EVOLUTIONARY DESIGN OF A GENETIC BASED SELF ORGANIZING NEURAL NETWORK FOR WIND SPEED PREDICTION

Transcription

1 4 th Internatonal Conference on Experment/Proce/Sytem Modelng/Smulaton/Optmzaton 4 th IC-EpMO Athen, 6-9 July, IC-EpMO EVOLUTIONARY DESIGN OF A GENETIC BASED SELF ORGANIZING NEURAL NETWORK FOR WIND SPEED PREDICTION Stylano Sp. Pappa *, Georgo E. Chatzarak, Valo C. Moua, Nkolao Pournara 3 Department of Electrcal Engneerng Eduacator Apete School of Pedagogcal and Technologcal Educaton N. Heraklon, 4 Athen, Greece. School of Technologcal Applcaton, Technologcal Educatonal Inttute of Athen, Ag. Spyrdona St., Egaleo, Greece. 3 Ctzen In Deed, -4 Meogeon Ave., Athen 5 7, Greece. * e-mal: pappa@aegean.gr, kroval@otenet.gr Keyword: Wnd power, tme ere predcton, genetc-baed elf-organng network (GBSON) method, artfcal neural network, genetc algorthm, evolutonary algorthm Abtract. Wnd peed predcton condered a the mot crucal tak n the mplementaton of an alternatve but at the ame tme relable and autonomou electrc power ource. Accurate wnd peed forecatng method are a gnfcant tool n overcomng a varety of economc and techncal problem connected to wnd power producton. Th paper addree the problem of wnd peed forecatng by applyng a technque baed on the Genetc-Baed Self-Organng Network (GBSON) method. Real data were ued and real cae were teted baed on the meaurement of the wnd peed provded by Veta Hella. The wnd peed tme ere predcton reformulated to a ytem dentfcaton problem, where the nput the pat value of the tme ere and the output the future value of a tme ere. Th method ha been appled n the pat to varou tme ere predcton problem gvng atfactory reult. INTRODUCTION One of the mot mportant factor cloely related to both economc and ocal development nowaday energy. However mot of the electrcal energy produced baed on the fol fuel, whch on one hand are hghly effcent but on the other are reponble for the emon of greenhoue gae and ther reerve, may not lat for much longer. A a reult alternatve way have been already adopted for electrc power producton baed on renewable ource uch a the wnd. It common knowledge that wnd power generaton drectly depended on wnd peed, whch gnfcantly affected by varou factor uch a the type of the terran, the heght, any obtacle preent and many more. Th mean that f the relablty of the wnd power generaton not at an acceptable level, wnd power cannot be ued n order to contantly upply electrcal energy to the power ytem. [, ]. Artfcal Intellgence amongt a et of method that offer robut oluton and can be effcently mplemented to varou real lfe problem n dvergng feld uch a medcne, chemcal procee, computer network topology, economc, fuel management and many more [3, 4, 5, 6, 7, 8]. In th paper artfcal ntellgence and more pecfcally the Genetc-Baed Self-Organzng Network (GBSON) method [9] appled to the wnd peed etmaton and predcton problem, ung real data provded by Veta Hella, gvng atfactory reult. The ame technque ha been already appled and teted to another real problem concernng the predcton of the thundertorm day n Greece, producng very good reult [, ]. METHOD PRESENTATION The technque propoed a hybrd method that combne the Group Method of Data Handlng (GMDH) and Genetc Algorthm []. It man advantage t ablty to overcome the drawback of the orgnal GMDH algorthm, nce they ue local earch technque to obtan an optmal oluton [3, 4]. The propoed method ue polynomal neural network to repreent the model of the ytem to be dentfed.

2 Each layer of the polynomal neural network regarded a a eparate optmzaton problem. The nput to the frt layer of the network the ndependent varable of the data ample. The output of each layer the peak node obtaned by the ue of a mult-modal Genetc Algorthm [5]. The peak node elected to be the output of a layer are alo the nput for the next layer. The populaton member of the GA are network node repreented by an eght-feld bt trng. The two frt feld are ued to repreent the node from the prevou layer connected to the preent node. The other x feld are ued to repreent the coeffcent of a quadratc functon that determne the output of the node y: y bz + = () + cz + dzz ez z where z, z are the output of the connected node n the prevou layer. The ftne meaure of a node gven by calculatng t decrpton length. The decrpton length gve a trade off between the accuracy of the predcton and the complexty of the network. The equaton ued by Kargupta and Smth for calculatng the decrpton length : D n I =.5nlog D n +.5mlog n () where the mean-quare error, m the number of coeffcent n the model elected and n the number of obervaton ued to determne the mean-quare error. The mult-modal GA ued n GBSON ncorporate the ftne-harng cheme, where the hared ftne gven by: f (3) m f = f the orgnal ftne of the node and m a nche count defned by: where m = N d = h( ) (4) h d ( d ) = σ a f d < σ otherwe (5) N the populaton ze and d the Hammng dtance between the member of the populaton and. The nche radu σ determned by the equaton: σ l l = = l q (6) where l the trng length and q the number of node n the prevou network layer. New populaton are obtaned after applyng the genetc operator of tournament electon, ngle-pont croover and pont mutaton. A matng retrcton alo appled to the member to be croed. If a member to be croed, t mate elected uch that d < σ. If no uch mate can be found then elected randomly. The GBSON procedure contnue untl the GA converge to a layer wth a ngle node. 3 THE WIND SPEED SERIES A number of mulaton wa conducted baed on the hour average of daly wnd peed recorded by the Veta Hella from November up to February. The obtaned tme ere dd not follow any perodc pattern and t wa alo preentng rregular ampltude, makng t hard to predct (Fgure ). The tak of th work to generate a ngle-tep predcton baed on pat obervaton. The data were normalzed to take value from zero to one, before ung them a nput data to the polynomal neural network. The nput pattern wa agned a (x(t-), x(t-), x(t-3)) and the dered output wa:

3 x ( t) = f (( x( t ), x( t ), x( t 3)) (7) From the 75 avalable data pont, 53 pont (77 to 3) were ued for the valdaton of potental model. The mulaton were run wth a populaton ze of 8 for 5 generaton, wth tournament ze 6, probablty of croover.95 and probablty of mutaton.. GBSON reulted to a network wth three layer to model the wnd peed tme ere. The mot gnfcant term n the partal decrpton, y = a + bx + cx + dx x + ex + fx (8) of the model wa the term x and the le gnfcant term wa the contant term. The pat value of the wnd peed ere, (x(t-), x(t-), x(t-3)), contrbuted equally to obtan the fnal model. The reult of the predcton can be een n Fgure. The actual error of the predcton hown n Fgure 3. The percent quare error (PSE) over the whole data et 3.% and the root mean quare error (RMSE).495. The PSE over the valdaton data et 3.5%. The dfference of the PSE over the whole data et and the valdaton data et mall, and thu the model obtaned can be condered to perform wth approxmately the ame accuracy n data pont that have not been ued n any part of the modellng proce. 4 CONCLUSIONS The paper ha preented the ue of artfcal ntellgence and more pecfcally artfcal neural network, genetc algorthm and evolutonary algorthm n the oluton of the tme ere predcton problem. The tme ere predcton problem ha been formulated a a ytem dentfcaton problem, where the nput to the ytem wa the pat value of a tme ere, and t dered output the future value of a tme ere. A method ha been developed baed on the well known from the lterature Genetc-Baed Self-Organng Network (GBSON) method and ha been appled to real data producng acceptable reult. However further work need to be done on the fne tunng of the neural network parameter n order to reduce the PSE. Addtonally the ablty of the propoed method for on-lne predcton hould be teted a well. 5 FIGURES Raw Data Wnd Speed n m/ Hourly Data Pont Fgure. Hourly averaged wnd peed tme ere, // //.

4 Global Data,9,8,7 Wnd Speed n m/ec,6,5,4,3 Raw Predcted,, Hourly Data Pont Fgure.The actual (normalzed) wnd peed ere and the predcted wth the propoed method. Global Error,6,5,4,3,, -, -, -, Fgure 3.The actual error for each pont of the wnd peed tme ere predcted by the propoed method.

5 REFERENCES [] Le, M, Shyan, L., Chuanwen, J., Honglng, L. and Zhang, Y. (9), A revew on the forecatng of wnd peed and generated power, Renewable and Sutanable Energy Revew, Vol. 3, pp [] Marcukat, M., Katna, V., Kavalauka A. (8), Wnd power uage and predcton propect n Lthuana, Renewable and Sutanable Energy Revew, Vol., pp [3] Chen-Hen, W., Valad, C.A. (998), Applyng hybrd artfcal ntellgence technque n watewater treatment, Engneerng Applcaton of Artfcal Intellgence, Vol., Iue 6, pp [4] Patel, V.L, Shortlffe, E.H, Stefanell, M., Szolovt, P., Berthold, M.R., Bellazz, R., Abu-Hanna, A. (9), The comng of age of artfcal ntellgence n medcne, Artfcal Intellgence n Medcne, Vol. 46, Iue, pp [5] Samuel, P. (993), Applcaton of artfcal ntellgence technque to computer network topology degn, Engneerng Applcaton of Artfcal Intellgence, Vol.6, Iue 5, pp [6] Elmar, C.A., Bala, A., Allan, T.W., Koc, L. (4), Marne ltter predcton by artfcal ntellgence, Marne Polluton Bulletn, Vc.ol. 48, Iue 5-6, p.p [7] Cheng, M., Ta, H.C., Lu, C.L. (9), Artfcal ntellgence approache to acheve trategc control over proect cah flow, Automaton n Contructon, Vol. 8, Iue 4, pp [8] Ioná Maghal Santo de Olvera, Schrru, R. (), Swarm ntellgence of artfcal bee appled to In- Core Fuel Management Optmzaton, Annal of Nuclear Energy, Vol. 38, Iue 5, pp [9] Kargupta, H., and Smth, R.E. (99), Sytem dentfcaton wth evolvng polynomal network, Proceedng of the 4th Internatonal Conference on Genetc Algorthm, pp [] Pappa, S.Sp., Ekonomou, L. (6), Comparon of Artfcal Intellgence Method for Predctng the Tme Sere Problem, Proceedng of the 6th WSEAS Internatonal Conference on Smulaton, Modellng and Optmzaton, Lbon, Portugal, September -4, pp [] Ekonomou L., Pappa S.Sp. (6), Implementaton of Artfcal Intellgence n the Tme Sere Predcton Problem, WSEAS Tranacton on Informaton Scence and Applcaton, Vol. 3, No., pp [] [3] Yao, X., Lu, Y. (998), Makng ue of populaton nformaton n evolutonary artfcal neural network, IEEE Tranacton on Sytem, Man, and Cybernetc-Part B: Cybernetc, vol. 8, no. 3, pp [4] Quaglarella, D., Peraux, J., Polon, C., Wnter, G. (998), Generc algorthm and evoluton tratege n engneerng and computer cence, Wley. [5] Pctet, O.V., Dacorogna, M.M., Dave, R.D., Chopard, B., Schrru, R., Toman, M. (995) Genetc algorthm wth collectve harng for robut optmzaton n fnancal applcaton, Neural Network World, vol. 5, no. 4, pp