APPLICATION OF ARTIFICIAL NEURAL NETWORK I N PROTECTIVE RELAYING OF TRANSMISSION LINES S.A. Khaparde, P.B. Kale Electrical Engineeri'ng Department I n d i a n I n s t i t u t e of Technology Bombay-400076, I N D I A S.H. Agarwal T e s t i n g Department T a t a Electric Company Trombay, Bombay, I N D I A That the A r t i f i c i a l Neutel Netwcork (Awl) can pi3o.m the pattern classification i n excellent f a s h i m is already established i n the literature. Here, we envisage the relay as a pattern classifying device. ThFsopensan~~dimensioninrelayphFlosophywhich d wi& investigatiaw. Keeping themirelay framKx3c intact, this peper repzts the findings about the feasibility of using Awl in pratection of tranmlission lins. " I moddl is explored f o r the application and is fand to yield -aging results. The inprt variables are quantified over the opera* range which eases the arithmetics of the .-im The traiaing is performed i n off-line mode and the camrged weight matrix is stceced f o r on-line use. -, this huable heginning has cope f a a lot of further refinemnts which are rrnder investigatia. 1. IFmCUXmm The techniques available f o r transmission line pmtection as inp1i n the analog and static relays include, directlcnal UlverSBctkne over current, d i r e c t i d distance, p i l o t -on w i t h a ammdcaticm channel betuem te"lS. -, in the area of d i g i t a l protech'on of transmission lines, distance relaying principle has received m n s i i k a b l e attentim. DigiW distance relaysmeasuretheinpedancebatvreenarelaylocatim and the fault point, and thus determFne i f a fault is internalorBRernaltoap" zone. The hpedmce is defined a t -tal frequency. Qnrent and voltage signals used as inplts d l y txntain the f " e n t a l culpn@nts of interrst and other signals of a wide f " c y qy=t" which may be terraed as noise " l 'a).Durllrgeachstmpling period, the signal sargles i n the data w i d o w are -by filter and then Befinetheimpedance to arrive at the t r i p signal. several algorithm are t r i p decision f r u a available to extract the signals minimizing the decision time [l,21. the -le These algorithm include mier algorithn with full or partial cycle, Kahm filter etc. trim The-im distance relays claim some new and unlqlm capabllitiea 121. cunnunication circuits provide local and "te access to settings, fault recards and other infarmation. Data prucessing f o r plrposes ather than p"ztim8 Like f a u l t locating, event * etc.c%nbeincl~. This paper is an atto explore haw A r t i f i c i a l m a l Network (ANN) can be m i e d to realize a distance relay in its sinplest form. framecnrk of themi-relay is retamed here. we envisage a relay to be a 'pattern c l a s s i f e r ' and pmvideaA"lrPdelwh.i.chcanperf'ormthistkskinan excellent manner. = s e c t u n describes H i c i a l Neural "As. sincethemotivaticnofprpsentwmkisbesedonthe f a c t that I\Ewmodel can be used as a pattan classifier andarelaycanbeviewedtobepatterncLassifk, it d d be wmthrtu'le to review the Awl mdel s h i c h caters to this task. The sinplemodel-byB. Widrav [SI, A[lALIEIE, is describ~~I here. o"E) -tially of adaptive threshold bgiC d m which can be used in trainable pattern recogniti.cn systen. Aaaption is effected byIMsorrewardand-algtxithn. Awptive LmEX 91TH0374-9/91/0000-0122$01.~1991 IEEE Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY BOMBAY. Downloaded on December 4, 2008 at 06:26 from IEEE Xplore. Restrictions apply. In its sinplest form AwLINE can be shorm to r e a l i z e linearly separable functicns. figure 1 shavs basic building block of weighted are t Wak I WUC, ... by set a of 4. Since a relay can be looked upon as a pattern classifier, the which essentially does the sane task can replace the cmnentl'aral relay. The input quantities are voltage and current values. Since the aim of the present work is to establish.the f e a s i b i l i t y of the pmpcsed ccmcept we have e the fra"rk of the micxqmcessor relay and lwdware aspects to their SLrplest farm. Figure 2 shows the schematic diaqran of t h e micmyxessar based neural relay. The m w set of wights which are worked i n off-line rmde are ttxs stored in the micruprocessor for on-line awlication. m. coefficient $ . mc4TIcN = T ,Wml , the elaaent produces an analog ~,theinnerprcductYk~<y(.Thebiasweight Wok is cavlected to a oo~lstentinput xo = +1 and it c"1s t h e threshold level. Decisicns are made by a 2-level quantizer. The quantizaticn of a l l the i n p t vector is prcpxed which is suitable far micmprocessar framwork described i n next section. Input patt- ern + yk *lk0 Analog output quantize each inplt almg its axis fran " I to Itlilxi" value and the n m b r of quanta w i l l depend upm the resolution of the A/D m e . This facilitates the redwt.~ 'on of the input vector to be descritized i n finite nunber of points. such discretized input can be readily processed nunerically by the mimqrocessor. we assign a binary string to each value of quantun. The input vedar is formed by a x l c a t i n g t h e c o d * e s ~on each axis. All quanta form a code m a t r i x which if f u l l rank has '2k0 '3k0 d e f i n i t e omvergeme. Chosen ccde shculd be linearly independent. (xrtofthe codes dis~ussedin ref. [6], we have chcsen 01 contrast here for our a@icatim. I l l u s t r a t i o n of the coding is given i n next section. Illustratim: Cansider a relay w i t h 2 inplts, voltage and current with their ranges 0-5 V and 0-1 A respectively. xet both of t h e n be quantized i n 5 parts each. Then the inpxt space would look like as si" i n Figure 3a. Fig. 1 Desired response input (training signal) I& us use 01 contml coding for the case, then the coding matrix is as sham i n Figure 3b. For input v = 2.5 volts and I = 0.3 A , we have input vedar [ O O ~ U O I . U ~ J ~ . In this case, we have 10weights, one to each b i t i n the input vector. Let the d g h t vector be [+1 -1 2 3 0 2-3 1 0 -11. We get the following e a c t e r i s t i c s for givencase (fig. 4). response' is a special input signal used to train t h e neuron. An e o n algorithm adjusts ueights so that the outplt to the gatterns w i l l be as clase as posslble to ttmr msFective desired "sp0"Ses. In other wxds the mdel can realize a l m e a r l y separable logic functicn d c h is a separating plane between the output patterns which divides t h e output space having different patterns. Ihe 'desired 4.2 Aproposed schenatic of Neural Relay is as shown i n figure 2. As shown in the Figure, the relay takes the inputs fran the pmer system, a c h are passed thrcugh surge f i l t e r s to zaaove high frecperq sut-gq. he wider, for two-dunens * ionalcase, Y = x1 w1 + % W 2 + W o = 0 (1) w1 %=-3 - i;s; *l B g a a t i m ( 2 ) r e p r e s e n t s a line in * i d plane. 'ihe r e p r e s s t a t i o n w i l l produce outplts w i t h qposite signs on either si& of the line. This sqgests that a caabina+ion of or netwark of muons may be used to realize &inear separating PLaneS r71. -Aspects sigmls are then caditimed to make measuranents. quantities are sanpled using -le and hold circuit and corresponiing d i g i t a l values are abtained by using analog to d i g i t a l cm-. lhese d i g i t a l values are fed to a miQ9ccrrFpltet. A b i t pattern coarespading to every input value is f d by pattern enccdfr, using pmper coding schele. The input vector thus formed is fed to the AIXUJNE and weighted sun is obtained, which is passed througfi hard limiter functim. "?E "p"t of t h i s pmduces the binary-to t r i p (+1) or no t r i p (-1) signal. m i s signal is fed back to power system The 123 Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY BOMBAY. Downloaded on December 4, 2008 at 06:26 from IEEE Xplore. Restrictions apply. . ------- ANALOG INPUTS(V, I -,PHASE - - - - -etc.) - ---T- --RELAY --l- SETTINGS OUTPUT - - - - - -DIGITAL -- f----- FILTER SIH CIRCUIT 5 - ' -L, PATTERN ----c WEIGHTED' AID CONVERTER I ENCODER + Tripping zone 1.0 - No trip zone I 0.8 4 SUMMER 0.6 I 3 0.4 2 1 t v 0.2 2 (0,O)1 Quntum no. -PI 2 4 3 3 v Fig. 4 5 4 Fig. 3a 1 1 1 1 1 1 0 1 1 0 0 1 0 0 0 4.3 ~ A s p e c t 8 Fig.3b 124 Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY BOMBAY. Downloaded on December 4, 2008 at 06:26 from IEEE Xplore. Restrictions apply. = presentweight = presentoutplt d = desiredoutplt x(k) = present inplt W(k) y carparable w i t h canventional relays even though the wallel processing is not implmted. The decision space is V-I plane rather than R-X plane. T h e d a y setting can be changed by changing weights and time &lay settings can also be achieved. softwamaagani zatial: sanplenet : To generate the training and sanple Set rain net Toolnet 4.4 : +- train the V using above generated training sanple s e t %e a w l i c a t i o n is yet i n the develqing stages leaving many refinanents. The adaptive relaying [8] can be a&eved using layered d e l where additicnal inpa l i k e direction of current and approximate value of fault resistance w x l d change the operating characteristics of the relay. - : To experiment on the weight matrix, and observe effect of a transformation of weight matrix on the - a m i S t i c s realized. phadke and J.S. n'lorp, canplter m Y Fmer Systems, John Wiley and Sons, 1988. A.G. Observations IE "rial b fa ~aurseon "Microprocessor Relays and protection Systems," NO. 88, M 0269-1-RiR. ~Ulcui.ng factors were studied and sane osrsenmtions were made as f o l l m : P.D. Wasserman, CaFPlling, 'Iheory and Practice," Van Nostrand Reinhold, 1989. Placewnt of training sample: This is a critical fador i n training since it f a r m basis f o r the training. R. x , i v , Introductim to C m ~ ~ i nw igt h Neural ~ e t s , " Assp Magazine, April 1987. order of presentation of training samples: I f t h e sanre training set is presented to a ADAtINE i n different order of training sanples, then they may B. Widrcw, R.G. Winter and R.A. Baxter, ''h- - realize s l i g h t l y different characteristics. order of quantization: As the quantization is immased, m ~ l r e91130th surfaces can be realised. s a . y a m d , R. subramaniani "Optimal codbes far Leanung Control~systenS",Fifth N a t i d SystanS COnferenoe, India, 1971. Copingschenes: Anunberofcodingschemeswere Identity coding and 21 contrast code did not prcduce encowaging results, while 01 contrast gave satisfadary cutputs. tried. F.W. anith,"A Trainable Nonlinear Function Generatnr", IEm, Trans. of Autanatic ccntrol, vol. 11, No. 2, pp 212-218, A p r i l ~ 1 6 6 . superposition of regions: A new weight vector can be farmed by adding 2 weight m a t r i c e s i n different proport5ms. mis may be useful i n cutting daun undesirable part of a regial etc. A.G. Phadke,. S.H. U t Z , "pdaptive e y i n g , " IEEE ccnprter Agplications i n Paver, Vol. 3, No. 3 PP 47-51, J d Y 1990. Using a different concept of sun of r u m (SRI and of colurns (sc), we can explain n?latialship of sane of the above factors on the characbxistics realized. W - S R and sc o~~cepts imrolves basically contrikrtion of SUR of weights along that axis f o r t h a t axis for that quantun. Fbr any cell, the analog output is fcnmd by adding this sua for the c*aresponding quanta along every axis. Hence the dichotany alarg any quantun i s depedent only ~1 the sun a t that quantun. l h i s can be used far shifting the characteristic etc. Also a non-closed curve can be gtaerated using this cc(Icept [Ref. 71. SUR i As expcted t h e pmposed AaAcINE model is able to .locate the cperating paint correctly i n the &cision space. This ccnfirms feasible use of W as relay. "he pmposed schaae considers two inputs viz. voltage andcurrentvalueS. Theschemecanbeextendedto three inplt variables to include the phase angle, it can be realized for of cross-cheduq. The suggested quantificatmn of mputs is suitable for micraprocessor f r a " k W weight matrix, calculated i n off-line mDde is stared in "y. ltle on-line time tfor arriving a t decision is -. 125 Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY BOMBAY. Downloaded on December 4, 2008 at 06:26 from IEEE Xplore. Restrictions apply.
© Copyright 2025 Paperzz