Deign, imulation and Implementation o a Full Bridge erie-parallel Reonant D-D onverter uing ANN controller Mohammad Jaari, Mohen Imanieh Department o Electrical Eng, Ilamic Azad Univerity-Faa branch Faa, Iran Mohammad_jaari@iauaa.ac.ir Zahra Malekjamhidi Department o Electrical Eng, Ilamic Azad Univerity-Marvdaht branch Marvdaht, Iran zmalekjamhidi@miau.ac.ir Abtract- A new method o control or high-voltage Full Bridge erie-parallel Reonant (FBPR D-D converter with capacitive output ilter, uing Artiicial Neural Network (ANN i propoed in thi paper. The output voltage regulation obtained via high witching requency and ot witching operation (Z and ZV technologie to decreae the loe and optimize the eiciency o converter. In the ollowing ection, a mall- ignal Model o FBPR converter on bae o irt harmonic analyi and the generalized averaging method i derived. Then the obtained model i ued to imulate the dynamic behavior o real converter uing Matlab otware. It wa alo ued to obtain ideal control ignal which are the deired ANN input and output and were aved a a training data et. The data et i then ued to train the ANN to mimic the behavior o the ideal controller. In act the ANN controller i trained according to the mall ignal model o converter and the ideal operating point. To compare the perormance o imulated and practical ANN controller, a prototype i deigned and implemented. The prototype i teted or tep change in both output load and reerence voltage at teady tate and under tranient condition. omparion between eperimental and imulation how a very good agreement and the reliability o ANN baed controller. Key Word: Full Bridge, erie-parallel Reonant onverter, Artiicial Neural Network, ANN, ZV, Z A. INTRUDUTION The deign and analyi o reonant converter i oten comple due to the large number o operating tate occurring within a pule period. In thi paper a Full Bridge erie-parallel Reonant (FBPR onverter uing an ANN controller i introduced and the dierent tep o deign, imulation and eperimental tet are dicued. The converter output power i controlled by duty-cycle variation while the witching requency hould be adjuted to enure that one bridge leg commutate at zero current (Z. The econd bridge leg operate under zero voltage witching condition (ZV and thi guarantee the ot-witching operation in the entire operating range []. To deign an appropriate controller, an accurate mall ignal model o converter i needed. Many controller are deigned by trial and error method and thi take much time to et the controller parameter []. In thi paper a generalized averaging method i ued to obtain the mall ignal model [], []. Thi method overcome the limitation o the traditional tate-pace averaging method becaue it doe not require that the waveorm have mall ripple magnitude. Thu, it i able to decribe arbitrary type o waveorm []. A thi model doen't need comple mathematical analyi, it impliie the controller deign or FBPR converter and minimize the deign time, epecially in the trial and error method. Dierent adaptive controller or the FBPR onverter i uggeted, uch a equential tate Machine [] [5], Gain cheduled controller [6][], Paivity Baed controller [7][8][], and Fuzzy ontroller. equential tate Machine i an abtract machine compoed o inite number o tate that determine tranition rom one tate to another. It provide the gate ignal or power witche according to the previou value o ome power circuit parameter [], [5]. Gain cheduling i a eed orward adaptation and it can be regarded a a mapping rom proce to controller parameter. The main advantage o gain cheduling i the at dynamic repone o the controller. The Paivity baed control i a very robut method but it dynamic repone i not a at a the repone o the gain cheduled controller becaue it depend on the peed o the etimate o the load [] [7]. In thi paper an ANN baed control trategy i employed to provide a ae and table repone or regulation o converter output voltage, due to any change in load, reerence voltage and input voltage. Figure how the tructure o the FBPR onverter. The block known a ontrol ircuit included a DP high peed microcontroller to provide the driving gate ignal according to the trained ANN command and change in everal input ignal rom power circuit.
Figure. tructure o purpoed FBPR converter In the Following ection, the teady tate analyi o FBPR onverter i reviewed in ection B, the mall ignal model i dicued in ection. ection D allocated to the ANN controller dicuion and imulation and eperimental reult are dicued in ection E and F. B. TEADY TATE ANALYI OF FBPR ONVERTER The FBPR onverter can operate in three commutation mode known a Natural, Forced and Mied mode []. In the natural mode tranitor operate with Zero urrent witching (Z in the turn o time. To reduce the turn o loe, at recovery diode hould be ued a the current pike take place during the turn o proce [][9]. In the orced mode all witche operate with Zero Voltage witching (ZV and turn on when their anti-parallel diode are in conduction mode and turn o with current []. In the mied mode, witche o one bridge leg (Q,Q work with the ZV during turning on and the witche o other leg (Q,Q operate in Z during turning o time. In thi mode the conduction loe are minimized and the converter eiciency increae. In our project the converter work in boundary between orced and mied commutation mode. In thi mode the witche Q and Q turn on and o with zero current and the witche Q and Q turn on with ZV according to operation above reonant requency []. The reonant current I L i almot inuoidal orm during the operation and o it pectrum contain only the irt harmonic component. However waveorm o V, I D and Vcp do not have inuoidal orm. The voltage G can be traner ratio o the reonant converter ( calculated a a unction o the output rectiier conduction angle (which i proportion to load variation and the normalized witching requency according to equation ( []. v K G( = o = n. ( vin π K In thi equation n i high requency tranormer ratio, K and K are deined a: AB tan ( P p [ (. o ω. p. Re K = p. ( o ( = 0.7 in ϕ K ( ] Where, ϕ i the conduction angle o output rectiier which change according to load variation, the witching requency and O denote i the erie reonant requency and i calculated according to equation (. / ( L π ( 0 = Figure preent the three dimenional graph o G ( a a unction o o variation ( ϕ. Vo / Vin 0 8 6.8.6. φ. 0.8.5 F,n and.5 load Figure. three dimenional graph o G( a a unction o and load variation ( ϕ O The reonant requency o the circuit ( o with the conduction angle ( change ϕ and load variation. Thi i becaue o the inluence P on the reonant requency. The converter behave a a erie reonant converter at lower requencie and a a parallel reonant converter at higher requencie []. When the reonant current low or only a mall part o the witching period through P and the load i increaed, the converter behave a a erie reonant converter. In thi cae the reonant requency i almot equal to the erie reonant requency ( 0. On the other hand, the converter behave a a parallel reonant converter in
low load and while current low almot the whole witching period through p [][0]. In the ollowing ection, the operation o the erieparallel reonant converter above reonance with variable requency and phae-hit control i eplained in detail. In thi paper, FBPR converter witche commutate in a boundary between mied and orced mode to combine the advantage o both commutation mode. Thi mode o operation wa decribed in []. Figure illutrate the waveorm o converter during a pule period rom t to t0 and it i obviou that reonant current i almot in inuoidal orm.. MALL IGNAL MODEL In order to ind the mall-ignal model or the FBPR converter with capacitive output ilter, the irt tep i to ind an equivalent circuit. In thi circuit the component in econdary ide o high requency tranormer reerred to the primary ide a hown in igure 5. Figure 5. The equivalent circuit o FBPR converter or mall ignal analyi The tate variable o the circuit conidering the undamental harmonic o IL and V are deined a [], []: Figure. Waveorm o converter during a period The teady tate trajectory or the FBPR converter operating at ull load i hown in igure. It i obviou that the hape o the trajectory i almot circular which mean that the tate variable o the converter circuit (IL, V are almot in inuoidal orm. The time hown in the diagram are reerred to igure. Figure. The teady tate trajectory or tate variable o converter In the net ection, the teady tate mall ignal model or FBPR converter i obtained and analyzed. I L j j 5 j6 = (5 V = (6 V P V O = = (7 (8 0 7 Where, and 5 are repreentative o coine component o waveorm and, and 6 or inuoidal part. The tate variable 5 and 6 are written a a unction o and a they are repreentative o parallel capacitor voltage which could not be conidered a a tate variable. 5 6 Where πω πω P [ δ γ] = (9 P [ δ γ] = (0 γ π ϕ in (ϕ δ in ( ϕ = ( = ( According to thee variable, the tate vector o circuit can be written a: [ ] T 5 = ( Where:
( Dπ d 5 Vinin = ω ( dt L L πl d 6 V in = ω ( on( Dπ (5 dt L L πl d = ω (6 dt d = ω (7 dt d dt π 7 7 =. ϕ In thee equation [ o( ] R. (8 ω (witching requency i u (irt control input variable and D (duty elected a cycle o witching pule a u (econd control input. On the other ide the output voltage, the amplitude o IL and V are elected a output variable y, y and y repectively. Thereore the relation between tate variable, control input and output can be written a ollow: u = ω (9 u = D (0 y = ( 7 y = ( y = ( To obtain a teady tate olution o the ytem and inding the mall ignal model, the derivative o Equation (-(8 hould be equal to zero. Furthermore, the equation providing the teady tate olution can be achieved uing ( ϕ, the teady tate value o ϕ according to below equation. ϕ =.tan π ω pro ( γ = π ϕ in(ϕ (5 δ in ( ϕ ω. p. V in K δ δ M = (8 p γ π ( πω. L. p = (6 = (7 K. in ( Dπ M = (9 K. M in( D. π = { o( Dπ } ( M M (0 ω. = ( ω. = ( [. δ. γ ] = ( 5 π. ω. p [. δ. γ ] = ( 6 π. ω. p R 7 = π [ o( ϕ ] (5 The mall ignal traner unction wa achieved ater linearization o model around the teady tate. The linearized model according to below equation can give G ( which i ratio o output voltage to duty cycle. = A B. u. (6 y =. D. u (7 Where A, B, and D are matrice o ytem parameter and, y and u are tate, output and input vector repectively. The ymbol mean mall change in repective parameter. G( V D y u O = = (8 To achieve a complete model o ytem, variation o witching requency can be modeled by a contant diturbance becaue o it intantaneou change. It automatically adjuted to enure Z operation o one bridge leg. The inalized model i hown in igure 6. It can be ued a an ideal model to produce ample data or training ANN controller in the net tage. Figure 6. Block diagram o ytem model D. ONTROLLER DEIGN TAGE In thi ection the deign procedure or Artiicial Neural Network (ANN baed controller will be decribed in brie. everal conduction mode took place during a complete witching period. Each tate ditinguihe by the tranitor and diode which
conduct during their repected time. In general there are normally our baic tate, but everal unepected or unwanted tate may alo take place which hould be conidered. In thi reearch, a multi-layer eed-orward artiicial neural network i employed to achieve realtime control. ince the output voltage i a nonlinear unction o duty cycle ( D and witching requency ( F, n, they are choen to be the output o the neural network controller. The input variable o ANN are, input voltage (Vin,output current(i0,output D voltage(vo, erie inductor current (IL and parallel capacitor voltage (Vcp a torage element. The converter output voltage hould be kept at a contant while the ZV and Z operation hould be guaranteed; otherwie, the duty cycle o control ignal hould be changed or both line and load regulation and the witching requency hould be changed or ZV operation. Thereore, the output o the ANN controller are the change in duty cycle and requency o driving ignal ( D, F, n. The overall tructure o deigned controller illutrated in igure 7. approimate any nonlinear unction. The nonlinear igmoid unction i choen a the activation unction [9]. = e ( (9 a Ater imulation o converter according to mall ignal model and training o ANN controller, enitivity baed neural network pruning approach i employed to determine an optimal neural network controller coniguration [9]. In thi approach, the contribution o each individual weight to the overall neural network perormance i indicated by a enitivity actor j w. ( The enitivity o a global error unction, with repect to each weight, can be deined a the ollowing [9]: = J ( w = 0 J ( w = J ( without w J ( with w In thee equation, = w (0 w i the weight o the neural network and w i the inal value o weight ater training. The equation (0 can be approimated by ( or the back-propagation algorithm [9]. n= η ( w N [ w ( n ] w w i ( Figure7. The overall tructure o deigned controller The ANN baed controller i trained according to an ideal imulated model o FBPR converter. The ideal model i imulated uing mall ignal equation o the ytem a decribed in ection (. At the irt tep, a imulated model uing MATLAB otware i developed to repreent the converter ytem and it i ued to evaluate the converter tate and output at any time. Then we apply change to the input controller parameter and earch by an oline iteration procedure, at each ampling point. Thi proce give u the eact value o the control variable needed to enure convergence o output ignal to the deired value at the net ampling point. The obtained ideal control ignal which are the deired ANN input and output are aved a a training data et. The data et i then ued to train the ANN to mimic the behavior o the ideal controller. In order to atiy thi requirement, a multilayer eed-orward ANN, wa elected to be trained. It i clear that a multi-layer eed-orward ANN can Where N i the number o training pattern or each ANN weight update, ηi the learning rate which i choen to be 0.6 and W i the weight update. The enitivity calculation were done on baed o Equation (. The weight are inigniicant and can be deleted i their enitivity actor wa maller than a deined threhold and alo a neuron can be removed when the enitivitie o all the weight related with thi neuron are below than the threhold [9]. Thi proce reduced the number o active weight and neuron and wa eective in development o ANN controller peed. A Matlab imulink model i developed on bae o mall ignal model o converter to train the neural network o-line. a three layer eed-orward neural network which ha one hidden layer with 0 neuron wa elected and the network weight are elected randomly with uniorm ditribution over the interval [-, ]. The total number o weight wa 8 at irt tage and thi reduced to the 7 Ater activation o pruning proce while till the ANN controller provide imilar perormance. The imulated FBPR converter and it load regulation abilitie are preented in igure 8 and 9. A illutrated in igure 9(a and 9(b the output current tep up and down by 0mA. It take about 5 to 6m or output voltage to ettle on it table value ater a %0.5 o rated value (00v over and underhoot. 5
Figure 8. imulated circuit o FBPR converter (A (A (B Figure 8. hange in output voltage and current or a tep change in load (A rom 60mA to 50mA and (B rom 50mA to 60mA. E. EXPERIMENTAL REULT In order to veriy the reult obtained theoretically, a 500W prototype wa built on bae o high peed Digital ignal Proceor (DP TM0F8 a i hown in igure 0. The nominal output voltage wa 00V and the maimum output current wa A. The ANN otware wa developed in language and the bit, high peed PU provide an acceptable eicient proceing peed in both load and line regulation. Figure 0. (B The implemented FBRP converter (A, and Wave orm o VAB or ull and light load (B Inductor are made o errite core and the capacitor are made o plain polyeter. Power MOFET IRF80 are ued a active witche. The at recovery chottky diode FR07 are ued a rectiier. Eperimental reult how that the ANN controller in cae o mall change in load or reerence voltage provide good ytem repone o ettling time, Overhoot and Rie time. The reult are hown in igure (A to (D. 6
(A (B rom igure (A and (B that a 0mA tep up or down change in load current caue a 00mv under and overhoot in output voltage which i about %0.5 o rated value (00v. It take about 6m or controller to ettle on it table value (00V. The igure ( and (D how the change in output current in cae o an increae or decreae about %0.5 o rated value (00V, in reerence voltage. Thee caue a 0mA over and underhoot in output current and a 000mv over and underhoot in output voltage which compenated ater about 6m. The negligible dierence between imulation and eperimental reult i due to o-line training o ANN controller on bae o ideal imulink model. To ine-tune the weight o ANN and reduce the dierence an online ine-tuning train can be ueul. Eperimental reult how good agreement with imulation which approve the reliability o mall ignal model and ANN controller. F. ONLUION The ANN controller propoed in thi paper can be a reliable alternative to claic controller or FBPR converter. In general the ANN controller provide good characteritic in term o overhoot, rie-time and ettling time. omparion between imulated and eperimental reult howed good agreement which approved reliability o propoed controller. G. AKNOWLEDGMENT The author would like to thank the Reearch center o Ilamic Azad Univerity-Faa Branch or the inancial upport o thi reearch project. ( (D Figure.. The regulation o converter output voltage or tep up and down change in load and reerence voltage. The output current and voltage ignal were quite noiy due to high requency witching element. It i obviou H. REFERENE []. G. Ivenky, A. Kat,. Ben-Yaakov, A Novel R Model o apacitive-loaded Parallel and erie-parallel Reonant D-D onverter. Proceeding o the 8th IEEE Power Electronic pecialit onerence, t. Loui, Miouri, UA, vol., pp. 958 96, 997. []. F.. avalcante and J.W. Kolar, Deign o a 5kW High Output Voltage erie-parallel Reonant D-D onverter, in Proceeding o the th IEEE Power Electronic pecialit onerence, Acapulco,Meico, 00, vol., pp. 807-8. []. G. Garcia oto, J. Gaye, G. W. Baptite, Variable ampling Time erial-reonant urrent onverter ontrol or a High-Voltage X-ray Tube Application. Proceeding o the 0th European Power Quality onerence, Nuremberg, Germany, 00, pp. 97-977. []. J. A. abate, F.. Y. Lee, O-Line Application o the Fied- Frequency lamped-mode erie Reonant onverter. IEEE Tranaction on Power Electronic, vol.6, no., pp. 9. 7, Jan. 99. [5]. H. Aigner, Method or Regulating and/or ontrolling a Welding urrent ource with a Reonance ircuit, United tate Patent 68988, February 005. [6]. K. J. Atrom, B. Wittenmark, Adaptive ontrol. econd Edition, Addion-Weley Publihing ompany, Inc, 995. [7].. ecati, A. Dell'Aquila, M. Lierre, et al., A Paivity-Baed Multilevel Active Rectiier with Adaptive ompenation or Traction Application, IEEE Tranaction on Indutry Application, vol.9, no. 5, [8]. R. Ortega, A. Loría, P. J. Nicklaon, H. ira-ramirez, Paivitybaed ontrol o Euler-Lagrange ytem: Mechanical, Electrical and Electromechanical Application. pringer-verlag, London, UK, 998. [9]. Xiao-Hua Yu, Weiming Li, Tauik'' Deign and implementation o a neural network controller or real-time adaptive voltage regulation 7