A NEW ACIVE POWER LINE CONDIIONER FOR COMPENSAION IN UNBALANCED/DISORED ELECRICAL POWER SYSEMS Jesús R. Vázquez, Patrco Salmerón, F. Jaer Alcántara, Jame Preto Electrcal Engneerng Department, Escuela Poltécnca Superor, Unersdad de Huela Huela, Span azquez@uhu.es; Patrco.Salmeron@dfae.uhu.es; FcoJaer.Alcantara@dfae.uhu.es; Jame.Preto@dfae.uhu.es Abstract A new control desgn of an Acte Power Lne Condtoner (APLC) for unbalanced/dstorted electrcal power system compensaton s presented. he APLC s bult by an dcac power conerter and a control crcut. In ths paper, the control crcut s a new desgn based n two types of artfcal neural networks (ANNs). An adapte lnear neuron (ADALINE) estmates the reference compensaton current and a multlayer feedforward network carres out a pulse wdth modulaton (PWM) control. he resultant compensaton system elmnates harmoncs, reacte power and unbalanced currents wth a quck dynamc response. Last, two dgtal smulatons to check the proposed control performance are presented. Keywords: Power Qualty, Harmoncs, Power factor correcton, Acte power flters, Adapte control, Neural network INRODUCION In the last years, the ndscrmnate use of nonlnear loads hae supposed an ncrease of oltage and current waeform dstorton n electrcal power systems. On the other hand, the utlzaton of larger snglephase consumers mply serous unbalance problems n threephase systems. hus, n general, threephase electrcal power systems need a set of equpments to correct the asymmetry and the dstorton that systematcally pollute the electrc power qualty. he APLC s a method, between other ones, based on power electronc that permts to correct the electrc power qualty defcences, [3]. he APLC s a compensaton system bult by two basc crcuts: a dcac power conerter and a control crcut. Nowadays, the power conerter of major utlzaton s the named oltagesource nerter based n a brdge confguraton wth s IGBs, fg.. wo capactors are connected n the dc sde wth a common dc busbar, that ensures the presence of the neutral wre wth a reduced number of electronc deces, [3]. he control crcut s bult by the reference current generator and the PWM control, manly, the ramp comparator or the hysteress band comparator. he ramp comparator method compares the error between actual and reference compensaton currents to a trangular waeform to generate the nerter frng pulses. he adantage s that the nerter swtchng s lmted to the frequency of the trangular waeform; howeer, there are phase and ampltude errors. In the hysteress band controller method, the currents wll stay n a band around the reference currents; ths scheme prodes ecellent dynamc performance but s not possble to know the swtchng frequency. In short, the sutable sequence of IGBs frng sgnals allows to synthesze a compensaton current that follows the reference waeform. he major research works of the last twenty years are related wth control crcut desgns. he target s to obtan relablty control algorthms of the reference current and a quck response procedure to get the control sgnal. he Artfcal Neural Networks (ANNs) hae been systematcally appled to electrcal engneerng, [46]. Nowadays, ths technque s consdered as a new tool to desgn APF control crcuts. he ANNs present two prncpal characterstcs. It s not necessary to establsh specfc nputoutput relatonshps but they are formulated through a learnng process or through an adapte algorthm. Moreoer, the parallel computng archtecture ncreases the system speed and relablty, [7]. In ths paper, a new APF control method based on Neural Networks wll be presented. Load oltages and currents are sensed, the control block calculates the power crcut control sgnals from the reference compensaton currents, and the power crcut njects the compensaton current to the power system. Net tems descrbe the topology of the Artfcal Neural Networks (ANNs) used n the APF control. In secton, a new method to control an acte power flter wth ANNs s presented, and sectons 3 and 4 wll complete the descrpton of ANNs blocks ncluded. In secton 5, t wll be shown how the electrcal system and ts compensaton can be smulated n MatlabSmulnk applcaton. he results of two practcal case wll be presented, the compensaton by shunt APF of a threephase unbalanced ac regulator and the compensaton of a controlled threephase conerter.
PRINCIPLES OF AN ACIVE POWER FILER WIH NEURAL NEWORKS A system wth a nonlnear threephase load wth oltage supply s consdered. A shunt Acte Power Flter s used to generate the compensaton current. he nonlnear load current L s the sum of the source current S and the compensaton current C. he target s to get a balanced supply current wthout harmonc and reacte components. he sutable compensaton current njected by the shunt APF corresponds to the load current nonacte component. he APF power crcut proposed s a threephase IGBs brdge nerter wth a splt capactor n dc sde, to compensate threephase, fourwres, unbalanced nonlnear loads, [3], fgure. he actual load currents are sensed, ther estmated nonacte components are the reference compensaton currents, whch are compared wth the actual compensaton currents. he dfferences are the nputs of hysteress comparator blocks, and ther output sgnals are used to turn on/off the nerter swtches. he compensaton current wll stay n a band around the reference sgnal. S L Nonlnear Load Source C APF Power Crcut o IGBs APF Control Crcut Fgure : A threephase fourwre system wth shunt Acte Power Flter. 3 3 L, L In ths paper, a shunt APF wth a new hysteress band control s used to compensate the nonlnear loads. he target s to control the compensaton currents to follow the reference ones. he swtchng strateges of the threephase nerter wll keep the currents nto the hysteress band. he control block generates the IGBs trgger sgnals. A basc scheme of hysteress band control s shown n fgure. Here, a feedforward ANN s proposed to work as the habtual nonlnear relay. Reference compensaton currents Hysteress comparator Fgure : Dagram of ANN Hysteress band control. Actual compensaton currents,, 3, 3 Fgure 3: Control reference and actual currents. he control s bult by two blocks. he frst one s deeloped wth adapte networks (Adalne neurons), whch allow to make onlne estmaton of control reference compensaton currents. he second one s a feedforward network. After a tranng process, t works onlne as comparator between reference waeforms and actual compensaton currents, fgure 4. L, actual L, actual C, Adapte Network Block Error C, actual Feedforward Network Block IGBs swtchng functons APF (PWM control) Fgure 4: Block dagram of Acte Power Flter control. he nputs of the frst block are load oltages ( L,actual ) and load currents ( L,actua ). hs block estmates the compensaton currents whch are gong to be used as reference n the control system, ( C,ref ). he nputs of the second block are the dfferences between actual and reference compensaton currents, sgnals Error n fgure 4. If the error alue, poste or negate, s greater than the hysteress band, the trgger sgnals change, and the compensaton current wll decrease or ncrease respectely, fgure 3. he actual compensaton currents are the result of controllng the swtchng logc of power crcut deces. he proposed control allows an ecellent flter dynamc response, and the compensaton currents can be adapted quckly to any change of load current. he results of two practcal cases wll be presented n secton 5.
3 COMPENSAION REFERENCE CURREN 3. Adapte neural network prncples A perodc waeform can be epanded by Fourer analyss as sum of cosne and sne frequency components he followng model of sgnal to be estmated s proposed: [ X n cos ( n t) Yn sn ( n t) ] f ( t) ω ω () n,..., N where X n e Y n are the ampltude of cosne and sne components of the ordern harmonc. In ectoral form: he weght adaptaton algorthm s a modfcaton of the WdrowHoff (WH) algorthm, [7],[9], whch mnmze the aerage square error between actual and estmated sgnals. It can be wrtten as follows: e( ( W ( k ) W( (4) ( ( Equaton (4) s the WH rule. he scalar product ( ( s the norm of the ector (. So, n each teraton, weghts are adjusted proportonally to the error and they follow the ( untary drecton. A modfcaton of WH rule can be wrtten as follows: where: f ( t) W ( t) () α e( y( W( k ) W( (5) ( y( [ X Y X Y ] W! ( t) cosω t sn ω t... cos Nω t sn Nω t he sgnals are sampled wth unform rate, t. So, the tme alues are dscrete, k t wth k,,... he dot product presented n () s carred out by an Adalne network, where W s the network weghts ector. After the ntal estmaton, an adapte algorthm updates the weghts. hus, the estmated sgnal conerges to the actual one. Fgure 5 shows the network topology and the weghts update algorthm. ( s the cosne/sne ector and f actual ( s the actual sgnal. he neurons, takng nto account ther weghts W(, carry out an estmaton f est (. he error e( s the dfference between the actual sgnal and ts estmaton. An algorthm allows to get the weghts to be used n the net teraton W(k), whch mnmzes that error. After ths terate process the estmated sgnals adapt to the actual sgnals. f (k t) (k t) N N and sn ωk t cos ωk t sn ωk t cos ωk t X ( Y ( X ( Y ( sn N ωk t cos N ωk t X N ( Y N ( Fgure 5: Adapte network topology. Xn (k), Yn (k) f est (k t) error (k t) Weghts update algorthm (3) In (5), y( s the sgn of (,.e., y(sgn((). As ( are snusodal sgnals, f sgnals sgn s consdered, the learnng rate for the weght correcton wll ncrease. he conergencesettlng tme decreases, though the conergence s less stable. he authors hae consdered an aerage between the sgnal and the sgnal sgn, []. hus, t reduces the conergence problems. y(.5 sgn( ( ).5 ( (6) Moreoer, a learnng parameter α s ntroduced to get a more stable conergence. he α parameter s modfed as shown n the followng equaton. α ce c e α (7) hus, α parameter, whch depends on the lnear error and ts derate, mproes the algorthm conergence. Both correctons nfluence the conergence n opposte way; ths commtment must be acheed to get stable and fast enough conergence. he ntal eoluton of estmated sgnals depends on the ntal weght choce. he eoluton from another change doesn t depend on that ntal choce. 3. Estmaton of Voltage and Current Waeforms As nonlnear loads are present n the power system, load oltages and load currents can be epressed as shown n equatons (8) and (9): L L [ Vn cos( n t) Vn sn ( nω t) ] n,...,n ω (8) [ I n cos ( n t) I n sn ( nω t) ] n,...,n ω (9) where ω s the fundamental pulsaton, V n and V n are cosne and sn coeffcents of harmonc components of
the load oltage and I n and I n are cosne and sn coeffcents of harmonc components of the load current. wo Adalne neurons estmate, per phase, the fundamental components of load oltage and current. Each acte current s estmated from those fundamental components. Wth fundamental frequency coeffcents of L and L estmated per phase, the load acte current can be calculated wthout computng any ntegraton. In fact, the target source acte current per phase,.e. s: act, L P V he result of () s: act act, L, L V V L, ( V ω t) V sn( ω t) ) he dfference between actual load currents and ther estmated fundamental acte components are the nonacte currents. hey are used as reference compensaton currents of the APF control crcut. L, L, V L, cos(, V comp, ref, L L act, L () (3) he source currents of the compensated system become balanced and snusodal. 4 CONROL OF COMPENSAION CURRENS In the preous sectons, the method to obtan the control reference currents wth Adalne networks was presented. In ths secton, the control crcut based on feedforward ANNs s deeloped. 4. Feedforward neural network prncples he Artfcal Neural Networks (ANNs) consst of a large number of strongly connected elements. he artfcal neurons represent a bologcal neuron abstracton carred out n a computer program. he artfcal neuron model s shown n fgure 6. I L, L, L,, I dt dt () () he nput data (), (), (3),, (n) flow through the synapses weghts and they are accumulated n the node represented as a crcle. he weghts amplfy or attenuate the nputs sgnals before ther addton. Once added, the data flow to the output through a transfer functon f, that may be the threshold one, the sgn one, the lnear threshold one or the pure lnear one. Otherwse, t may be a contnuous nonlnear functon such as the sgmod one, the nerse tangent one, the hyperbolc one or the gaussan one. he neurons are connected conformng dfferent layers. he most commonly adopted arqutecture s feedforward, as shown n fgure 7. () () (3) (n) INPU DAA () () (3) (n) Fgure 6: Artfcal neuron model. INPUS W j ( W j * (j) ) b b HIDDEN LAYERS OUPU LAYER Fgure 7: Feedforward Neural Network arqutecture. he neural arqutecture conssts of three layers: the nput one, the hdden one and the output one. In the fgure, the crcles represent neurons. hus, the feedforward arqutecture computes the nput data n parallel way, faster than the computer sequental algorthms. hs network can be traned to supply an output target when the correspondng nputs are appled. he most commonly used method s the backpropagaton tranng algorthm. he ntal weghts are random. he ntal output pattern s compared wth the current output, and the weghts are adjusted by the algorthm untl the error becomes small enough. he tranng process s carred out by a program that uses a large number of nput/target data, whch can be obtaned from smulatons or epermental results. f RANSFER FUNCION NEURON OUPU
4. Neural PWM Control A feedforward Neural Network works as hysteress comparator n the PWM control, fgure 8. hs network s desgned wth two nputs and two layers, the hdden wth 4 neurons and the output layer wth neuron. he actaton functons are logsgmod n the hdden layer and lnear n the output layer. he tranng algorthm used s backpropagaton. he comparator outputs depend on the nputs and ther eoluton. he chosen confguraton has two nputs, the error sgnal n t and ts alue n the preous tme t. he network topology s shown n fgure 8. c,actual c,ref Delay error (t) error (t) 4 o IGBs A threephase snusodal oltage supply, a nonlnear load block and the power crcut of an acte power flter were deeloped n the same software, computeraded by the Power System Blockset, to get load oltages and currents of the electrcal system. In fgure 9, the Smulnk dagram of the compensated electrcal system s shown. he crcut wll be smulated wth two load block, a model of a threephase conerter and a model of a threephase acregulator. he control block s presented n fgure. he adapte network block estmates load oltages and currents. he dfference between actual currents and ther estmated acte components are the nonacte current used as compensaton reference currents. On the other hand, the feedforward block works onlne as hysteress comparator, ts nputs are reference and actual compensaton currents, and ts outputs are the trgger sgnals of the power crcut IGBs. Load oltage Fgure 8: opology of feedforward Network. o f the network weghts, t s necessary to compare the network outputs wth the outputs of a real electrcal system. 5 RESULS OF PRACICAL CASES 5. Smulaton Models o check the proposed desgn, the electrcal system, control block ncluded, was smulated n Matlab Smulnk software. Load current Adapte ANN block Estmaton of Acte current 3 Compensaton current VL error IL error Ic ref Fgure : Smulnk control block. Feedfor. ANN block Hysteress comparator o IGBs 3 Fgure 9: Smulnk dagram of electrcal system.
o tran the feedforward network, t s necessary to know real system nput and output sgnals. So, the electrcal system was smulated n Smulnk usng a relay block as comparator to obtan nputs and targetoutput sgnals. o tran the feedforward network, the Matlab Neural Networks toolbo was used. A tranng program usng matlab routnes was deeloped. Intff ntalses Network weghts, ranlm carres out the tranng process to f the fnal weghts usng pattern sgnals. After the tranng process, Smuff allows the network to work onlne as comparator. he error of ths comparator block was specfed at, % as mamum goal. he error sgnal durng the tranng process s shown n fgure. error, % 4 6 8 teraton Fgure : Error of tranng process. 5. Man results Case. he proposed desgn was appled n a electrc system wth a threephase balanced and snusodal oltage supply and a controlled threephase rectfer. Net, t s shown the eoluton of seeral waeforms per phase when the load s eperenced a 5% change. Fgure a presents load oltage. Actual and estmated load current are shown n fgures b and c. a) 5 5.5..5..5.3 3 he estmaton of equalent conductance, fgure 3a, shows the transent process eoluton. he compensaton current, fgure 3b, s the result of hysteress band control used. Fgure 3c presents the source current of the compensated system. he sgnal eoluton depends on the ntal weghts of the adapte networks. hose weghts hae been chosen null n ths eample. he dynamc response of power flter s one and a half perod appromately. a).3....5..5..5.3.5..5..5.3.5..5..5.3 me (s) Fgure 3: Case : Controlled threephase conerter. a) Estmated equalent conductance; Compensaton current; Source current Case. On the other hand, a second practcal case was smulated, an unbalanced threephase acregulator. As aboe, fgures 4 and 5 show the man results of ths case. Besdes, the neutral current s null n the compensated system, fgure 5d. a) 5 5.5..5..5.3 5 5.5..5..5.3.5..5..5.3 5 5.5..5..5.3 me (s) Fgure : Case : Controlled threephase conerter. a) Load oltage; Load current; Estmated load current.5..5..5.3 me (s) Fgure 4: Case : hreephase acregulator. a) Load oltage; Load current; Estmated load current
a) d).8.6.4..5..5..5.3 5 5.5..5..5.3 5 5 5.5..5..5.3 5.5..5..5.3 Fgure 5: Case. hreephase acregulator. a) Estmated equalent conductance; Compensaton current; Source current; d) Neutral source current 6 CONCLUSIONS me (s) A new control method of an APLC has been presented. he control method s desgned wth two ANNs blocks. he frst one ncludes Adalne networks, whch estmate load oltage and current harmonc components. A new method to get fundamental acte currents and reference compensaton currents has been presented. In the hysteress band control used, the usual comparators hae been substtuted by Feedforward networks traned by the backpropagaton algorthm. he use of neural technologes presents adantages wth regard to classcal solutons, t ncreases the control speed and relablty. he results of two practcal cases n MatlabSmulnk hae been presented. he control desgn proposed allowed to get an ecellent flter dynamc response to load changes. Nowadays, the authors are deelopng a practcal mplementaton wth dspace software. A controllerboard wth a realtme processor allow to carryout the control operaton. REFERENCES [] H. L. Jou, J. C. Wu, H. Y. Chu, New snglephase acte power flter, IEE Proc. Electr. Power Applcatons, ol. 4, no.3, May 994. [] B. Sngh, K. AlHaddad, A. Chandra, Acte Power Flter for Harmonc and Reacte Power Compensaton n hreephase, FourWre Systems Supplyng Nonlnear Loads, EEP Vol. 8, no., March/Aprl 998. [3] Maurco Aredes, Jürgen Häfner, and Klemens Heumann, hreephase FourWre Shunt Acte Flter Control Strateges, IEEE ransactons on Power Electroncs, ol., no., March 997. [4] BorRen Ln, Rchard G. Hoft, Power Electroncs Inerter control wth neural networks, IEEE echnology Update Seres, ol. Neural Networks Applcatons, pp. 7, 996. [5] S. Fukuda and H. Kamya, Current Control of Acte Power Flters asssted by Adapte Algorthm, Power Electroncs and Varable Speed Dres, Septembre, Conference Publcaton No. 475, pp. 37, IEE. [6] YaowMng Chen and Robert M. O Connell, Acte Power Lne Condtoner wth a Neural Network Control, IEEE ransactons on Industry Applcatons, Vol. 33, No. 4, pp. 336, July/Agost 997. [7] P. K. Dash, S. K. Panda, Baburam Mshra, and D. P. Swan, Fast Estmaton of Voltage and Current Phasors n Power Networks Usng an Adapte Neural Network, IEEE ransactons on Power Systems, ol., no 4, Noember 997. [8] Qun Wang, Nng Wu and Zhaoan Wang, A Neuron Adapte Detectng Approach of Harmonc Current for APF and Its Realzaton of Analog Crcut. IEEE ransactons on Instrumentaton and Measurement, Vol. 5, No., pp. 7784, February. [9] P. K. Dash, S. K. Panda, A. C. Lew, B. Mshra, R. K. Jena, A New Approach to Montorng Electrc Power Qualty, Electrc Power System Research, 46, 998. []J.R. Vázquez, P. R. Salmerón, Neural Network Applcaton to Control an Acte Power Flter, Proceedngs of 9 th Power Electroncs and Applcatons Conference. Graz, Austra, August.