Electric Power Systems Research 76 (2005) 115120

ART artificial neural networks based a.Y.

ngchua

3 May 2

Abstract

This pape ry (ARdynamically ractersthe disadvan twork b(ARTNN) h ce andselector has ase selevarying pow n be trand experim s and e 2005 Else

Keywords: A

1. Introduction

Character of conventional phase selectors varies underdifferent padapt dyna[1]. Pattern(ANN) isto the varyrelay reseastrated AN[36] had e(MFNN) bfication relof these fative relaysprotective ragation alghas some dlocal extremfrom all the

CorresponE-mail ad

weights of the network will be retrained with all the trainingpatterns, even if only a new pattern needs to be stored inthe network. In other words, BPNN cannot get a good trade-

0378-7796/$doi:10.1016/jower system conditions, because they cannotmically to the power system operating conditions

recognizer based on artificial neural networksa good classifier and adapts itself dynamicallying environment, which has aroused protective

rchers interests [1,2]. Many papers have demon-Ns applications in protective relays. Referencesxplored Multilayer Feed forward Neural Networkased fault direction discriminator and fault classi-ay and distance protection relay, respectively. Allcts have shown the good performance of protec-based on ANN. Presently, most of ANN basedelays are designed by utilizing Error Back Prop-orithm based MFNN (BPNN). However, BPNNisadvantages [7] such as: low learning efficiency,um and unstable weights of the network derivedtraining patterns. Unstable weights mean that the

ding author.dress: jdyy@sjtu.edu.cn (Y. Yang).

off between weights stability and plasticity. These questionshave hindered the further applications of ANN in protectiverelays. In order to solve these questions, based on analyz-ing and comparing several neural network architectures, anadaptive phase selection relay is proposed and designed byutilizing the neural network architecture based on adaptiveresonance theory (ART). Simulation results and experimen-tal field data tests have validated the way correctness andeffectiveness.

2. Basic theory

2.1. The principle of adaptive resonance theory basedneural network (ARTNN)

The basic principle of ARTNN is an adaptive resonancetheory introduced by Carpenter and Grossberg [8]. Comparedwith other neural network architectures, ARTNN could bet-ter solve the above-mentioned questions, particularly about

see front matter 2005 Elsevier B.V. All rights reserved..epsr.2005.05.006Y. Yang , N.L. Tai, WDepartment of Electric Power System, Shanghai Jiao Tong University, Do

Received 5 November 2004; received in revised form

r introduces a new phase selector based on adaptive resonance theoto the power system operating conditions, it presents different chatage, an adaptive phase selector, which utilizes artificial neural neas some advantages such as no local extremum, quickly convergenbetter performances compared with other neural networks based pher system operation conditions. Furthermore, the phase selector caental field data tests have illustrated the phase selectors correctnesvier B.V. All rights reserved.

daptive resonance theory; Neural networks; Adaptive phase selectiondaptive phase selectorYun Road, No. 800, Shanghai 200240, PR China

005; accepted 3 May 2005

T). Because conventional phase selector cannot adaptunder different power system conditions. To overcomeased on ART, is designed. ART based neural networkso on. Therefore, the proposed ARTNN based phase

ctor, and the new selector can adapt dynamically to theained and learned on-line. A lot of EMTP simulationsffectiveness.

116 Y. Yang et al. / Electric Power Systems Research 76 (2005) 115120

the question of unstable weights that cannot be handled withother neural network architectures. A better tradeoff betweenstability and plasticity of weight can be attained by the coop-eration of the two subsystemsattentional subsystem andorienting subsystem within ARTNN. That is to say, depend-ing on the calculated vigilance, which denotes the extend ofsimilarity between two patterns, the pattern, whose vigilanceis above the threshold value set in advance, will be thought tobe similar with a pattern stored in the network database, andbe dealt with the attentional subsystem. At the same time, thepattern, whose vigilance is lower than the threshold value, isencoded and stored into the network by the orienting subsys-tem. It can be seen that ARTNN could learn new knowledgeand avoid modifying the patterns stored in the network, whichmakes ARTNN obtain the ability of learning on-line. Conse-quently, the special character could make the phase selectionrelay based on ARTNN obtain better performance underdifferent power system conditions. Moreover, learning algo-rithm of ARTNN is a linear iterative process, which ensureshigh learning efficiency and no local extremum. Architectureof ART2-A [9] shown in Fig. 1 is responsible for arbitrarysequences of analog input patterns, which has advantages ofrapid category and recognition learning. It consists of atten-tional subssubsystemgory repres

In termsis supervisattain exceOppositelyso it can atraining paworks. Accterns compembody es

2.2. Simple description of ART2-A algorithm

The structure and algorithm of ART2-A are complicatedand have be[9], so therFig. 1 as fo

i. Importwork.

ii. Calculafield F

If I0I is the

I = Iwhere

P = f

where

iii. Inputteegory r

Thes follo

j ={

here2 and

he fielode.ompeatchiThe

s the wis the

et threevel. Differea) If

thanet

b) If Ywo

new

c) If Ywo

eightEqua

j,i =

herej,i =ystem and orienting subsystem. The attentionalconsists of input representation field F1 and cate-entation field F2.of training patterns, learning strategy of BPNN

ed, so it needs many training patterns in order tollent fault-tolerance and generalization capability., learning strategy of ARTNN is non-supervised,utomatically abstract inherent characteristics oftterns by adaptively modifying parameters of net-ording to that, ARTNN needs fewer training pat-ared with BPNN, but the training patterns shouldsential of the analyzed question as possible.

Fig. 1. Schematic diagram of ART2-A.

a

Y

w

Ftn

iv. Cm

iJs

ld(

(

(

v. W

Z

w

Zen described and illuminated in detail in referencee only a reduced algorithm is described based onllows:

ing a vector (M dimension) as input vector of net-

ting the vectors (I, P) of the input representation1.is an M-dimensional input vector of network, thennormalized vector of I0 as follows:0

() = ()/() and the vector P is:I

f ()i ={

()i, if ()i > , 0 < 1/

M

0, otherwised value of the output nodes (i.e. nodes of the cat-epresentation field F2).inputted value of the jth output node of F2 is givenws:

0, if the jth node is inactiveP Zj,i, if the jth node is active

Zj,i stands for the weight vector of the jth node ofP (p1, . . ., pi, . . ., pM) denotes the input vector ofd F2. Yj is called matching value of the jth output

ting among output nodes and verification ofng level.output node, which has the biggest sum of input,inner among all output nodes, i.e. YJ = max(Yj),index of winning node. (named vigilance) is a

shold in advance and is utilized to judge matchingepending on the relationship between YJ and ,

nt processes are dealt with as follows:YJ > , the input vector is classified to the typet the node J denotes, which has been stored in thework, then network turns into the weight learning.J < and there is an inactive node J in the net-

rk, then set ZJ,i = P, where node J denotes thissort and the algorithm back to step i.

J < and there are no inactive node, then the net-rk directly back to step i.learning.tions of modifying weights are:{Zj,i/

Zj,i , if j = JZj,i, if j = J

j stands for the index of network output nodes andZj,i + (P Zj,i), is learning step.

Y. Yang et al. / Electric Power Systems Research 76 (2005) 115120 117

The above five steps formed an integral algorithm, whichconsists of discrimination, classification and weight learning.Compared with the BPNN, its speed is quicker as the wholeiterative alg

3. Design

Learninof ARTNNIt is impracages of tranphase selecdent featurnodes. It sethe numbeever, it willsame note.

When ato a dynama fault systeposition thfundamentthe measurin the powetion. The fiin polarityfictitious sotime. It resmeasured ccurrents anThat is to ster reveal tselects faulto reduce osification aphase selec

(a) Ir =Ur =

(b) Is =Us =

(c) IAB C UAB = U

(d) r =(e) p =where I anrespectivelof voltage(A, B andand zero sesequences.

Therefoponent as

A sing

tion rles netrainin

Simul

e thremulatK in

iven in

Param

ccordiode itrainihas

tilizingcteris

improlassifiries o

resu

KN uble 1.A to

tive apointesultsve the.

ramet.05. Itcondy lots

Traini

ainingoundhree ptive angles = 0 and 45; fault location distance fromion of relay L = 1, 10 and 40 km; fault path resistance, 20, 50, 80 and 100 . Sampling frequency is set toHz. In this paper, the vector of current and voltage canorithm is linear.

of ARTNN based phase selection relay

g of ARTNN is non-supervised, so training patternshould try to embody nature of research question.tical that directly utilizing phase currents and volt-smission line as input vectors of ARTNN basedtion relay, because these quantities have no evi-

es so that there will form too many classificationems that by decreasing the value of vigilance ,

r of classification nodes could be reduced. How-result in sorting different kind of patterns into the

fault occurs in the electric power system, it leadsic transition from the normal system condition tom condition. This case can be analyzed by super-

eorem. The superposition theorem is one of theal tools in circuit analysis. It allows consideringed currents and voltages as the sum of all sourcesr system and a fictitious source at the fault loca-

ctitious source is equal in magnitude and oppositeto the prefault voltage at the fault location. Thisurce is applied to the system at the fault inception

ults in changes in the magnitude and phase of theurrents and voltages. The changes in the measuredd voltages are directly related to the fault type [10].ay, the components generated by fault could bet-he nature of the fault types. Therefore, this papert components as input vectors of network in orderutput nodes and not to debase precision of clas-t the same time. Input vectors of ARTNN basedtion relay are shown as follows:

postfault Ir prefault Irpostfault Ur prefault Ur

postfault Is prefault Ispostfault Us prefault Us

= IA IB, IBC = IB IC, ICA = I= UA UB, UBC = UB UC, UCApostfault r prefault rpostfault p prefault p

d U represent amplitude of current and voltage,y, and represents the angle between the phaseand current, the subscript r represents a phaseC), s a sequence component (positive, negativequences) and p represents positive and negative

re, there are 23 input vectors. Selecting fault com-input vector ensures that ARTNN based phase

Fig. 2.

selecenablittle

3.1.

Thfor sipointare g

3.2.

Aput nfromANNby ucharatherthe cA selationlinein Taphaseincepfromtion rachierelay

Pa = 0whenfied b

3.3.

Trto grand tinceplocatR = 02000IA

C UA

le line diagram of the power system used for simulation testing.

elay has fewer output classification nodes andtwork to attain stable classification ability withg patterns.

ation model of power system

e phases power system, shown in Fig. 2, is chosenion testing. Phase selection relay is located at theKN line. The details of the sample power system

Appendix A.

eters of ARTNN

ng to the fault types, the number of ARTNN out-s set to 10. The value of vigilance is derivedng patterns. Generally, the approach based ongood generalization capabilities. In this paper,

fault components, which can better reflect thetics of different fault types, ARTNN can fur-ve generalization capabilities. At the same time,ed precision of network could still keep fine.f simulation tests have verified this idea. Simu-lts of phase A to ground fault occurred in thender various fault conditions, have been shown

Training patterns of ARTNN are generated byground faults occurred under the conditions: faultngles = 0 and 45, fault location distance LK is 15 km, fault path resistance R = 0 . Simula-have indicated that fewer training patterns couldtraining purpose of ARTNN based phase selection

ers of ARTNN are set as follows: = 0.2; = 0.85;should be noted that these parameters do not varyitions of power system vary, which has been veri-of simulation tests.

ng patterns set

patterns are generated by simulating single phasefaults, two phase or two phase to ground faultshase faults under different fault conditions: fault

118 Y. Yang et al. / Electric Power Systems Research 76 (2005) 115120

Table 1Simulation results with phase A to ground fault occurring

Patterns Fault condition

L = 90 km, R = 100, = 0 L = 45 km, R = 10 L

= 54 = 90

1 N = 1, S = 0.9465 N = 1, S = 0.9466 N = 1, S = 0.9483 N2 N = 1, S = 0.9442 N = 1, S = 0.9441 N = 1, S = 0.9455 N3 N = 1, S = 0.9481 N = 1, S = 0.9482 N = 1, S = 0.9493 N4 N = 1, S = 0.9494 N = 1, S = 0.9549 N = 1, S = 0.9568 N5 N = 1, S = 0.9563 N = 1, S = 0.9620 N = 1, S = 0.9635 N6 N = 1, S = 0.9601 N = 1, S = 0.9654 N = 1, S = 0.9670 N7 N = 1, S = 0.9591 N = 1, S = 0.9642 N = 1, S = 0.9658 N8 N9 N

10 N

S: matching le A to g

Table 2Simulation rebefore fault in

Fault type

Phase AB faPhase ACg

be calculatalgorithmconsecutiveas the training is 2 3patterns is

As mentalgebra proafter 6080should be ptraining is

4. Verifyinselection r

Verifyinunder varioscillation,relays respand 3.

Table 3Simulation reafter 10-ms fa

Fault type

Phase AB faPhase ACg

0 denotes no

Gener

peratAfte

ompoithmecisioeeds 2opingIn te

rbitraNN b

re alwn thisresenccurr

n theN = 1, S = 0.9564 N = 1, S = 0.9612 N = 1, S = 0.9629N = 1, S = 0.9557 N = 1, S = 0.9603 N = 1, S = 0.9619N = 1, S = 0.9569 N = 1, S = 0.9612 N = 1, S = 0.9627

vel of verifying pattern. N: output of verifying pattern. N = 1 denotes phase

sults of using patterns corresponding to ten sampling pointsception

Verifying pattern

1 2 3 4 5 6 7 8 9 10

ult 0 0 0 0 0 0 0 0 0 0round fault 0 0 0 0 0 0 0 0 0 0

ed according to the sampling point and the Fourierwithin half cycle. For every simulation test, six

vectors after 10-ms fault duration are selecteding patterns. So the number of simulation test- 5 10 = 300, and the number of total training

2 3 5 10 6 = 1800.ioned above, the algorithm of ARTNN is a simplecess, so weights of ARTNN quickly achieve stable

4.1.

(1) O

c

r

dn

(2) C

a

Aa

Ipo

i

iterative loops. It should be noted that this trainingerformed before the network is carried out, i.e. theoff-line.

g performance of ARTNN based phaseelay

g patterns are generated by different type faultsous system conditions, including normal anddifferent source impedances, etc. Parts of the

onses to verifying patterns are shown in Tables 2

sults of using patterns corresponding to ten sampling pointsult duration

Verifying pattern

1 2 3 4 5 6 7 8 9 10

ult 0 1 1 1 1 1 1 1 1 1round fault 1 0 1 1 1 1 1 1 1 1

output; 1 denotes correct output.

mentioknowleterns stin Sectto thevectorditionsrepresedescribperformwill ha

5. Experim

The propprototype.processor sDSP proceto the lowthis paper,on its good= 90 km, R = 100, = 144 L = 45 km, R = 10, = 216

= 1, S = 0.9500 N = 0, S = 0.8431= 1, S = 0.9470 N = 0, S = 0.8418= 1, S = 0.9501 N = 0, S = 0.8432= 1, S = 0.9511 N = 0, S = 0.8429= 1, S = 0.9576 N = 0, S = 0.8459= 1, S = 0.9609 N = 1, S = 0.9614= 1, S = 0.9589 N = 1, S = 0.9589= 1, S = 0.9558 N = 1, S = 0.9541= 1, S = 0.9552 N = 1, S = 0.9524= 1, S = 0.9561 N = 1, S = 0.9529

round fault; N = 0 denotes no output.

al condensed summary

ion speed of ARTNN based phase selection relay.r fault occurs, at least 10 ms is necessary for faultnent calculation when half-wave Fourier algo-is utilized. For reliability, the network makes an by judging four points consecutively, whichms, so the whole operation time is about 12 ms.with mal-output of pattern.

rms of theory, artificial neural network may reflectry complex function, but in practical applications,ased protection relays cannot be perfect, so thereays mal-operation zones of protection relays.case, ARTNN based phase selection relay will

t good performance. For example, after a faulted, which is not embodied in the patterns storednetwork, the relay will have a mal-output. Asned in Section 2.1, ARTNN could learn fast newdge and simultaneously avoid modifying the pat-ored in the network, so according to the algorithmion 2.2, an inactive output node could be addednetwork and its weight vector is set to the inputP, which is derived from the mal-operation con-, i.e. the added output node denotes a sort, whichnts the above-mentioned conditions. Based on theed algorithm, this process is very fast and can be

ed on-line. Then, the relay could be rectified and

ve a correct output if later similar conditions occur.

ental and eld results

osed algorithm has been tested in an experimentalThe hardware of this prototype [11] is a multi-ystem comprising a master controller and severalssors. The DSP card has been designed accordingcost principle and sufficient computing ability. InTMS320F206 has been used in the relay, based

performance and flexibility to meet the needs

Y. Yang et al. / Electric Power Systems Research 76 (2005) 115120 119

al relay

Table 4The field dataCompany Lim

Date Faul

20014 590529

20014 590529

20018 54088

20017 54077

of signal pOne of typdog has beeoperation.processingof the neur

Some fiof East ChTable 4. Thin Table 5.

Table 5The experime

Output

Phase AgrouPhase BgrouPhase Cgrou

() A correctselection.

onclu

newFig. 3. Block diagram of one of typic

derived from numerical fault recorders of East China Gridited

6. C

A

t phase Data file Sampling

frequency (Hz)Length (km)

-C Ping4386.x 01 3840 143

-A Ping438a.x 01 3840 143

-B 011882.eve 2400 138.69

-B 012921.eve 2400 139

rocessing and control applications in protection.ical relay hardware is shown in Fig. 3. A watch-n fitted to the master microprocessor to check its

According to the hardware design, digital signalalgorithms are used to calculate the input vectorsal network.eld data derived from numerical fault recordersina Grid Company Limited have been shown ine test results with those field data have been shown

ntal test results with the field data

5905-A 5408-B 5407-B 5905-C

nd

nd

nd

output of fault phase selection; ( ) no output of fault phase

in this papnents are caand then Afeatures ambetween thscheme hasing patternneural netwscheme canby dynamiand experimcorrectness

Acknowled

The authsupport givogy FoundExcellent Y

Appendix

Paramet

Positive s

Z1 = 0.8C1 = 0.0Zero sequ

Z0 = 3.6C0 = 0.0hardware.

sion

fault phase selection scheme has been proposeder. Firstly, the power frequency fault compo-lculated by utilizing half-wave Fourier algorithm,RT based neural networks are used to extractong those fault components and discriminate

e fault phase and the non-fault phase. The newsuch advantages as higher flexibility, fewer train-

s and quicker training speed compared with BPork based protection schemes. Furthermore, thisbetter adapt itself to varying system conditions

cally adding output node. Extensive simulationsental field data tests have validated this scheme

.

gement

ors would like to express their thanks to financialen by the Power Electrical Science and Technol-

ation of XJ Group and University Foundation foroung Teacher in Shanghai

A

ers of perfect transposition line (Clarke model):equence impedance:

6 mh/km, R1 = 0.027 /km,123F/km

ence impedance:

mh/km, R0 = 0.1948 /km,051F/km

120 Y. Yang et al. / Electric Power Systems Research 76 (2005) 115120

The parameter of sources is not fixed; there are two set ofparameter for each source.

The first set of parameters of source 1:

Ump = 330 kV, f = 50 HzPositive sequence impedance:Z1 = j45.149 Zero sequence impedance:Z0 = j23.321

The first set of parameters of source 2:

Ump = 330 kV, f = 50 HzPositive sequence impedance:Z1 = j45.149 Zero sequence impedance:Z0 = j23.321

The second set of parameters of source 1:

Ump = 330 kV, f = 50 HzPositive sequence impedance:Z1 = j6.06 Zero sequence impedance:Z0 = j7.07

The second set of parameters of source 2:

Ump = 330 kV, f = 50 HzPositive sequence impedance:Z1 = j4Zero seqZ0 = j7

References

[1] S.K. Chakravarthy, C.V. Nayar, N.R. Achuthan, Applying patternrecognition in distance relaying. Part I: concept, IEE Proc.-C 139(July (4)) (1992) 301305.

[2] S.K. Chakravarthy, C.V. Nayar, N.R. Achuthan, Applying patternrecognition in distance relaying. Part II: feasibility, IEE Proc.-C 139(July (4)) (1992) 306314.

[3] T.S. Sidhu, H. Singh, M.S. Sachdev, Design, implementation andtesting of an artificial neural network based fault direction discrimi-nator for protecting transmission lines, IEEE Trans. Power Deliv. 10(April (2)) (1995) 697706.

[4] A.G. Jongepier, L. van der Sluis, Adaptive distance protection ofdouble-circuit lines using artificial neural networks, IEEE Trans.Power Deliv. 12 (January (1)) (1997) 97105.

[5] T.D. Alstein, B. Kulicke, Neural network approach to fault classifi-cation for high speed protective relaying, IEEE Trans. Power Deliv.10 (April (2)) (1995) 10021011.

[6] Y.Q. Duan, J.L. He, Distance relay protection based on artificialneural network, Proc. Chin. Soc. Electric. Eng. 19 (May (5)) (1999)6770.

[7] P.F. Yan, C.S. Zhang, Artificial Neural Network and EvolutionaryComputation, Tsinghua University Press, Beijing, China, 2000.

[8] G.A. Carpenter, S. Grossberg, ART 2: self-organization of stablecategory recognition codes for analog input patterns, Appl. Opt. 26(December (23)) (1987) 49194930.

[9] G.A. Carpenter, S. Grossberg, D.B. Rosen, ART2-A: an adaptive res-onance algorithm for rapid category learning and recognition, NeuralNetw. 4 (1991) 493504.

[10] A. Apostolov, Implementation of a transient energy method for direc-tional detection in numerical distance relays, in: Proceedings ofthe IEEE Power Engineering Society Transmission and DistributionConference, vol. 1, 1999, pp. 382387.

. Tai, Z.J. Hou, X.G. Yin, X.H. Li, D.S. Chen, Wavelet-basedround fower S4.1 uence impedance:9.4

[11] NgPault protection scheme for generator stator winding, Electricyst. Res. 62 (1) (May, 2002) 2128.

ART artificial neural networks based adaptive phase selectorIntroductionBasic theoryThe principle of adaptive resonance theory based neural network (ARTNN)Simple description of ART2-A algorithm

Design of ARTNN based phase selection relaySimulation model of power systemParameters of ARTNNTraining patterns set

Verifying performance of ARTNN based phase selection relayGeneral condensed summary

Experimental and field resultsConclusionAcknowledgementAppendix AReferences