Application of Feed Forward Neural Network to Differential Protection of Turbogenerator

Transcription

1 16th NATIONAL POWER SYSTEMS CONFERENCE, 15th-17th DECEMBER, Appliction of Feed Forwrd Neurl Network to Differentil Protection of Turbogenertor Amrit Sinh Dept. of Electricl Engg., Ntionl Institute of Technology Ptn, Indi D. N. Vishwkrm Dept. of Electricl Engg., Institute of Technology Bnrs Hindu University Vrnsi, Indi Abstrct This pper discusses the ppliction of Multi- Lyer Feed Forwrd Neurl Network (MFNN) for the differentil protection of the turbogenertor bsed on pttern clssifiction. The cses of ll the possible internl fults in the sttor of the genertor with lp winding hve been simulted using Modified Winding function Approch. The simulted fult currents in the phses nd their prllel pths t the terminl nd the neutrl end hve been considered for trining nd testing of the proposed MFNN. Different networks hs been ccordingly trined nd tested to detect, identify nd clssify the internl fult in the sttor. From the test results it is cler tht the proposed networks re cpble of correctly identifying nd clssifying the fult signl. Keywords- Synchronous genertor; Turbogenertor; Differentil Protection; Pttern Clssifiction; Artificil Neurl Network; Feedforwrd Neurl Network. I. INTRODUCTION Lrge turbogenertors which re synchronous mchines nd form n importnt prt of the modern power system usully hve lp-connected sttor windings in order to increse the current cpcity of the mchine. The knowledge bout the internl fults in the sttor windings of lrge synchronous mchines is very essentil s the lrge fult currents my cuse severe dmge to the windings nd possibly to the shft nd coupling of the mchine. In order to design n pproprite numericl protection scheme for synchronous genertors to protect ginst the internl fults, it is importnt to get the fult dt. However, it is imprcticl nd very difficult to get ctul synchronous genertor fult dt. Therefore, detiled nlysis of vriety of situtions is done by simulting synchronous genertor under internl fult conditions i.e. turn-to-turn, turn-to-frme etc. which is more verstile nd cost effective. A digitl differentil protection scheme bsed on restrining nd sensitivity fctor hs been proposed in [1]. An incomplete differentil protection computer technique bsed on wvelet trnsform hs been proposed in [2] for genertor. An online digitl computer technique tht uses the field winding current to detect fult in the rmture winding hs been proposed in [3]. The direction of the negtive sequence power flow t the mchine terminls discrimintes the internl nd externl fult. In [4], generlized digitl technique hve been developed for detecting open circuit nd short circuit in the sttor windings of the synchronous genertor bsed on positive nd negtive sequence models. A neurl network bsed digitl differentil rely hs been implemented in [5], nd uses seprte modules for fult detection nd fult clssifiction. The presence of fundmentl nd/or second hrmonic component in the field current hs been used for differentiting the three genertor sttes (norml, internl fult nd externl fult). Artificil neurl network bsed fult dignosis schemes hve been presented in [6], to detect the genertor winding fult, clssify the type of fult, nd lso identify the fulted phses. A neuro-fuzzy bsed scheme hve been proposed in [7], tht discrimintes between the three operting sttes of the synchronous genertor nd opertes only in the cse of internl fults in the sttor winding. This pper presents n intelligent numericl differentil protection scheme for synchronous genertor bsed on feedforwrd neurl network. The synchronous mchine model bsed on Modified Winding Function Approch hs been used for simulting internl nd externl fults under different conditions. Using simulted numericl fult dt, vrious rchitectures of feed-forwrd neurl networks hs been trined nd tested nd the most suitble rchitecture of ANN which will give the correct informtion regrding occurrence of the internl fult in the turbogenertor hs been identified. After receiving the signl of occurrence of the fult, the proper tripping commnd cn be issued to the circuit breker to isolte the fulty turbogenertor. II. FAULT SIMULATION The synchronous mchine model bsed on Modified Winding Function Theory requires ccess to the sttor winding rrngement so it cnnot be generlized. The Modified Winding Function Approch hs been used to simulte different types of internl nd externl fults using mchine electricl prmeters. The clcultion of inductnce is derived directly from the originl wveforms using winding functions of the ctul mchine winding distribution; hence the spce hrmonics re tken into ccount. The mchine electricl prmeters hve been used to simplify the clcultion of the mchine inductnces [8]. A Modified Winding Function Theory s expressed in eqution (1) & (2) nd discussed by [9], is used in this pper to tke into ccount the effect of the non-uniform ir gp nd non-sinusoidlly distributed windings in cse of internl fults in lrge synchronous genertors.

2 16th NATIONAL POWER SYSTEMS CONFERENCE, 15th-17th DECEMBER, L rl g n n xy = 2 πμ ( φθ, ) x( φθ, ) y( φθ, ) 2πμ rl g (, φ θ) n (, φ θ) g (, φ θ) n (, φ θ) x g (, φθ) Where, μ is the permebility of the free spce, r is the verge rdius of the ir gp, l is the xil stck length of the mchine, g ( ϕ, θ) is the inverse ir gp length, nd n ( ϕ, θ ) nd ( ϕ, θ ) re the turns function of the x n y windings x nd y, respectively. Here, ϕ is the ngle long the inner surfce of the sttor, nd θ is the ngulr position of the rotor with respect to the sttor reference xis. Opertor f is defined s the men vlue of function f over [,2π] s follows: 1 = 2π 2π f (2) f ( ϕ) dϕ For 2p slient-pole synchronous mchine nlysis, the inverse ir gp length is usully pproximted s [1], g ϕ, θ ) = α α cos(2 p θ ) (3) ( 2 Where, the minimum ir-gp length is (α + α 2 ) -1 nd the mximum ir-gp length is (α α 2 ) -1. In the cse of turbogenertors (non-slient pole synchronous genertors), the ir gp is uniform nd therefore in (3) the term α = 2 [1] nd the inverse ir gp length is pproximted s. g ( φ, θ) = α (4) Substituting (4) into (1), mutul inductnce L xy of the sttor windings is given by L [ ] xy = K nn x y nx ny (5) K = 2πμrlα where The self nd mutul inductnces of non-slient pole synchronous mchine in bc phse vribles re given respectively s [11] L = L nd L b = -L b (6) There is very smll mount of flux round the ends of windings which does not cross the ir-gp. Hence it cn be pproximted s L b = L /2 (7) The inductnces in dq vribles cn be represented in bc phse vribles for round rotor synchronous mchines s L md = L mq = L + L b (8) y (1) The mchine geometricl constnts cn be represented by solving (1-13) for K in terms of mchine electricl prmeters 2Lmd For self inductnces K = 3( nn n n ) (9) x x x x Lmd For mutul inductnces K = 3( nn x y nx ny ) (1) Substituting (9) & (1) bck into (5) completes the expression to clculte the self nd mutul inductnces for rbitrry windings without the use of mchine geometricl prmeters. As the internl fults in the sttor winding do not ffect the winding distribution nd the electricl prmeters of the rotor windings. However, since the winding distribution of the fulted windings on the sttor hs chnged, the mutul inductnces between the rotor nd the sttor should be clculted. The mutul inductnce between n rbitrry sttor winding x nd nd the dmper windings kd, kq re given s [8] L fdx nx sin( p( ϕ θ )) = Lmd (11) Lkdx n sin pϕ 2 + n cos pϕ L kqx ( ) ( ) 2 ( p( ϕ θ )) nx cos = Lmq (12) n sin ( pϕ) 2 + n cos( pϕ) 2 The synchronous mchine modeling during n internl fult under vriety of conditions cn be done using direct phse quntities s in [12]. The elements of the inductnce mtrix cn be clculted for turbogenertors using (5, 9-12). The rotor winding prmeters re directly clculted from the mchine stndrd electricl prmeters. The turbogenertor of 2MW, 15.7KV, 5Hz hs been considered for the modeling, simultion nd designing of protection scheme nd its electricl prmeters re given in ppendix. The circuit representtion of the turbogenertor is s shown in fig.1. The different types of internl fult i.e. turn-toground, inter-turn, turn-to-turn-to-ground fults nd externl fult with different inception ngles hve been simulted using Modified Winding Function Approch. Fig.1 Circuit representtion of the turbogenertor

3 16th NATIONAL POWER SYSTEMS CONFERENCE, 15th-17th DECEMBER, Fig.2 Exmples of simulted fults for phse fult detection nd clssifiction. Fig. 3. Exmples of simulted fults for prllel pth fult detection nd clssifiction.

4 16th NATIONAL POWER SYSTEMS CONFERENCE, 15th-17th DECEMBER, The mchine model hs been simulted in the MATLAB/SIMULINK environment nd hs been solved using TR-BDF2 solver (n implicit Rung-Kutt formul with first stge tht is trpezoidl rule step nd second stge tht is bckwrd differentition formul of order two). Since the elements of inductnce mtrix depends on the position of the rotor θ, the mtrix L nd its inverse hve to be evluted t ech step in order to determine the current. Typicl exmples of different types of simulted fults with phse currents nd prllel pth currents re s shown in figure 2 nd 3. III. MULTILAYER FEEDFORWARD NETWORK The informtion in the multilyer feedforwrd network flows in the forwrd direction nd during norml processing; there is no feedbck from the outputs to the inputs. In this the neurons re connected in lyered structure nd neuron in given lyer receives inputs from the previous lyer nd sends their output to the next lyer. Fig.4 represents multilyer feedforwrd network. The number of input nd output neurons is decided by the problem itself, while the complexity of the problem governs the number of hidden lyers nd the number of neurons in ech hidden lyer. A number of rchitectures hve been trined nd tested using bck propgtion lgorithm. Using the neurl network toolbox of MATLAB, n intelligent protection scheme hs been designed to detect n internl fult, nd differentite it with stedy stte condition nd externl fult. The Multi-Lyer Feed-Forwrd Network hs been used s pttern clssifier to trin nd test the network. A set of 2 sets of dt for vrious cses including 2 stedy stte condition cses hs been used for trining, testing nd vlidting the different feed-forwrd neurl network rchitectures. 2% of the totl number of dt sets comprising vrious cses hs been used for testing s well s 2% for vlidting nd the remining 6% hs been used for trining, for creting network. The feed-forwrd network hs been trined using the grdient descent with momentum bck propgtion trining function nd grdient descent with momentum weight/bis lerning function. The lerning rte hs been chosen s.1 nd the momentum constnt s.9. The hyperbolic tngent trnsfer function hs been used for ll the hidden lyers nd the output lyer. The sigmoidl function hs been used becuse it helps in producing n rbitrry decision boundry with smooth curves nd edges. Input lyer Hidden Lyer Output Lyer Fig.4 A Multilyer Feedforwrd Network The trining ws done on P83, core2 duo-processor, 2.4GHz for 5, epochs for the different protection schemes. Concerning the ANN rchitecture, prmeters such s the number of hidden lyers nd the number of neurons in the hidden lyers were decided empiriclly. A. MFNN for Detection of Internl Fult The three line currents in per unit hve been smpled t 8Hz (i.e. 16 smples per phse per cycle) t both the terminl end nd the neutrl end. These smples hve been represented by 96 input neurons to the network being trined for detection of internl fult bsed on pttern recognition. The three output neurons represent the norml condition, internl fult nd externl fults respectively. This trining process involves experimenttion with vrious network configurtions of three lyer network. The feed-forwrd network with 18 to 144 neurons in the hidden lyer ws trined nd the performnce observed. The trining time vried from 25 to 51 minutes. On the bsis of performnce, the best rchitecture which hs minimum trining, testing nd vlidting error in combintion s tbulted in Tble 1, hs been selected for testing. B. MFNN for Phse Clssifiction of Internl Fult 96 input neurons represents the three line currents in per unit with smpling frequency of 8Hz i.e. 16smples/cycle/phse t the terminl nd the neutrl end for differentil protection scheme. The five output neurons represent the norml condition, internl fults in the three phses, nd externl fults respectively. A three lyer network with 2 to 132 neurons in the hidden lyer ws trined nd the performnce observed for the detection nd clssifiction of fults. The trining time vried from 27 to 46 minutes. On the bsis of performnce, the best rchitecture which hs minimum trining, testing nd vlidting error in combintion hs been selected for testing. The trining dt hs been tbulted in Tble 2. C. MFNN for Prllel Pth Clssifiction of Internl Fult The currents in two prllel pths of the three phses t both the ends of the sttor winding hve been represented by 192 input neurons with smpling frequency 8Hz i.e.16smples/cycle/phse/pth hve been used for detection nd clssifiction of internl fults in one or more prllel pths. The eight output neurons represent the norml condition, internl fults in the two prllel pths of ech of the three phses, nd externl fults respectively. A three lyer network with 36 to 252 neurons in the hidden lyer ws trined nd the performnce observed. The trining time vried from 35 to 97 minutes. The best rchitecture which hs minimum trining, testing nd vlidting error in combintion hs been selected for testing on the bsis of performnce. The trining dt hs been tbulted in Tble 3. The moving dt window hs been used for the finl testing of ll the networks with the selected rchitecture nd the results hve been tbulted in Tble 4 from the instnce of fult inception. The results show tht the fult cn be correctly identified nd clssified ccurtely.

5 16th NATIONAL POWER SYSTEMS CONFERENCE, 15th-17th DECEMBER, Tble 1: Neurons in hidden lyers nd errors for MFNN Fult Detection (Neurons in Input Lyer = 96, Neurons in Output Lyer = 3) Network Trining Architecture Testing Vlidtion Trining Time in mins. 96 x 18 x x 42 x x 6 x x 9 x x 18 x x 144 x Tble 2: Neurons in hidden lyers nd errors for MFNN Fult Detection & Clssifiction of phses (Neurons in Input Lyer = 96, Neurons in Output Lyer = 5) Network Trining Testing Vlidtion Architecture Trining Time in mins. 96 x 2 x x 36 x x 55 x x 84 x x 15 x x 12 x x 132 x Tble 3: Neurons in hidden lyers nd errors for MFNN Fult Detection & Clssifiction of Prllel pths (Neurons in Input Lyer = 192, Neurons in Output Lyer = 8) Network Architecture Trining Testing Vlidtion Trining Time in mins. 192 x 48 x x 84 x x 12 x x 168 x x 24 x x 228 x x 252 x IV. CONCLUSION This pper hs presented the ppliction of MFNN for the differentil protection of turbogenertors ginst internl fults. The MFNN hs been used s pttern clssifier for detection, identifiction nd clssifiction of internl fults in phses s well s their prllel pths. The simulted phse currents nd prllel pth currents t the terminl nd the neutrl end hve been used for trining nd testing of the different rchitectures of the network. Vrious rchitecture of the network hve been tested nd compred s regrds trining nd testing errors. Finlly suitble rchitecture i.e. hs been selected nd the network hs been tested on moving dt window s when the fult is identified. After identifiction the genertor cn be isolted. The test results indicte tht the proposed MFNN is cpble of correctly identifying ll types of fults within cycle. APPENDIX Turbogenertor prmeters: Rted Power of Genertor 235 MVA Rted Voltge 15.7 KV Rted Frequency 5 Hz No. of poles 2 Armture Resistnce =.28 pu Armture Lekge Rectnce =.15 pu Direct Axis Synchronous Rectnce = 2.12 pu Direct Axis Trnsient Rectnce =.29 pu Direct Axis Sub-Trnsient Rectnce =.2 pu Qudrture Axis Synchronous Rectnce = 2.12 pu Qudrture Axis Trnsient Rectnce =.7 pu Qudrture Axis Sub-Trnsient Rectnce =.28 pu Direct Axis Trnsient OC Time Constnt 7 s Direct Axis Sub-Trnsient OC Time Constnt.24 s Qudrture Axis Trnsient OC Time Constnt.95 s Qudrture Axis Sub-Trnsient OC Time Constnt.12 s REFERENCES [1] P. K. Dsh, O. P. Mlik, nd G. S. Hope, Digitl Differentil Protection of Generting Unit Scheme nd Rel Time Results, IEEE Trns. Power App. & Sys., vol. 96, pp , Mr/Apr [2] T. Nenling, A. Qin, Y. XingGen, C. Deshu, New Genertor Incomplete Differentil Protection bsed on Wvelet Trnsform, Electric Power Systems Reserch, vol. 69, pp , 24. [3] P. K. Dsh, O. P. Mlik, nd G. S. Hope, Fst Genertor Protection Aginst Internl Asymmetricl Fults, IEEE Trns. Power App. & Sys., vol. 96, pp , Sep/Oct [4] T. S. Sidhu, B.Sung, nd M. S. Schdev, A Digitl Technique for Sttor Winding Protection of Synchronous Genertors, Electric Power Systems Reserch, vol. 36, pp , [5] A. I. Meghed, nd O. P. Mlik, An rtificil neurl network bsed digitl differentil protection scheme for synchronous genertor sttor winding protection, IEEE Trns. Power Delivery, vol. 14, pp , Jn [6] H. A. Drwish, A. I. Tlb, nd T. A. Kwdy, Development nd implementtion of n ANN-bsed fult dignosis scheme for genertor winding protection, IEEE Trns. Power Delivery, vol. 16, pp , Apr. 21. [7] B. Bhlj, R. P. Mheshwri, S. Nem, nd H. K. Verm, Neurofuzzy-bsed Scheme for Sttor Winding Protection of Synchronous Genertor, Electric Power Comp. & Sys., vol. 37, pp , My 29. [8] X. Tu, Louis-A. Dessint, Nicols Fllti nd Bruno De Kelper, Modeling nd Rel-Time Simultion of Internl Fults in Synchronous Genertors with Prllel-Connected Windings IEEE Trnsctions on Industril Electronics, vol. 54, pp , June 27. [9] J. Fiz nd I. Tbtbei, Extension of Winding Function Theory for Uniform Air-Gp in Electric Mchinery, IEEE Trnsctions on Mgnetics, vol. 38, pp , Nov. 22. [1] P. C. Kruse, O. Wsynczuck, nd S. D. Sudhoff, Anlysis of Electric Mchinery nd Drive System, IEEE Press, New York, [11] P.Kundur, Power System Stbility nd Control, Electric Power Reserch Institute, McGrw Hill, Inc., [12] A. Sinh, D. N. Vishwkrm nd R. K. Srivstv, Modeling nd Rel Time Simultion of Internl Fults in Turbogenertors, Elec. Power Comp. & Sys., Vol. 37, Sept. 29, pp

6 16th NATIONAL POWER SYSTEMS CONFERENCE, 15th-17th DECEMBER, Tble 4: Test Result of MFNN for the selected rchitecture using moving dt window from the instnt of inception of fult Test Result of 96x9x3 network for fult in pth 1 t 73% from neutrl. Test Result of 96x9x3 network for externl line to ground fult. Test Result of 96x15x5 network for fult between pth c1 t 75% & c2 t 75% from neutrl. Test Result of 192x228x8 network for fult between pth 2 t 45% & b2 t 92% from neutrl. Smple No. Norml Int. Fult Ext. Fult Norml Int. Fult Ext. Fult Norml phse phse b O 4 phse c O 5 Ext. Fult Norml pth pth O 4 pth b O 5 pth b O 6 pth c O 7 pth c O 8 Ext. Fult