Artificial Neural Network Maximum Power Point Tracker

download Artificial Neural Network Maximum Power Point Tracker

of 5

Transcript of Artificial Neural Network Maximum Power Point Tracker

  • 7/27/2019 Artificial Neural Network Maximum Power Point Tracker

    1/5

    T S I N G H U A S C IE N C E A N D T E C H N O L O G YI S S N 1 0 0 7 - 0 2 1 4 1 2 / 2 3 p p 2 0 4 - 2 0 8V o l u m e 1 0 , N u m b e r 2 , A p r i l 2 0 0 5

    Artif ic ial Neural Network Maximum Power Point Trackerfor Solar Electric Veh icleT h e o d o r e A m i s s a h O C R A N , C A O J u n y i

    C A O B i n g g a n g ( f ^B J ) , S U N X in gh ua ( ^ )

    R e s e a r c h a n d D e v e l o p m e n t C e n t e r f or E le c t r i c V e h i c l e , X i a n J i a o t o n g U n i v e r s i t y , X i a n 7 1 0 0 4 9 , C h i n a

    A b s t r a c t : This paper proposes an artificial neural network maximum power point tracker (MPP T) for solarelectric vehicles. The MPPT is based on a highly efficient boost converter with insulated gate bipolar transistor (IGBT) power switch. The reference voltage for MPPT is obtained by artificial neural network (ANN) withgradient descent momentum algorithm. The tracking algorithm changes the duty-cycle of the converter sothat the PV-module voltage equals the voltage corresponding to the MPPT at any given insolation, temperature,and load conditions. For fast response, the system is implemented using digital signal processor (DSP ).The overall system stability is improved by including a proportional-integral-derivative (PID) controller, whichis also used to match the reference and battery voltage levels. The controller, based on the information supplied by the ANN, generates the boost converter duty-cycle. The energy obtained is used to charge the lithium ion battery stack for the solar vehicle. The experimental and simulation results show that the proposedscheme is highly efficient.K e y w o r d s : artificial neural network; maximum pow er point tracker (MPPT); photovoltaic module; digital

    signal processor; solar electric vehicle

    I n t r o d u c t i o nPhotovoltaic (PV) generation is gaining increased importance as a renewable source of energy. The undesirable rapid changes of solar power, which usually occurin a running vehicle, arise as a result of shade frombuildings, large trees, and clouds in the sky. Conventional PV systems have difficulties in responding torapid variations due to shade. The main drawbacks ofPV systems are that the initial installation cost is considerably high and the energy conversion efficiency(from 12% to 29% ) is relatively low. Furtherm ore, inmost cases, PV systems require a power conditioner

    R ece i v ed : 2 0 0 4 -0 3 -2 4* * T o w h o m co r r e s p o n d en ce s h o u l d b e ad d re s s ed .

    E-mai l : cao jy@mai l s t .x j tu .edu .cn ;T e l : 8 6 -2 9 -8 7 5 7 0 2 5 2

    (dc-dc or dc-ac converter) for load interface. Therefore,the overall system cost could be reduced drastically byusing highly efficient power conditioners, such as themaximum power point tracker (MPPT), to extract andmaintain the peak power from the PV module evenwhen the above-mentioned unfavorable conditionsoccur.

    Various methods of maximum power tracking havebeen considered in PV power applications^ 1 12^. Amongthe hill climbing methods [ 1 _ 5 ], the perturb and observe(P&O) method tracks maximum power point (MPP) byrepeatedly increasing or decreasing the output voltageat MPP of the PV module. This method requires calculation of dP/dFto determine the MPP [ 1 '2 ' 4 ] . Though themethod is relatively simple to implement, it cannottrack the MPP when the irradiance changes rapidly.Also, the method oscillates around the MPP instead of

  • 7/27/2019 Artificial Neural Network Maximum Power Point Tracker

    2/5

    Theodore Amissah OCRAN et al: Artificial Neural Network Maximum Power Point Tracker 205

    d i r ec t ly t r ack in g it. The i n c r e m e n t a l c o n d u c t a n c et e c h n i q u e (ICT) is the m o s t a c c u r a t e [ 6 7 ] a m o n g theo t h e r m e t h o d s . T h i s m e t h o d g i v e s g o o d p e r f o r m a n c eu n d e r r a p i d l y c h a n g i n g c o n d i t i o n s . H o w e v e r , the c o m p l e x c a l c u l a t i o n of dl/dV and the c o m p l i c a t e da l g o r i t h m r e q u i r e use of a d ig i ta l s ig n a l p ro cesso r( D S P ) , w h i c h w i l l u s u a l l y i n c r e a s e the to ta l sy s temcoEhe MP P t r a c k i n g m e t h o d u s i n g the shor t circui tc u r r e n t of t h e PV m o d u l e u t i l iz e s the fact thatthe opera t in g cu r r en t at MPP of the PV m o d u l e is l in ear lyp r o p o r t i o n a l to the sh o r t c i r cu i t cu r r en t oft h ePV m o d u l e ^ . U n d e r r a p i d l y c h a n g i n g a t m o s p h e r i c c o n d i t i o n s ,t h i s m e t h o d has a f a s t r esp o n se sp e ed of t r a c k i n g theM P P ,but the co n t ro l c i r cu i t is c o m p l i c a t e d . The MPPt r a c k i n g m e t h o d u s i n g o p e n c i r c u i t v o l t a g e ofthe so larp a n e l [ 1 1 ] u t i l i zes the fact that the o p e r a t i n g v o l t a g e atM P P is a lm o s t l in ear ly p ro p o r t io n a l to o p e n c i r cu i tv o l t a g e at MPP of th e PVm o d u l e ( u s i n g 76 of o p e nc i r cu i t v o l tag e as the MPP v o l t a g e ) . T h i s m e t h o d isv e r y s i m p l e and co s t - e f f ec t iv e , but the r e f e r en ce v o l t a g e d o e s not c h a n g e b e t w e e n s a m p l i n g s . M P P T s u s i n gt h e f u z z y l o g i c [ 1 3 1 4 ] and the a r t i f ic ia l n eu ra l n e tw o rk( A N N ) [ 1 5 ] h av e b een r ep o r ted . Th ese s tu d ies sh o w th a ts u c h m o d e r n c o n t r o l a l g o r i t h m s arec a p a b l e of i m p r o v in g the t r a c k i n g p e r f o r m a n c e as c o m p a r e d tot h e c o n v e n t i o n a l m e t h o d s .

    In th i s p ap er , we p r o p o s e an ANNM P P T for so lare lec t r ic v eh ic le PV s y s t e m s . The t r ack in g a lg o r i th mc h a n g e s the d u ty r a t ioof the c o n v e r t e r so t h a tthe PVm o d u l e v o l t a g e e q u a l sthev o l t a g e c o r r e s p o n d i n g to theM P P at t h a t a t m o s p h e r i c c o n d i t i o n . T h i s a d j u s t m e n t iscar r ied out by u s i n g the b a c k - p r o p a g a t i o n ANN. Ther e f e r e n c e v o l t a g e to the MPP is o b t a i n e dby an offlinet r a i n e d ANN. The co n t ro l le r g e n era tes the b o o s tc o n v e r t e r d u t y - c y c l e .1 PVM odu le C haracterist icsTh e so la r a r r ay ch arac te r i s t ics s ig n i f ican t ly in f lu en cet h e d e s i g n of the c o n v e r t e rand the co n t ro l sy s te m, soth e PV ch arac te r i s t ics w i l l be b r ie f ly r ev iew ed h e re .Th e so la r a r r ay is a n o n l i n e a r d e v i c e and can be r e p r e s e n t e das a c u r r e n t s o u r c e m o d e l ,ass h o w n inF i g .1.

    Th e t r ad i t io n a l IV ch arac te r i s t ics of a so lar ar ray ,w h e n n e g l e c t i n gthein te rn a l sh u n t r es i s tan ce , are g iv enb ythef o l lo win g e q u a t io n :

    Fig.1 Equ ivalent c i rcui tof a solar cel l

    ex p AkT(1)

    w h e r e 7 o a n d V0 are the o u tp u t cu r r en t and o u tp u tv o l t a g e of th e so la r a r r ay ; / is theg en e ra ted cu r r en tu n d e r a g iv en in so la t io n ; 7 s a t is the r ev er se sa tu ra t io ncu r r en t ; q is the c h a r g e of an e l e c t r o n ; k is theB o l t z m a n n c o n s t a n t ; A is the id ea l i ty f ac to r for a P-N j u n c t i o n ; is the a r r a y t e m p e r a t u r e ; Rs and Rsharethe in t r in s ic se r iesand sh u n t r es i s tan ces ofthe so larar ray .

    Th e sa tu ra t io n cu r r en t of th e so la r a r r ay v ar ies w i t ht e m p e r a t u r e a c c o r d i n gto thef o l lo win g e q u a t io n :

    ex p qE GBkI =17 + ^ _ r - 2 5 ) l

    8 L so /v ; J 1 0 0

    (2)

    (3)w h e r e Tr is the r e f e r e n c e t e m p e r a t u r e ; Ior is thesa tu ra t io n cu r r en t at Tr; EG is the b a n d - g a p energ y of t h e s e m i c o n d u c t o r u s e d in the so la r a r r ay ; isa lso an ideali ty factor ; 7s c is the shor t circuit curre nta t 2 5 C ; Kj is the sh o r t - c i r cu i t cu r r en t tem p era tu reco ef f ic ien t and is thei n s o l a t i o n in m W / c m 2 .

    E q u a t i o n s ( l ) - ( 3 ) are u s e d in the d e v e l o p m e n t ofc o m p u t e r s i m u l a t i o n s for the so la r a r r ay . The M a t l a bp r o g r a m m i n g l a n g u a g e is u s e d . F i g u r e 2 s h o w thes i m u l a t e d c u r r e n t - v o l t a g e and p o w e r - v o l t a g e c u r v e sfo r the so lar ar ray at d i f f e r en t in so la t io n s and d if ferentt e m p e r a t u r e s . T h e s e c u r v e s s h o w t h a t the o u tp u t ch ar ac te r i s t ics of the so la r a r r ay are n o n l i n e a r and g r ea t lyaffected by the so la r r ad ia t io n , t em p era t u re , and lo adc o n d i t i o n . E a c h c u r v e has a m a x i m u m p o w e r p o i n tTW), w h i c his the o p t im a l o p e r a t i n g p o i n tfor the ef

    f icient use ofthe so lar a r ray .

    2 Artif icial Neu ral Netw orkA N N t e c h n o l o g y has b e e n s u c c e s s f u l l y a p p l i e d to

  • 7/27/2019 Artificial Neural Network Maximum Power Point Tracker

    3/5

    2 0 6 Tsinghua Science an d Technology, April 2 0 0 5 , 1 0 ( 2 ) : 2 0 4 - 2 0 8

    5 - 1 0 0 0 W / m

    0 20 40 60 80U

    (c) Power-vol tage curvesF i g . 2 C u r r e n t - v o l t a g e a n d p o w e r - v o l t a g e c u r v e s f o rt h e s o l a r a r r a y a t d i f f e r e n t i n s o l a t i o n s a n d d i f f e r e n tt e m p e r a t u r e s S i s t h e s o l a r r a d i a t i o n )

    solve very complex problems. Recently, its applicationin various fields is increasing rapidly [ 1 6 1 7 ] .

    The instantaneous sum of error squares or error energy at iteration is given by

    0implies Vre{-V A)2 ^ 0 , and then the connect ingweights of the network are adjusted in such a way thatthe array voltage VA is identically equal to the max imum power point voltage Vmp. At this stage the reference voltage Fre f becomes equal to the maximumpower point voltage Vmp.

    The configuration of the proposed three-layer feedforward neural network function approximator isshown in Fig. 3. The neural network is used to obtainthe voltage of the maximum power Vmp n) of the solar panel. The network has three layers: an input, ahidden, and an output layer. The numbers of nodes aretwo, four, and one in the input, the hidden, and outputlayers, respectively. The reference-cell open circuitvoltage Voc{n) and the time param eter n) aresupplied to the input layer of the neural network. Thesesignals are directly passed to the nodes in the next hidden layer. The node in the output layer provides theidentified maximum power point voltage V n) .Thenodes in the hidden layer get signals from the input

  • 7/27/2019 Artificial Neural Network Maximum Power Point Tracker

    4/5

    Theodore Amissah OCRA N eta l: Artificial Neural Network Maximu m Power Point Tracker 2 0 7layer and send their output to the node in the outputlayer. The sigmoid activation function is utilized in thelayers of the network. The training program calculatesthe connecting weights ^{1,1} with the bias b{\}for the input to hidden layer mapping, the connectingweightsW L{2 1}with bias b{2} for the hidden layer to

    Input

    output layer map ping. During the training, the connecting weights are modified recursively until the bestfit is achieved for the input-output patterns in the training data. The training of the net was accomplishedoff-line using Matlab.

    Hidden Output

    ^ o c ( )T{n

    * \ ^ , { 1 , 1 }

    W L { 2 ,1 }JJ

    b{2}> >

    2 4 1F i g . 3 F e e d - f o r w a r d n e u r a l n e t w o r k f u n c t i o n a p p r o x i m a t o r

    3 E x p e r i m e n t a l / S i m u l a t i o n R e s u l t sA PV array used for the collection of the experimentaldata is JDG-M-45 (Germany) type modules. The module has a maximum power output of 45 W and a 20-Vopen-circuit voltage at an irradiation of 1000 W /m 2and a 25 C temperature. The PV module specifications provided by the manufacturer in Table 1.

    T a b l e 1 E l e c t r i c a l s p e c i f i c a t i o n s f o r t h e P V m o d u l e( J D G - M - 4 5 )

    7 = 0 . 1 ; momentum factor a =0.9; number of trainingiterations=10 000; error goal=0.000 001. The convergence error for the training process is shown in Fig. 4.

    M a x i m u m p o w e r, Pm ( W ) 4 5S h o r t c i r c u i t c u r r e n t , 7 S C ( A ) 2 . 9 3O p e n c i r c u i t v o l t a g e , Voc ( V ) 2 0V o l t a g e a t m a x p o w e r p o i n t , Vmp ( V ) 1 7 . 1C u r r e n t a t m a x p o w e r p o i n t , 7 m p ( A ) 2 . 6 4S i z e ( m m X m m X m m ) 9 7 1 X 4 4 1 X 2 8M a s s ( k g ) 5

    The feed-forward back-propagation ANN, as shownin Fig. 3 was trained with values obtained from experimental data of the reference cell. Gradient descentalgorithm was used in training as it improves the performance of the ANN, reducing the total error bychanging the weights along its gradient. The trainingparameters are as follows: learning rate parameter

    0 2000 4000 6000 8000 10 000I terat ions

    F i g . 4 C o n v e r g e n c e e r r o r f o r t h e n e u r a l n e t w o r kt r a i n i n g p r o c e s s

    The parameters used for the training of the ANN andthe output values given by the ANN after the trainingcan be found in Table 2. Various sets of reference cellopen circuit voltage Voc and a time parameter (asshown in Table 2) are supplied as the input to the ANN.In order to validate the learning capability of the ANN,other sets of Voc different from the one in Table 2 werealso supplied to the ANN, which gave out values ofV mp as expected. The software Matlab was used in thetraining of the ANN.

    T a b l e 2 T h e r e s u l t s o f t h e t r a i n e d d a t a g i v e n b y t h e A N NH o u r o f t h e d a y

    9 : 0 0 9 : 3 0 1 0 : 0 0 1 0 : 3 0 1 1 : 0 0 1 1 : 3 0 1 2 : 0 0 1 2 : 3 0 1 3 : 0 0 1 3 : 3 0 1 4 : 0 0T i m e p a r a m e t e r - 1 - 0 . 9 - 0 . 8 - 0 . 7 - 0 . 6 - 0 . 5 - 0 . 4 - 0 . 3 - 0 . 2 - 0 . 1 0V oc m e a s u r e d 0 . 5 5 5 0 . 5 1 9 0 . 5 1 6 0 . 5 2 9 0 . 5 1 7 0 . 5 1 4 0 . 5 1 6 0 . 5 1 0 0 . 5 1 1 0 . 5 0 9 0 . 5 0 8V mp m e a s u r e d 0 . 4 2 2 0 . 3 9 4 0 . 3 9 2 0 . 4 0 2 0 . 3 9 3 0 . 3 9 1 0 . 3 9 2 0 . 3 8 8 0 . 3 8 8 0 . 3 8 7 0 . 3 8 6V mp g i v e n b y A N N 0 . 4 2 1 0 . 3 7 9 0 . 3 8 8 0 . 4 1 6 0 . 3 9 5 0 . 3 9 4 0 . 3 9 2 0 . 3 8 8 0 . 3 8 8 0 . 3 8 6 0 . 3 8 5

  • 7/27/2019 Artificial Neural Network Maximum Power Point Tracker

    5/5

    2 0 8 Tsinghua Science and Technology April 2 0 0 5 , 1 0 ( 2 ) : 2 0 4 - 2 0 8

    The weigh ts to the hidden layer 1 from input 1 are asfollows:

    2 .1706 -5 .3865 -3 .2326 -5 .1766{1,1} =1 [109.8309 -27.9 435 97.7250 -8.5640_

    The w eights to the output layer 2 are ^{ 2, 1} =[-0.2039 0.0065 0.0561 - 0.3 02 9]. The bias tolayer 1 is b{\} [-60.0887 12.996 -54 .47 53- 1 . 5 9 6 9 ] . The bias to layer 2 is b{2}= [ -0 .0397] .

    4 C o n c l u s i o n sAn artificial neural network MPPT for charging thebattery stack of a solar (hybrid) vehicle has been proposed in this paper. An off-line ANN, trained using aback-propagation with gradient descent momentum algorithm, is utilized for online estimation of referencevoltage for the feed-forward loop. Experimental data isused for the offline training of the ANN, and softwareMatlab is used in the training of the net. The precisionof the estimation has been verified by the graph of theconvergence error. The proposed method has severaladvantages over the conventional methods, particularlyin that there is no need for voltage and current sensors,and in that it avoids a complex calculation of power.The experimental and simulation results show that theproposed scheme is highly efficient.References[ 1 ] H u a C , L i n J , S h e n C . I m p l e m e n t a t i o n o f a D S P - c o n t r o l l e d

    p h o t o v o l t a i c s y s t e m w i t h p e a k p o w e r t r a c k i n g . IEEE Trans.Ind. Electron., 1 9 9 8 , 4 5 ( 1 ) : 9 9 - 1 0 7 .

    [ 2 ] K o u t r o u l i s E , K a l a i t z a k i s K , V o u l g a r i s C . D e v e l o p m e n to f a m i c r o c o n t r o l l e r - b a s e d p h o t o v o l t a i c m a x i m u m p o w e rp o i n t t r a c k i n g c o n t r o l s y s t e m . IEEE Trans. Power Electron., 2 0 0 1 , 1 6 ( 1 ) : 4 6 - 5 4 .

    [ 3 ] K u o Y C , L i a n g J , C h e n J F . N o v e l m a x i m u m p o w e rp o i n t t r a c k i n g c o n t r o l l e r fo r p h o t o v o l t a i c e n e r g y c o n v e r s i o n s y s t e m . IEEE Trans. Ind. Electron., 2 0 0 1 , 4 8 ( 3 ) : 5 9 4 -6 0 1 .

    [ 4 ] S u l l i v a n C R , P o w e r s J . A h i g h - e f f i c i e n c y m a x i m u mp o w e r p o i n t t r a c k e r f o r p h o t o v o l t a i c a r r a y s i n a s o l a r -p o w e r e d r a c e v e h i c l e . I n : P r o c . I E E E P o w e r E l e c t r o n . S p e c .Conf. S e a t t l e , W A , U S A , 1 9 9 3 : 5 7 4 - 5 8 0 .

    [ 5 ] S i m o e s G , F r a n c e s c h e t t i . A r i s e - m i c r o c o n t r o l l e rb a s e d p h o t o v o l t a i c s y s t e m f or i l lu m i n a t i o n a p p l i c a t i o n s . I n :P r o c . I E E E A p p l i e d P o w e r E l e c tr o n . Conf. N e w O r l e a n s ,

    L A , U S A , 2 0 0 0 : 1 1 5 - 1 1 5 6 .[ 6] T s e , H o , C h u n g S , H u i S Y . A n o v e l m a x i m u m

    p o w e r p o i n t t r a c k e r f o r P V p a n e l s u s i n g s w i t c h i n g f r e q u e n c y m o d u l a t i o n . IEEE Trans. Power Electron., 2 0 0 2 ,1 7 ( 6 ) : 9 8 0 - 9 8 9 .

    [ 7 ] B r a m b i l l a A , G a m b a r r a M , G a r u t t i A , R o n c h i F . N e w a p p r o a c h t o p h o t o v o l t a i c a r r a y s m a x i m u m p o w e r p o in t t r a c k i n g . I n : P r o c . I E E E P o w e r E l e c t r o n . S p e c . Conf. C h a r l e s t o n , S C , U S A , 1 9 9 9 : 6 3 2 - 6 3 7 .

    [ 8 ] M u t o h N , M a t u o T , O k a d a K , S a k a i M . P r e d i c t i o n - d a t a -b a s e d m a x i m u m p o w e r p o i n t t r a c k i n g m e t h o d f or p h o t o v o l t a i c p o w e r g e n e r a t i o n s y s t e m s . I n : P r o c . I E E E P o w e rE l e c t r o n . S p e c . Conf. C a i r n s , A u s t r i l i a , 2 0 0 2 : 1 4 8 9 - 1 4 9 4 .

    [ 9 ] N o g u c h i T , T o g a s h i S , N a k a m o t o R . S h o r t - c u r r e n t p u l s e -b a s e d m a x i m u m p o w e r p o i n t t r a c k i n g m e t h o d f o r m u l t i p l ep h o t o v o l t a i c a n d c o n v e r t e r m o d u l e s y s t e m . IEEE Trans.Ind. Electron., 2 0 0 2 , 4 9 ( 1 ) : 2 1 7 - 2 2 3 .

    [ 1 0 ] L e e D Y , N o h J , H y u n D S , C h o y I . A n i m p r o v e dM P P T c o n v e r t e r u s in g c u r r e n t c o m p e n s a t i o n m e t h o d f o rs m a l l s c a l e d P V - a p p l i c a t i o n s . I n : P r o c . I E E E A p p l i e dP o w e r E l e c t r o n . Conf. D a l a s , T X , U S A , 2 0 0 2 : 5 4 0 - 5 4 5 .

    [ 1 1 ] E n s l i n J R , W o l f S , S n y m a n D B , S w e i g e r s W . I n t e g r a t e d p h o t o v o l t a i c m a x i m u m p o w e r p o i n t t r a c k i n g c o n v e r t e r . IEEE Trans. Ind. Electron., 1 9 9 7 , 4 4 ( 6 ) : 7 6 9 - 7 7 3 .

    [ 1 2 ] C h e n Y , S m e d l e y K , V a c h e r F , B r o u w e r J. A n e w m a x i m u m p o w e r p o i n t t r a c k i n g c o n t r o l l e r f o r p h o t o v o l t a i cp o w e r g e n e r a t i o n . In : P r o c . I E E E A p p l i e d P o w e r E l e c t r o n .Conf. M i a m i B e a c h , F L , U S A , 2 0 0 3 : 5 6 - 6 2 .

    [ 1 3 ] N a f e h A , F a h m y F H , E l - Z a h a b A . E v a l u a t i o n o f ap r o p e r c o n t r o ll e r p e r f o r m a n c e f o r m a x i m u m p o w e r p o i n tt r a c k i n g o f a s t a n d - a l o n e P V s y s t e m . International Journalof Numerical Modelling, 2 0 0 2 , 1 5 ( 4 ) : 3 8 5 - 3 9 8 .

    [ 1 4 ] V e e r a c h a r y M , S e n j y u S , U e z a t o K . F e e d f o r w a r dm a x i m u m p o w e r p o i n t t r a c k i n g o f P V s y s t e m s u s i n g f u z z yc o n t r o l l e r . IEEE Trans. Aerosp. and Electron. Syst., 2 0 0 2 ,3 8 ( 3 ) : 9 6 9 - 9 8 1 .

    [ 1 5 ] H i y a m a T , K o u z u m a S , O r t m e y e r T . E v a l u a t i o n o f n e u r a ln e t w o r k b a s e d r e a l t i m e m a x i m u m p o w e r t r a c k in g c o n t r o l l e r f o r P V s y s t e m . IEEE Trans, on Energy Com., 1 9 9 5 ,1 0 ( 3 ) : 5 4 3 - 5 4 8 .

    [ 1 6 ] B o s e . A r t i f i c i a l n e u r a l n e t w o r k a p p l i c a t i o n s i n p o w e re l e c t r o n i c s . I n : P r o c . I E E E I n d . E l e c t r o n . Conf. D e n v e r ,C O , U S A , 2 0 0 1 : 1 6 3 1 - 1 6 3 8 .

    [ 1 7 ] H a y k i n S . N e u r a l N e t w o r k s A C o m p r e h e n s i v e F o u n d a t i o n , 2 n d E d i t i o n . N e w Y o r k : P r e n t i c e H a l l I n c . , 1 9 9 9 .