1,... n) and a set of J hidden units (V, , j = 1,... J Thus, the hidden unit Vj ) . receives a net input and produces the output:

The final output is then produced:

Y, = F xWlmVm where i = l,..n


The neurons are normally connected to each other in a speclfied fashion to form the ANNs. These arrangements of interconnections could form a network which is composed of a single layer or several layers. As mentioned before, the ANN models must be trained to work properly. The desired response is a special input signal used to train the neuron. A special algorithm adjusts weights so that the output response to the input patterns will be as close as possible to the respective desired response. In other words, the ANNs must have a mechanism for learning. Learning alters the weights associated with the various interconnections and thus leads to a modification in their strength. 111. THE BACKPROPAGATIONMETHOD is: The backpropagationalgorithm is central to much current work on learning in neural networks. It was invented independently several times, by Bryson and Ho (1969), Werbos (1974), Parker (1985) and Rumelhart, Hinton, and Williams (1986). A closely related approach was proposed by Le Chun (1985). The backpropagationmethod works very well by adjusting the weights which are connected in successive layers of multi-layer perceptrons [ 101. The algorithm gives a prescription for changing the weights 1 "

L, J

(3)

F[.] is a non-linear transfer function which can be of various forms. Backpropagation networks often use the logstic sigmoid as the activation transfer function. The logistic sigmoid transfer function maps the neuron input from the interval (-oo,+oo) into the interval (0,+1). The logistic sigmoid, shown in (4), is applied to each element o f the proposed ANN.
(4)

Where: n - summation output& -bias adjust. The usual error measure or cost function for the process

(5)

and now becomes,

104

This is clearly a continuous differentiable function of every weight, so we can use a gradient descent algorithm to learn appropriate weights. Since the weight errors are successfully back-propagated from the output layer, this specific training algorithm is known as error backpropagation[lOJ. IV. APPLICATION OF THE BACKPROPAGATION METHOD TO DISTANCE PROTECTION , This paper proposes an ANN approach in order to simulate a distance relay by implementing the backpropagation method. The ANN relay is supposed to i d e n m faults located in the first protection zone of the line (covering 80% of the line length), having the magnitudes of the voltage and current phasors corresponding to the postfault fundamental frequency as inputs. In order to obtain the magnitudes and phase angles of such waves, the half cycle Discrete Fourier Transform @FT) filter was utilized. For this approach, only the fundamental frequency magnitudes of voltage and current (not the phase angles) were utilized as the inputs of the network. In t h s way, through three phase voltage ( Iv, , v b , V, ) and current ( Ia I b I , I, I magnitudes seen at the busbar, the scheme should discriminate between faults lying within 80% of the line length and faults outside that zone, giving 0 and 1 answers respectively for the situations described. For faults located within 80% of the line length, the relay should send a trip signal to the equivalent circuit breaker. It should be mentioned that the input variables have to be normalized in order to reach the ANN input level (kl). Due to the necessity o different scaling for voltages and currents, the normalized current must be- divided by an additional factor. Concerning the ANN architecture, parameters such as the number of inguts to the network. as well as the number of

60 Hz Mag.

I Val
lvbl

VCI

+ trip/

no trip
Ira1

lIbl
I I C I

Fig. 3 - ANN architecture used by Johns and Aggarwal [ll]. It should be mentioned that although the techmque described is based on Computer Aided Design (CAD) studies, practical considerations such as the effect of transducers, analogue filters and quantisation on primary system fault data are also included in the simulation so that the data obtained is very close to that found in practice. The technique also considers the physical arrangement of the conductors (Fig. 4.b), the characteristics of the conductors, mutual coupling. the effect of earth return path and the effects of frequency dependent parameters. Perfect line transposition has been assumed. The 100 km, 400 kV transmission line used to train and test the proposed ANN is shown in Fig. 4.a More details about the transmission line parameters are presented in Appendix I.

I I I I I

I I, I

the logistic sigmoid described in the earlier section. The use of ANNs in distance relays results in a considerable advance in the correct diagnosis of operation. ANNs can solve the overreach and the underreach problems which are very common m the conventional distance relay design. V . THE POWER SYSTEM DIAGRAM USED In order to test the applicability of the scheme proposed above, a simulation of the transmission line in a faulted condition was utilized. This paper makes use of a digital simulation of faulted EHV transmission lines developed

................... ?

Vertical single-circuit line:

..:9% . 6 8 2.
............... ;b

: 10lm

1
I

10.2m 8.3m

x 8.8m .:,
M

i 104m

..........

'i

12.2m

//////////

(b)

Fig. 4 Transmission line used for the ANN studies

105

Fig. 5 typifies the data of a faulted phase as attained from an accurate model of the system, for an a to earth fault at 50 km from A. Information from only one side of the line was used in the referred method (from busbar A).

1
400000

"

"

"

"

10000

5000

200000
h

5
0

0
0

>
-200000

r,
h

-5000

-400000
O

_ _ _c u r r e n t
I

( /
0.03
0,04

J
-10000 0.05

ltag e
0.01 0.02

0.00

tim e ( s )

Fig. 5 Voltage and current of the faulted phase

3
current signal Analog V,I input signals voltage signal Antialias. Multiplexer (clock) Converter
~~

........

Fig. 6 shows the schematic diagram for the hardware needed in an ANN implementation, including the microprocessor based neural relay. The voltage and current signals are taken from the transmission line, passed through low pass filters and then conditioned to make measurements. The waveforms are sampled and digitized using analogue to Qgital converters. The digital values are then fed to a microprocessor. The first step of the software is to detect the fault. In case a fault exists, the voltage and current signals are fed to half cycle DFT filters for extraction of the fundamental phasors. As mentioned before, the magnitudes of the phasors (voltage and current) are the input quantities to the proposed ANN. The converged set of weights, which are worked in an off-line mode, are stored in the microprocessor for on-line application. The scheme works in a sample frequency of 4kHz. As half cycle DFT filters are used, approximately 32 samples of each quantity are utilized to feed the filters and their outputs (magnitude phasors of the signals) are the inputs for the ANN process. It should be mentioned that, after the fault occurrence only these 32 data samples are utilized in the decision making process. Concerning the time response for the algorithm, it is estimated that the ANN alone would introduce a serial processing time in the range of 11 ps for the archltecture shown in Fig.3, considering a Pentium 90 MHz. Due to the use of the magnitude phasors, a simple neural network was created. As a consequence, the processing time to be added to the time spent by the half cycle DFT routines is very short. In fact, for practical applications, the FFT (Fast Fourier Transform) should be employed. The FFT is a fast algorithm used for efficient implementation of the DFT. It is estimated that for all studied cases, considering the use of FFTs and the ANN, the relay operates in less than 11 ms &er the fault occurrence.

V . THE TRAINING PROCEDURE AND TEST RESULTS I


FOR THE PROPOSED ANN The Neural Network Toolbox from the software MatlabTM
[12] was used to create the ANN diagram, train it and obtain

[ H H F A' eTT H [
Fig. 6 - Block diagram of the ANN distance relay.

the weights as output. The initial weights as well as the initial bias employed random values between 0- 1. On-line process ,...... A total number of 2,144 different faulted cases in various locations of the transmission line were used in order to train and test the proposed ANN. In the aforementioned cases several fault resistance values (between 0 100 0)and distinct trained off-line network conditions were included. Concerning the fault inception, two angles related to the voltage wave (90" and 0") were chosen for the training process. The main objective was to characterize two well-defined conditions: the worst case in output Digital i Fourier transfer Transf. function 0/1 i the voltage high frequency components (90") and a very ..................................................... pronounced exponentially decaying DC offset in current

signals (0').

In the referred scheme, a protection of 80% of the line length was chosen as the reach of the first zone of the relay, as

106

mentioned before. Points next to the region where triplno trip condition exchanges (80 km for the line used) had special treatment. In those cases, a lower degree of sparsity was chosen between locations used for training. As mentioned earlier, the backpropagation algorithm was used for training. The learning process converged in about 80,000 cycles and 2 hours of computing time on a SUNSPARCstation 20. This training 6procedureis only executed once and a generalization capability of the ANN is expected. The momentum factor was kept at 0.95 and the final error was less than 0.001. Tables I to V show the results of the ANN model used as a &stance relay. The ANN answer is shown, compared to the expected ones, for faults along the transmission line. The correct answer column illustrates the expected trip (0) and no trip (1) decisions. The tables indicate the exact values as attained from the ANN output. For practical applications, the interval 0-0.499 can be considered as a trip decision and 05-1 as a no trip decision. Considering that, the (*) symbol denotes the incorrect answers. It should be noted that the cases utilized for the tests are different from the ones used for the training. Table I shows the estimated output of ANN distance relay subjected to hfferent fault resistance$ -Rf- 17 and 70 ! ) (3, 2 for the operation condition shown in Fig. 4 (power transfer angle 6 of IS"). Table I1 characterizes the relay output subjected to different fault inception angles -FIA- (30", 60" and 90"). The operation condition of the line is the same as for the preceding Table.

Table 111 also illustrates the estimation output of the ANN relay, but for a different network condition. It is equivalent to Table I, but with the power transfer angle F of 20" rather than is0. Table IV shows the ANN performance for the operation condition illustrated in the earlier case, but with the source capacities at end A of 14, 8 and 6 GVA respectively.
TABLE II Estimated Output of the ANN Distance Relay with Different Fault InceptionAngles (6=15" and Rf=17R)

TABLE I Estimated Output of the A J Distance Relay wlth & ' Different Fault Resistances (6=15" and FIA= 0 ) "

TABLE III Estimated Output of the ANN Distance Relay wth a Different Network Condition (6=20 and FIA= 0')
Fault
Distance Correct

Rf=3n

Rf=17Q

Rf=7052

Answer

107 Finally, Table V illustrates the ANN estimation with varying source impedance. In practice, the source impedance can change significantly and the relay should not be affected by such a variation. The aforementioned table indicates the results for faults when the impedance of source A is 5 , 10 and 90 times the impedance of source B. The results presented in the Tables tp the efficiency of ye the proposed scheme. It should be pointed out that about 95% of almost all of the 1,000 cases tested estimated the expected answer for the ANN distance relay. Moreover, the inaccurate answers appear in 98% of the cases in the regon tripho trip between 78 and 82 km. It is well known that the half cycle DFT technique is effective in rejecting the high frequency components but only provides poor filtering of the low frequency components, particularly the DC offset [13]. As a consequence, overreaching can occur when using such a filter, considering the conventional distance relays. It is shown however, that the proposed scheme canceled out the overreach trend of the half cycle DFT, keeping the ANN relay reach accuracy, as can be observed in Table I and I11 (FIA=O"). Moreover, the method retains the reach accuracy when considering different operational conditions (Table I11 to V). It is clear from the results that the relay practically did not suffer from underreaching or overreaching in all the situations tested. This is the main advantage of an ANN relay in comparison to a conventional one. The ANN relay has the ability to learn aspects related to the fault condition and network configuration in order to keep the reach accuracy. Thus, the use of ANNs can make it possible to extend the first zone reach of the relays. It should also be noted that this study takes into account phase a to earth faults only. The same set of converged weights can be used for correct diagnosis of b to earth and c to earth faults. In order to extend the proposed scheme to real situations, a fault type classifier similar to the one presented in [14] should be used and a similar training process made for the other types of faults. At present, t h s extension is under development. VI1 . CONCLUSION The use of an ANN as a pattern classifier to simulate a distance relay was investigated in this paper. The results obtained in t h s scheme are very encouraging. The ANN relay can provide a fast and precise operation, keeping its reach accuracy when faced with different fault conditions (even in the presence of the DC offset in the current waveforms) as well as network changes. This is an improvement in performance if compared to conventional distance relays. Thus, the use of ANNs can make it possible to extend the first zone reach of the relays, enhancing system security. It should also be remembered that this process involved training and testing different network confgurations until satisfactory performance was achieved. It must however be pointed out that this tool opens a new dimension in relay philosophy which should be widely investigated, allowing one to solve some of the various problems related to the distance protection of transmission lines.
WIT. ACKNOWLEDGMENT

Fault Distance

Correct

oun)
25 35
45

14GV

8GVA

6GVA

Answer (0 - 1)

86 88 91 93 94
95

1
1

1 1

1 1 1
1

1 1 1
1

1 1 1 1 1
1

1 1

1 1
1 1

TABLE V Estimated Output of the ANN Distance Relay wt ih Varying Source Impedance.(S=2O0 Rf=70R) and

The authors acknowledge the financial support given by CAPES (Brazilian Government Institution).

108

IX. REFERENCES
[l] AG. Phadke and J.S. Thorp, Computer Relaying for Power Systems John Wiley and Sons, 1988.

[13] H J Altuve, I D i u , E Vasquez Fourier and Walsh Digital Filtenng Algonthms for Distance Protection, IEEE Trans on Power Systems, Vol 11, no 1, February 1996 pp 457-462 [14] T Dalstein, B Kuhcke Neural Network Approach to Fault Classlfication for High Speed Protective Relaymg, IEEE Trans on Power Delivery, Vol 10,no2.April 1995,pp 1002-1011

[2] M Ramamoorty, Application of Digital Computers to Power System Protection, Journal o f h s t . Eng., India,Vol 52, n10, June 1972, pp 235238. [3] A A Girgis and R G Brown, Application of Kalman Filtering in Computer Relaymg, IEEE Transactions on Power Apparatus and Systems, PAS-100(7), July 1981, pp.3387-95 [4] C N Lu, H T Wu and S Vermuri Neural Network Based Short Term Load Forecasting, IEEE Transactions on Power Systems, Vol8, no 1, February 1993, pp 336- 342 [5] M A El-Sharkiiwi, R J Marks, and S Weerasooriya, Neural Networks and Their Application to Power Engmeering, Control and Dynamics Systems ,Vol 41, 1991, pp 359-401 [6] S K Chakravarthy, C V Nayar and N R Achuthq Applying Paitem Recogmtion i Distance Relaying - Part 1 Concept and Part 2 n Feasibility, Proceedings Generahon, Transmission and Disbibuhon, Vol 139, no 4, July 1992, pp 301-314 [7] S A Khaparde,P B Kale, and S H Aganval, Application of Artificial Neural Network in Protective Relaymg of Transmission Lines, First Internatzonal Forum on Applications of Neural Networks to Power Systems, Seattle,WA July 1991, pp 139-143 [8] T S Sidhu, H Smgh, and M S Sachdev, Design, Implementation and Testing of An Artificial Neural Network Based Fault Direction D i s c m a t o r for Protecting Transmission Lines, IEEE Transactions on PowerDelivety, Vol 10, no 2, April 1995, pp 647-706 [9] P D Wassennan, Neural Computing Theory and Practice, Van Nostrand Remhold, New York,1989

APPENDIX I - TRANSMISSION LINE? PARAMETERS


The line simulation used in this work was a typical 400 kV vertical construction line. The relevant parameters used are: (i) Earth resistivity (assumed homogeneous)= 100 Om. (ii) Source XIR ratio = 10 Source sequence impedance ratio Z,JZizsI 1.5 = ($Phase conductor =4x54/7/0.33s.c.a. with 0.305 bundle spacing earth wire = 54/7/0.33 s.c.a. (s.c.a.=steel core aluminium) (iv)Nominal CVT and CT ratios = 400/0.11 and 2000/1, respectively. Various other factors such as effect of fault type, differing source capacities at the terminals, different fault resistances and different fault inception angles were also included.

BIOGRAPHIES
Denis V. Coury was bom in Brazil, in 1960. He received a B.Sc. degree in Electrical Engineering from the Federal University of Uberlandia, Brazil in 1983, a MSc degree from the University of SLo Paulo, Brazil in 1986 and a Ph.D. degree from Bath University, England in 1992. He worked for the Technological Research Institute (IPT), SLo Paulo, Brazil from 1985 to 1986. He joined the Department of Electrical Engineering, University of SHo Paulo, SLo Carlos, Brazil in 1986, where he is presently an Assistant Professor in the Power Systems Group. His areas of research interest are power system protection as well as new techniques for power system control and protection including the use of Expert Systems and Artificial Neural Networks. Email coury@sel.eesc.sc.usp.br

[lo] J Hertz, A Krogh, and R G Palmer, Introduchon to the Theory of Neural Computation, Adison-Wesley Pub Co,1991
Ill] A T Johns, and R.K Aggarwal, Simulation of Faulted EHV Transmssion Lines with Particular Reference to Very High Speed Protection, IEE Proceeding, Voi 123, April 1976, pp 353-359 [12]H. Demuth and M. Beale, Neural Network - Toolbox - For Use with Matlab y 1992.

David C. Jorge was bom in Brazil, in 1971. He received a B.Sc. degree in Electrical Engineering at the University of SLo Paulo, SLo Carlos, Brazil in 1993. He is at present in the final stage of his MSc degree in Electrical Engineering at the same university. His research area is power system protection.

Vous aimerez peut-être aussi