An Improved DTC of an Induction Motor Drive with Neural Network Controller

Transcription

1 Internatonal Journal of Mechancal & Mechatronc Engneerng IJMME-IJENS Vol:14 No:2 54 An Iproved DTC of an Inducton Motor Drve wth Neural Network Controller Wahyu Mulyo Utoo, Sy Y S, Zanal Ala Haron, Azuwen Ada Bohar, Nooradzane Muhaad Zn, Rolna Mat Arff, Waluyo Ad Swanto Unvert Tun Huen Onn Malaya Batu Pahat, Johor, Malaya wahyu@uth.edu.y, yy@hotal.co, zanalal@uth.edu.y, azuwenadabohar@yahoo.co, adzane@gal.co, rolnaatarff@yahoo.co, waluyo@uth.edu.y Abtract A pace vector odulaton baed drect torque Control trategy uggeted and an ntellgence controller degn baed on th trategy preented. A neural network controller propoed to replace the conventonal PID controller to prove the drve perforance nce the perforance of an electrc drve really depen on the qualty of a peed controller. The neural network controller wa traned and realze for a peed controller. The controller wa utlzed n the feed-back loop of the control yte. The control yte rendton a well a the onlne learnng technque of the neural network are decrbed n th paper. The coparon wth the conventonal PID drect torque controller reveal the effectvene of the propoed chee by proved the perforance of tranent repone preented. A ulaton odel repreentng the coplete neural network baed drect torque control chee of nducton otor drve developed and verfed ung S-Functon Sulnk block progra. Index Ter Drect Torque Control, Inducton Motor Drve, Neural Network Control. I. INTRODUCTION The eergence of the vector control granted the dynac perforance of the nducton otor (IM) wth a glut proved at the begnnng of 197[1]. Theoretcally, the vector control that baed on Fleng' law [2] proved the control perforance of IM vrtually cloe to the d.c. otor where the flux and torque are decoupled and henceforth could be controlled eparately. The vector control baed on the relatonhp vald for dynac tate, ntead of only the frequency and agntude takng nto conderaton, the ntantaneou poton of voltage and current are alo controlled. Thu, the VC work on the poton of the pace vector for both tranent and teady tate to furnh ther exact orentaton. Th enure the fat flux and torque to be dynacally and be part of the hgh-perforance control carry out n cloed-loop etho. [1], [3]. The vector control can be coplyng n varou etho but only everal bac chee that are offered n the arket. The Feld Orented Control (FOC), Drect Torque Control (DTC) and Drect Torque Control-Space vector Modulaton (DTC-SVPWM) are the ot welcoe control etho. The ue of DTC tratege ha becoe ore unveral and popular for nducton otor drve and ee ha a very rapd growth n the developent of t. The DTC provde a very accurate and wft torque repone excludng the coplex feld-orented block and nner current regulaton loop [4], n contrat to vector control. The DTC wa frt ntroduced by Takahah and ha found great ucce wth the noton to enhance the dynac and exactne of the torque and flux repone whle at the ae te ubde the IM paraeter dependency [5]. In conventonal SVPWM-DTC, the peed, flux and torque controller wa baed on PID, the ajor proble of PID baed DTC wa the proper tranent repone wth hgh overhoot, a well a the long ettlng te. Furtherore, for a hgh perforance DTC IM drve ung the PID peed controller n peranent ode frequently coe wth the conundru of an overhoot durng tart up repone and a poor load turbance rejecton [6]. To overcoe th proble, a neural network baed DTC wa propoed to replace the PI n peed controller of the DTC of IM to obtaned a better perforance of tranent repone by reduce the overhoot and ettlng te. The propoed controller evaluated by ulaton and the reult copared wth the conventonal PID-DTC. The developent of neural network for DTC ethod ha been condered due to ther varou advantage over conventonal one. The pleentaton of the off-lne learnng algorth could reduce the uperorty of the peed control IM drve ethod. Th due to the tranng data eploy fxed otor paraeter. Conequently, for a real te applcaton the perforance of the controller wll decreae, becaue the otor paraeter vary agant the teperature and agnetc aturaton [7]. Th paper propoe an proveent of the neural network control degn for peed controller of DTC nducton otor drve. The neural network controller odel developed baed on onlne learnng algorth ung Back propagaton chee. The controller degned to generate peed control gnal. In the followng ecton the bac chee of the DTC yte decrbed. Developent of the propoed neural network SVPWM-DTC wll be explaned n ecton III. Sulaton block odel and the tetng reult wll be preented n ecton IV. The lat ecton wll be a cuon and concluon IJMME-IJENS Aprl 214 IJENS

2 Internatonal Journal of Mechancal & Mechatronc Engneerng IJMME-IJENS Vol:14 No:2 55 II. DIRECT TORQUE CONTROL METHODS The SVPWM-DTC produce the tator reference voltage vector by adopt of both torque and flux error, then feed to the SVPWM algorth to odulate. The contant wtchng frequency gnal then exported to the three phae nverter [8] a llutrated n Fg.1. 1 Ψ q Ψ tan Ψ Then, the electroagnetc torque etated a: 3 p T Ψ e q Ψ q 2 2 Fg. 1. SVPWM-DTC nducton otor drve. The nducton odel n the tator-fxed d-q reference frae decrbed by [9]. V d R Ψ dt III. DTC-SVPWM NEURAL NETWORK SPEED CONTROLLER Inpred by the ucceful functon of the huan bran, the artfcal neural network (ANN) wa developed for olvng any large cale and coplex proble. Baed on ablty to proce oe nforaton and alo to analyze the nput and output ultaneouly, t ake ANN utable for dynac and nonlnear yte. The developent of the tructure and learnng algorth of the Neural Network Drect Torque Control (NN-DTC) explaned a follow [7]. Th paper propoed a DTC IM drve wth the neural network technque baed on SVPWM to reduce the overhoot and torque rpple. The NN control added to the peed controller to produce the torque reference. The block dagra of the propoed NN-DTC of nducton otor drve hown n Fg.2. V d Rrr Ψr jrψ dt r r Ψ L L r Ψ L L r r Fg. 2. The block dagra of propoed NN-DTC. The echancal equaton are gven a below. The nducton otor tator flux and torque are calculated n the flux and torque calculator a follow: Ψ d ( V R ) dt Ψq ( V q R q ) dt A. The Propoed NN-DTC Generally, bac concept of the plant requred n degn neural network. The nput nuber of the NN controller tructure can be deterned baed on the avalable nput gnal nuber of the yte. Further, the nuber of hdden layer and the total neuron depended on the coplexty of the yte and the requred tranng accuracy. To pleent peed controller of an nducton otor drve, a ultlayer perceptron neural network control developed In vew of the type of the tak to be perfored, the tructure of the propoed NN-DTC a hown n Fg.3. Ψ 2 2 Ψ Ψq IJMME-IJENS Aprl 214 IJENS

3 Internatonal Journal of Mechancal & Mechatronc Engneerng IJMME-IJENS Vol:14 No:2 56 Fg. 3. Dagra block of neural network peed control for DTC nducton otor drve. The controller cont of nput layer, hdden layer and output layer. Baed on nuber of the neuron n the layer, the NN-DTC defned a a network tructure. The frt neuron of the output layer ued a a torque reference gnal (a 2 1= f ). The connecton weght paraeter between j th and th neuron at th layer gven by w j, whle ba paraeter of th layer at th neuron gven by b. Tranfer functon of the network at th neuron n th layer defned by: S 1 n w a b j1 j 1 j The output functon of neuron at th layer gven by: ( a f n ) B. Learnng Schee of the Propoed NNDTC After the neural network archtecture odeled, the learnng odel deterned n next tage to update network paraeter. By th learnng capablty, t ake the ANN utable to be pleented for the yte wth otor paraeter whch are dffcult to defne and vary agant wth envronent. The tranng proce nze the error output of the network through an optzaton ethod. Generally, n learnng ode of the neural network controller a uffcent tranng data nput-output appng data of a plant requred. Snce the otor paraeter of the nducton otor drve vary wth teperature and agnetc aturaton, the onlne learnng Back propagaton algorth developed. Baed on frt order optzaton chee, updatng of the network paraeter are deterned. The perforance ndex u of quare error gven by: 1 2 F e 2 e t a where t target gnal and a output gnal on lat layer. The gradent decent of the perforance ndex agant to the connecton weght gven by: F w j F n n w j Where f actvaton functon of the neuron. In th degn the actvaton functon of the output layer unty and for the hdden layer a tangent hyperbolc functon gven by: 2 f ( n ) 1 2n 1 e Updatng of the connecton weght and ba paraeter are gven by: w b j F( k) ( k 1) wj w j F( k) ( k 1) b b where k aplng te, α learnng rate, and F perforance ndex functon of the network. The entvty paraeter of the network defned a: F n F a a n Gradent the tranfer functon agan to the connecton weght paraeter gven by: n w j a 1 Fro ubttuton equaton (18) and (2) nto (13) the updatng connecton paraeter gven by: 1 j 1 w ( k 1) w a IJMME-IJENS Aprl 214 IJENS

4 otor tator current Ia,Ib,Ic (A) otor tator current Ia,Ib,Ic (A) Speed ( rad/) Internatonal Journal of Mechancal & Mechatronc Engneerng IJMME-IJENS Vol:14 No:2 57 Wth the ae technque the updatng ba paraeter gven by: 1 b ( k 1) b copared. Wth the ae peed and load torque reference, the ulaton of both etho are run ultaneouly. The ulaton tart at the peed on 8 rad/ wth a contant load appled. The tart-up peed repone of both yte hown n Fg.5. Meanwhle, the tart-up otor tator current repone hown n Fg. 6 and Fg. 7. IV. SIMULATION AND RESULTS Sulaton wa carred out to nvetgate the perforance of the NN-DTC. In th ecton the dynac odel of a threephae nducton otor, pace vector PWM and neural network control odel have been developed. The ulaton developed ung Borland C++, and then ebedded a S- functon n Sulnk-Matlab. The paraeter for the otor are gven n Table 1 whle the ulaton block dagra of the propoed IM drve yte hown n Fg NN Controller PID Controller te () Fg. 5. Start-up peed repone coparon between conventonal PID-DTC and NN-DTC controller te () Fg. 6. The otor tator current (Ia, Ib, Ic) for NN-DTC controller. Fg. 4. The ulnk block dagra of the propoed NN-DTC. 25 TABLE I. MOTOR PARAMETERS Motor Paraeter Frequency, f Value 5 Hz Pole, p 4 Stator retance, R.5Ω Rotor retance, Rr.25 Ω Stator nductance, L.415H Rotor nductance, Lr.412H Mutual nductance, L.43H Cobned of nerta, J.1kg- 2 To verfy perforance of the propoed NNDTC, the ulaton reult for a conventonal DTC-SVPWM and the neural network DTC-SVPWM propoed controller are te () Fg. 7. The otor tator current (Ia, Ib, Ic) for PID-DTC controller. Fg. 5 llutrated the tart-up repone of the yte. Fro the fgure, t how that NN-controller have a tart-up repone proved fro the conventonal PID controller. It clearly IJMME-IJENS Aprl 214 IJENS

5 otor tator current Ia,Ib,Ic (A) otor tator current Ia,Ib,Ic (A) Speed ( rad/) Speed ( rad/) otor tator current Ia,Ib,Ic (A) Internatonal Journal of Mechancal & Mechatronc Engneerng IJMME-IJENS Vol:14 No:2 58 explan that the tart-up peed repone ha a great prove by reduce the overhoot fro 94 rad/ to no overhoot. Bede, t alo how a fat enhanceent of the ettlng te. In Fg. 6 the otor three phae tator current (Ia, Ib, Ic) are pleented by ung NN controller whle Fg. 7 are pleented by ung PID controller. The overhoot reducng fro 25A to 13A whle the ettlng te greatly proved fro.5 econd to.18 econd. The ulaton tetng carry on by vary the peed reference fro 8 to 11 rad/. The perforance of two yte oberved. The peed trajectory of the otor when peed vare at the te of 1.5 econd hown n Fg. 8. The tep-up otor tator current repone at the te of 1.5 econd are hown n Fg. 9 and Fg. 1. The tranent repone of the peed when the peed reference vared llutrated n Fg. 8. Wth the contant load appled, the tep up repone of the peed trajectory agan how the fat reduce n overhoot a well a the ettlng te. The overhoot reoved fro 126 rad/. Fg. 9 cloe the proveent of the overhoot and ettlng by pleented a NN controller when copared wth Fg. 1 where t pleented by ung PID controller. The overhoot reducng alot 5% that fro 7A to 35A. The NN controller acheve t teady tate condton at the te of 1.6 econd whle a PID controller take a longer te, whch 1.7 econd te () Fg. 1. The otor tator current (Ia, Ib, Ic) for PID-DTC controller. The ulaton tetng contnung by tep down the peed fro 11 to 9 rad/ at the te of 3 econd. The perforance of two yte oberved. The peed trajectory of the otor when load appled to the yte at the te of 3 hown n Fg.11. The otor tator current (Ia, Ib, Ic) of both yte are hown n Fg.12 and Fg NN Controller PID Controller NN Controller PID Controller te () te () Fg. 8. Step up peed repone cparon between conventonal PID SVPWM and NN-SVPWM controller when the peed reference vared fro 8 to 11 rad/ te () Fg. 9. The otor tator current (Ia, Ib, Ic) for NN-DTC controller. Fg. 11. Step down peed repone coparon between conventonal PID- DTC and NN-DTC controller when the peed reference vared fro 11 to 9 rad/ te () Fg. 12. The otor tator current (Ia, Ib, Ic) for NN-DTC controller IJMME-IJENS Aprl 214 IJENS

6 otor tator current Ia,Ib,Ic (A) Internatonal Journal of Mechancal & Mechatronc Engneerng IJMME-IJENS Vol:14 No: te () Fg. 13. The otor tator current (Ia, Ib, Ic) for PID-DTC controller. A llutrated n Fg. 11, the ettlng te decreaed wth the overhoot alo greatly reduce fro 83 rad/ to no overhoot by apply of the propoed neural technque, thu, prove the yte perforance at the ae te. It can be een that there alot no overhoot coe wth an proveent ettlng te wth the propoed peed controller. Referrng to the Fg. 12 and Fg. 13, an overhoot of the otor current durng tep-down reducng fro 6A to no overhoot whle the ettlng te taken alo booted up by the NN controller when copared wth the PID controller, whch fro 3.2 econd by PID controller to 3.8 econd by NN controller. V. CONCLUSION The neural network controller for SVPWM-DTC peed controller nducton otor drve yte ha been preented n th paper. The propoed ethod eploy a frt order onlne learnng back propagaton algorth to generate the torque reference. A conventonal PID controlled DTC IM drve and a NN controlled DTC IM drve have been teted and copared. The ulaton reult how that the PID-DTC anage to produce output pee that follow the gven dered reference peed. However, there an overhoot on the dynac repone before t reache teady tate condton. The proper yte perforance can be olved by the propoed NN-DTC. Propoed chee valdate the effectvene of the ethod by proved the yte perforance. The proper tranent repone wth hgh overhoot a well a the ettlng te proble of the conventonal PID-DTC can be olved. In addton, the controller doe not requre otor paraeter data. The reult how that the perforance of tranent repone proved by reduce the overhoot and ettlng te. REFERENCES [1] Z. Boulghaoul, A. Elbacha, E. Elwarrak, D. Youf, Cobned Vector Control and Drect Torque Control an experental revew and evaluaton, Internatonal Conference on Multeda Coputng and Syte ~ICMCS 211~, pp. 1-9, Aprl 211. [2] I. Takahah, Y. Ohor, Hgh-perforance drect torque control of an nducton otor, IEEE Tranacton on Indutry Applcaton, vol. 25, n. 2, pp , Mar/Apr [3] M. Perron, Hoang Le-Huy, Full Load Range Neural Network Effcency Optzaton of an Inducton Motor wth Vector Control ung Dcontnuou PWM, IEEE Internatonal Sypou on Indutral Electronc, pp , 26. [4] S.K. Ln, C.H. Fang, Sldng-ode drect torque control of an nducton otor, The 27th Annual Conference of the IEEE Indutral Electronc Socety, ~IECON 21~, vol.3, pp [5] I. Takahah, T. Noguch, A New Quck-Repone and Hgh-Effcency Control Strategy of an Inducton Motor, IEEE Tranacton on Indutry Applcaton, vol. IA-22, no.5, pp , Sept [6] A. Mloud, E.A. Alradad, A. Draou, A New Control Strategy of Drect Torque Fuzzy Control of a PWM Inverter Fed Inducton Motor Drve, IEEE Internatonal Sypou on Indutral Electronc, vol.3, pp , July 9-13, 26. [7] A.H.M. Yat, W.M. Utoo, Effcency Optzaton of Varable Speed Inducton Motor Drve Ung Onlne Backpropagaton, IEEE Internatonal Power and Energy Conference, ~ PECon 26~, pp , Nov , 26. [8] B.C. Babu, C. Poongotha, Hgh Perforance Drect Torque Controlled Inducton Motor Drve for Adjutable Speed Drve Applcaton, Frt Internatonal Conference on Eergng Tren n Engneerng and Technology, ~ ICETET '8~, pp , July 16-18, 28. [9] I. Choy,S.H. Kwon, J.Y. Cho., J.W. K, and K.B. K, On-Lne Effcency Optozaton Control of a Slp Angular Frequency Controlled Inducton Motor Drve Ung Neural Network, 13 annual Conference, ~IECON 1996~ Proceedng pp , 1996 ACKNOWLEDGMENT The author would lke to grattude Unvert Tun Huen Onn Malaya for any valuable upport durng conductng th reearch and n preparng th anucrpt IJMME-IJENS Aprl 214 IJENS