SIMULATION OF FAULT DETECTION FOR PROTECTION OF TRANSMISSION LINE USING NEURAL NETWORK

Transcription

1 Internatonal Journal of Scence, ngneerng and Technology Research (IJSTR), Volume 3, Issue 5, May 04 SIMULATIO OF FAULT DTCTIO FOR PROTCTIO OF TRASMISSIO LI USIG URAL TWORK Smrt Kesharwan #, Dharmendra Kumar Sngh # # MTCH Scholar, # Head of department Dept. of, Dr. C. V. Raman Unversty, Blaspur, Inda Abstract Transmsson lne among the other electrcal power system component suffer from unexpected falure due to varous random causes. Because transmsson lne s qute large as t s open n envronment. A fault occurs on transmsson lne when two or more conductors come n contact wth each other or ground. Ths paper presents a proposed model based on Matlab software to detect the fault on transmsson lne. The output of the system s used to tran an artfcal neural networ to detect the transmsson lne faults. Fault detecton has been acheved by usng artfcal neural networ. Keywords: Transmsson lne, Faults, Protecton, neural networ. I. ITRODUCTIO Transmsson lne s the most lely element n the power system to be exposed especally when ther physcal dmenson s taen nto consderaton []. Ths paper has concentrated on understandng the behavour of the transmsson lne phase voltages and currents as a consequence of faults. The objectve of ths wor s to study and employ neural networ technques as a relable tool to dentfy or detect faults n a transmsson lne system. Artfcal neural networ are a powerful to use n transmsson lne fault dentfcaton, classfcaton and solaton. The parallelsm nherent n neural networs enables them wth faster computatonal tme than tradtonal technques. Hence, usng ths technology n transmsson lne fault dagnoss does valdate ts usefulness and encourages engneer to use ths technque n other power system applcatons. The man objectve of ths paper s to develop neural networ based autonomous learnng system that acqure nowledge ncrementally n real tme, wth as lttle supervson as possble. To deploy effectve strateges for practcal applcaton of such system for fault dentfcaton and dagnoss. For protecton of transmsson lne the fault dentfcaton, classfcaton and locaton plays an mportant role. Due to lmted avalable amount of practcal fault data, t s necessary to generate examples of fault data usng smulaton [6]. To generate data for the typcal transmsson system, ths paper used to generate the fault current and voltage for all types of transmsson lne fault. The output of ths paper s used to generate smulaton data for the model of transmsson lne n normal and faulty condton to detect the fault. II. URAL TWORK Artfcal eural etwors (As), or smply called neural networs, use the neurophysology of the bran as the bass for ts processng model []. The bran conssts of mllons of neurons nterconnected to each other through the synapse. In the learnng process, the weght of the synapse s ncreased, decreased or unchanged. A. euron Model euron (also called node or perceptron) s modelled as follows. X j W j net y =f ( net ) b Fg. euron Model [] net W X +b j ach node has nputs connected to t and weghts correspondng to each nput. ach node only has one output. The above neuron, based on the above notaton, s called neuron. It has j nputs X j and one bas b. ach nput correspond to a weght W j, thus there are j weghts n the neuron. The output of the neuron y s produced by a functon of net where net Wj X b j Ths functon s called actvaton functon. There are many types of actvaton functons; two examples are hard lmt and log-sgmod functons. j All Rghts Reserved 04 IJSTR 367

2 Internatonal Journal of Scence, ngneerng and Technology Research (IJSTR), Volume 3, Issue 5, May 04 f(net ) 0 net Har dl m t Func t on f(net ) 0 net Log- s gmo d Func t on Fg Actvaton Functons [7] Hard lmt functon s defned as 0 net f (net ) net and the log-sgmod functon s defned as F net B. etwor Model e net very neuron can be nterconnected to other neurons by usng output of neurons as nputs to other neurons. The nterconnecton of neurons forms layers of neural networ. A neural networ conssts of three types of layer, nput layer, hdden layer, and output layer [4]. I nput s Input layer Hdden Layer Fg.3 A Layers [7] Output layer Out put s The number of nputs n a neural networ s equal to the number of nodes n the nput layer. Smlarly, the number of outputs n the neural networ s equal to the number of nodes n the output layer. The number of hdden layers and the number of nodes n the hdden layer are varyng dependng on ts applcaton. There are two dstnctve networ topologes wth regard to the way neurons are connected namely feedforward and feedbac networ. In the feedforward networ, an output n a layer (except output layer) s an nput n the next layer. In feedbac networ, an output n a layer can be ts own nput or nput of neuron n the prevous layer. C. A Learnng To get the ntended outputs for the gven nputs, the networ weghts need to be adjusted. The process of weghts adjustment or called networ learnng/tranng s done teratvely by presentng a set of nput data and desred output data. Ths type of tranng s called supervsed learnng [8]. The weght update can be done n ether batch or ncremental mode. In batch (epoch) mode, the weghts are updated after all tranng data n the tranng set has been presented. The ncremental mode updates the weghts every tme a data n the set s presented. Two ssues n updatng weghts are: when to stop updatng,.e. when to stop the tranng, and how the weghts are changed. The tranng can be stopped n two ways: usng maxmum epoch and usng a cost functon. An epoch refers to a complete tranng data set. Tranng data for 0 epochs means the weghts are updated wth the learnng rule contnuously untl nput data set has been presented for 0 tmes. A cost functon s a performance measurement. etwor tranng often uses the Mean Square rror (MS) as the cost functon. MS s defned as follows (d (n) y (n)) n s the number of pattern n data set, d (n) and y (n) are the desred output and the output at layer for n th tranng pattern respectvely. When there s more than one output, the functon becomes ( d (n) T y (n)) ( d (n) y (n)) n Where d and y are column vectors of desred output and output respectvely. The tranng adjusts the weghts by mnmsng over all the tranng set. The tranng stops when a specfed value of cost functon s reached. The weght (and bas f applcable) update follows a certan optmsaton technque. The weghts are ether ncreased, decreased or the same. The change n a weght s as follows W (t+) = W (t) + W The W s the weght correcton. The weght correcton s a functon that mnmses the error. In the gradent descent algorthm, the weght correcton s the negatve gradent of an mmedate square error for a pattern (n) ΔW (n) η W Where (n) s the mmedate square error at n th pattern and s the coeffcent of change or the learnng rate. Square error at n th pattern s n e n ( d n y n) All Rghts Reserved 04 IJSTR 368

3 Internatonal Journal of Scence, ngneerng and Technology Research (IJSTR), Volume 3, Issue 5, May 04 D. Bacpropagaton Bacpropagaton networ (BP) s a feedforward networ traned usng bacpropagaton algorthm. The bacpropagaton algorthm developed by McClelland and Rummelhart (986) used gradent descent learnng rule to update the weghts [5]. Durng tranng, each nput s forwarded through the ntermedate layer untl outputs are generated. ach output s then compared to the desred output to get the errors that wll be transmtted bacwards to the ntermedate layer that contrbutes drectly to the output. Based on these errors, the weghts are updated. Ths process s repeated layer by layer untl each node n the networ has receved an error sgnal that descrbes ts relatve contrbuton to the total error [7]. The gradent descent algorthm suffers from slow tranng tme. Some other fast algorthms such as Levenberg-Marquardt, Quas-ewton, and conjugate gradents algorthms have been used to optmse the learnng rules n BP [0]. III. MODLLIG TH POWR TRASMISSIO LI SYSTM The analyss of fault currents wll gve nformaton about the nature of the fault. Let us consder a faulted transmsson lne extendng between two power system as shown n fg. A 400 KV transmsson lne system has been used to develop and mplement the proposed A s. Fgure shows a one lne dagram of the system that has been used throughout the wor. The system consst of two generators of KV each located on ether ends of transmsson lne along wth a three phase smulator used to smulate faults at md poston on transmsson lne. Let us consder a faulted transmsson lne extendng between two sources as shown n fg. The faulted transmsson lne s represented by dstrbuted parameters. As an applcaton of 00 Km overhead transmsson lne wth the parameter of the transmsson lne model shown n fg. = 8.5+j308 Postve sequence Capactance = 3nf/m Zero sequence Capactance = 8.5nf/m Length = 00 Km. Ths above sngle lne dagram was modeled by usng MATLAB009a and smulated usng the smpower system toolbox n Smuln. A snapshot of power system model used for obtanng the tranng and test data sets for neural networ s shown n fg. Fg 5 Snapshot of transmsson lne system In the above fgure three phase V-I measurement bloc s used to measure the voltage and current sample at source end. The transmsson lne s dvded nto two lnes lne & lne each lne s 00 Km long. Model of three phase fault smulator s used to smulate varous types of fault. In transmsson lne faults are classfed as sngle lne to ground fault, lne to lne fault, double lne to ground fault and three phase fault. Fg. 6 Current waveform of healthy networ Fg.4 Sngle lne dagram of transmsson lne system Source voltage:kv(both) Source Impedance : Postve Sequence =.3+j5.0 Zero Sequence=.33+j6.6 Frequency = 50 Hz Transmsson lne mpedance: Postve Sequence =8.5+ j 94.5 Zero Sequence Impedance Sngle lne to ground fault occur when one of the phase s shortened to the ground. Durng the fault the mpedance, Z s not necessary zero, but t mght have a non zero mpedance, but stll much smaller than the lne mpedance. The magntude of current n a faulty lne rse sgnfcantly hgher than the normal operatve current and the voltage does not go through sgnfcant change n magntude [9]. The followng waveform shows the rse of current when sngle lne to ground fault occur on transmsson lne: All Rghts Reserved 04 IJSTR 369

4 Internatonal Journal of Scence, ngneerng and Technology Research (IJSTR), Volume 3, Issue 5, May 04 the nput layer, 0 neurons n the hdden layer and 6 neuron n the output layer. Ths fgure called as the performance plot of the neural networ. Fg 7 Current waveform of faulty networ The value of the three phase voltages and currents are measured and modfed accordngly and then fed nto the neural networ tool called as neural fttng tool (nftool) as nput. The Smpower system toolbox has been used to generate the entre set of tranng data for the neural networ n both fault and non fault cases. In ths paper for desgn and modelng of transmsson lne MATLAB (R009a) s used. IV. DSIG URAL TWORK FOR FAULT DTCTIO In order to desgn a neural networ for addressng the fault detecton problem several dfferent topologes of mult layer perceptron are studed. The crtera used to mplement and select an approprate MLP neural networ for the problem of fault detecton does tae nto consderaton the factors such as the networ sze, the sutable learnng rule and the sze of the tranng data[8]. A. Tranng procedure and learnng rule The bac propagaton learnng rule used n perhaps n over 80-90% of practcal applcaton. However the standard BP tranng algorthm s slow [8]. Snce t s generally requres small learnng rate for stable learnng process so that change n the networ weght usng the steepest descent algorthm reman small. Some technques to mprove the standard bac propagaton method such as the addton of momentum terms and adaptve learnng rate as well as alternatve methods to the gradent descent such as Levenber- Marquadt optmzaton routne can also be used. In the frst stage whch s the fault detecton phase, the networ taes n sx nputs at a tme, whch are the voltages and currents for all the three phases. Here the value we tae for no fault and sngle lne to ground fault condton. The output of the neural networ s that target value s equal to wth the nput or not. When target and nput values are same that shows the power system s healthy. If any changes n the target value t shows that any transmsson lne fault has occur on the system. Fg shows the tranng process of the neural networ wth confguraton ( 6 neuron n Fg 8 Mean Square rror performance of the networ Selectng the rght structure sze of the networ reduces not only the tranng tme but also sgnfcantly mpacts the generalzaton and representatonal capabltes of the traned networ [0]. The no of hdden layer and neuron n these layers are mportant factor n determnng the optmal sze and structure of the networ. The networ presented here represents only a sample of those that we modeled and correspond to the best results that were obtaned after extensve tral and error procedure. B. Testng the neural networ for fault detecton A test set was created to analyze the performance of the proposed networ. The selected networ s to recognze and classfy correctly both the normal condton as well as fault condton. The test data has 47 samples. The nput has 47x6 matrx structures wth three values of voltage and three values of current n a column. Once the neural networ has been traned ts performance has been tested by plottng the best lnear regresson that relates the target to the output as shown n followng fgure: Fg 9 Regresson ft of the output vs. target for the healthy networ All Rghts Reserved 04 IJSTR 370

5 Internatonal Journal of Scence, ngneerng and Technology Research (IJSTR), Volume 3, Issue 5, May 04 eural networ are ndeed a relable and attractve method for transmsson lne faults scheme especally n vew of ncrease complexty of the modern power systems. Bac propagaton networ are very effcent when a suffcent large no. of data set s avalable. The results show that the method s sutable for desgn a protectve scheme for transmsson lne base on artfcal neural networ. As the method s easy applcable and flexble, t can be used for modelng of any other transmsson lne. Fg 0 Regresson ft of the output vs. target for the faulty networ Fg 0 shows the snapshot of the traned A wth the confguraton and t s to be noted that the number of teraton requred for tranng process were 9. It can been seen that the mean square error n fault detecton acheved by the end of the tranng process value was shown n fg. and the number of valdaton chec 6 by the end of the tranng process. Fg Overvew of the A ( ) chosen for fault detecton V. COCLUSIO Ths paper presents a smulaton model usng by MATLAB (R009a) along wth Smpower system toolbox n Smuln for detecton of fault on transmsson lne. The methods employed mae use of the phase voltage and phase current as nput to the neural networ. In transmsson lne four types of fault namely sngle lne-ground, lne-lne, double lne-ground and three phase faults have been taen nto consderaton nto ths wor and only sngle lne ground fault has been show on ths paper wth ther proposed neural networ structures. ACKOWLDGMT I would le to express my sncere grattude and deep sense of ndebtedness toward my respected supervsor Dharmendra Kumar Sngh HOD (Dept. of ) of Dr. C. V. Raman Insttute of Scence & Technology who has always nspred me and extended hs full co-operaton n my wor. I would le to than all the Faculty of lectrcal & lectroncs Department. RFRCS [] sa Basher M Tayeb 03, eural networ approach to fault classfcaton for hgh speed protectve relayng Amercan Journal of engneerng research (AJR) volume-0, pp []S. Saha, M. Aldeen, C.P.Tan, Fault detecton n transmsson networs of power systems, Scnce Drect lectrcal Power and nergy Systems 33, pp , 0. [3]H.Sngh, M.S. Sachdev, T.S. Sdhu "Desgn, Implementaton and Testng of an Artfcal eural et- wor Based Fault Drecton Dscrmnator for protectng Transmsson Lnes," I Transactons on Power Delvery, Vol. 0, o., 995, pp [4]Abhjt A Dutta, A.K.adu & M.M.Rao 0 Intellgent control for locatng fault n transmsson lnes' Internatonal Journal of Instrumentaton, Control & Automaton (IJICA) ISS: volume, Issue-. [5] Thomas Dalsten, Brend Kulce 995, I Transacton on Power Delvery, volume0,issue-, pp [6]Rajveer Sngh 0, Fault detecton of electrc power transmsson lne by usng neural networ,volume-0, Issue-. [7]S..Svanandam, S.Sumath, S..Deepa Introducton of eural etwor Usng MATLAB 6.0 TMH Pbs. [8]Ibrahm Farahat,Dept. of lectrcal & computer Scence engneerng, Concorda Unversty, Canada. [9] I.J. agrath and D.P. Kothar, Power System engneerng", Tata McGraw-Hll, ew Delh, 994. [0]Venatesan R, Balamurugan B, A real-tme hardware fault detector usng an artfcal neural networ for dstance protecton,i Trans. on Power Delvery, vol. 6, no., pp. 75 8, 007. All Rghts Reserved 04 IJSTR 37