ANFIS based Prediction of Monthly Average Global Solar Radiation over Bhubaneswar (State of Odisha)

Transcription

1 ANFIS based Predcton of Monthly Average Global Solar Radaton over Bhubanesar (State of dsha) Sthtapragyan Mohanty Research Scholar, Department of Computer Scence & Engneerng, CET Bhubanesar Abstract -- The paper presents an adaptve neuro-fuzzy nference system (ANFIS) based modelng approach to predct the monthly global solar radaton (MGSR) n Bhubanesar. Comparsons of the predcted and measured value of monthly global solar radaton (GSR) on a horzontal surface are presented. The nput parameters of the model used n ths paper are sunshne duraton, temperature, humdty, clearness ndex and the Global solar radaton s taken as the output. An adaptve neuro-fuzzy nference system based modelng s used to predct the monthly Global solar radaton for Bhubanesar for fve years The solar radaton data for forty months are used for tranng the ANFIS and the data for telve months s used for testng. The purpose of the study s to compare the accuracy of ANFIS th the measured value Calculated by usng Angstrom s equaton and some ntellgent technques (Neural Netork & SVM) Index Terms-- Global Solar radaton, Sunshne duraton, ANFIS I. INTRDUCTIN Among all Reneable energy sources, solar radaton s consdered as the most mportant parameter n the desgn and evaluaton of solar energy devces. Many developng natons solar radaton measurements data are not easly avalable. Therefore t s mportant to elaborate methods to estmate the solar radaton on the bass of meteorologcal data. Global solar radaton (H G ) s the most mportant component of solar radaton snce t gves the total solar avalablty at a gven place. Global radaton s measured only at a fe locatons due to the hgh cost nvolved n the purchase of varous equpments and mantenance thereof. The amount of solar radaton potental n the partcular locaton s mportant for solar energy system desgn such as stand-alone PV and hybrd systems. Global solar radaton data s taken as the most mportant factor for szng of PV system. The total solar radaton receved at any ponts on earth of the PV panel n the form of drect and dffuse radaton. Dffuse solar radaton s not observed expermentally n any meteorologcal staton. For ths some clmatologcal parameters are needed to develop and estmate the global &dffuse solar radaton. To estmate the amount of solar 97 energy ncdent on a horzontal surface, many models ere developed to relate the global solar radaton (H G ) th varous parameters such as relatve humdty, sunshne duraton, temperature, lattude, longtude etc.many models have been proposed to predct the amount of solar radaton n some ctes usng varous meteorologcal / clmatologcally parameter [1-6]. Applcaton of Neuro -Fuzzy Technques for Solar Radaton W.A. Rahoma, U Al Rahama and A.H Hassan Journal of Computer Scence,7 (10): , 011 ANFIS neuro-fuzzy system as dscussed as because t combnes fuzzy logc and neural netork technques to gan more effcency. The structure of the TS fuzzy model [10] s dentfed usng a method hch permts to determne the optmal structure on automatc manner. Hargreaves et al. s Model [8, 9, and 15]. Hargreaves et al. ere the frst to propose a procedure to estmate the global solar radaton by usng the dfference beteen daly maxmum and daly mnmum ar temperature and extraterrestral radaton. A clear reve of ANN applcatons for reneable energy systems has been reported by Kalogrous [18, 19] and Mellt et al. [0] for photovoltac systems. M. Rzan, M. jaml and D. P. Kothar [1] used GNN, a modfed approach of artfcal neural netork (ANN), s proposed to estmate solar energy to overcome the problems of ANN such as a large number of neurons and layers requred for complex functon approxmaton. Benghanem & A.Mellt[] used four RBF models for predctng the DGSR usng meteorologcal data at AI-madnah(Saud Araba). In ths present ork e appled the Takag-Sugeno fuzzy systems for modellng the daly solar radaton data [6-7]. The Anfs model s used to estmate the global solar radaton at Bhubanesar th avalalable clmatc parameters of sunshne hour, temperature, humdty and compare the result th the measured value calculated by usng Angstrom equaton and neural netork.. METHDLGY:-.1. Data collecton: The meteorologcal parameter sunshne duraton measured by Bhubanesar from 000 to 004, ere appled for predctng monthly GSR usng dfferent Neuro-

2 Fuzzy technques. Here e have appled the data of 40 months for tranng and the data of 1 months for testng. 1985; Sugeno and Kang, 1988) fuzzy nterface system, a fuzzy model conssts of to rules... Anfs Model:- The man objectve of ths ork s to predct the Global solar radaton by usng attrbutes such as temperature, relatve humdty, sunshne duraton, clearness ndex by usng ANFIS model and compare th other models. Layer 4 Layer Layer 3 x y. The basc structure of a FIS conssts of three conceptual components: a rule base, hch contans a selecton of fuzzy rules; a database, hch defnes the membershp functons (MF) used n the fuzzy rules; and a reasonng mechansm, hch performs the nference procedure upon the rules to derve an output. The ANFIS uses a hybrdlearnng rule combnng back-propagaton, gradent-descent, and a least-squares algorthm to dentfy and optmze the Sugeno system s parameters. x y A 1 A B 1 B 1 1 N N 1 f 1 Layer 5 f As t s dffcult to mathematcally defne the relatonshp among dfferent clmatologcal parameters used for predcton of global solar radaton. ANFIS can be used to map nonlnear relatonshp for predcton of output (Jang 1991; 1993). ANFIS neuro-fuzzy system as consdered, as t combnes fuzzy logc and neural netork technques that are used to gan more effcency. A neural netork can learn from both the data and feedback thout understandng the pattern nvolved n the data. But, the fuzzy logc models are easy to compare the pattern because they use lngustc terms n the form of IF-THEN rules. A neural netork th ther learnng capabltes can be used to learn the fuzzy decson rules; thus creatng a hybrd ntellgent system. A fuzzy nference system conssts of three components. These are (a) rule base, contans a selecton of fuzzy rules. (b) data base, defnes the membershp functons used n the rules and, (c) reasonng mechansm, to carry out the nference procedure on the rules and gven facts. Ths combnaton merges the advantages of fuzzy system and a neural netork. Jang (1991) proposed a combnaton of a neural netork and fuzzy logc popularly knon as called an adaptve neuro-fuzzy nference system. ANFIS Archtecture:- A typcal adaptve netork shon n Fgure 1 s a netork structure consstng of a number of nodes connected through drectonal lnks. Each node s characterzed by a node functon th fxed or adjustable parameters. Learnng or tranng phase of a neural netork s a process to determne parameter values to suffcently ft the tranng data. The basc learnng rule method s the back propagaton method, hch seeks to mnmze some error, usually sum of squared dfferences beteen netork s outputs and desred outputs. Generally, the model performance s checked by the means of dstnct test data, and relatvely good fttng s expected n the testng phase. Consderng a frst order (Takag and Sugeno, 98 Fg 1-ANFIS Archtecture Rule 1: If x s A 1 and y s B 1 then f 1 =p 1 x+q 1 y+r 1 (1) Rule : If x s A and y s B then f =p x+q y+r () If f 1 and f are constants nstead of lnear equatons, e have zero order TSK fuzzy-model. Node functons n the same layer are of the same functon famly as descrbed belo. It s to be noted that j denotes the output of the th node n layer j. Layer 1: Each node n ths layer generates a membershp grade of a lngustc label. For nstance, the node functon of the th node mght be j 1 =µa (x)= b x-c 1+ a here x s the nput to the node I, and A s the lngustc label (small, large) assocated th ths node; and {a, b, c } s the parameter set that changes the shapes of the membershp functon. Parameters n ths layer are referred to as the Premse Parameters. Layer : Each node n ths layer calculates the frng strength of each rule va multplcaton: = =µa (x) µb (y), =1, (4) Layer 3: The th node of ths layer calculates the rato of the th rule s frng strength to the sum of all rule s frng strengths: = = 3 1+, =1, xy (3) (5)

3 For convenence outputs of ths layer ll be called normalzed frng strengths. Layer 4: Every node n ths layer s a squared node th a node functon 4 =f = (p +qy+r ) (6) here, W s the output of layer 3, and s the parameter set. Parameters n ths layer ll be referred as Consequent Parameters. Layer 5: The sngle crcle node computes the overall output as the summaton of all ncomng sgnals.e. 5 = verall output = n f = f Thus, an adaptve netork s presented n Fgure s functonally equvalent to a fuzzy nterface system. The basc learnng rule of ANFIS s the back propagaton gradent decent hch calculates error sgnals (defned as the dervatve of the squared error th respect to each nodes output) recursvely from the output layer backard to the nput nodes. Ths learnng rule s exactly the same as the beck-propagaton learnng rule used n the common feed-forard neural netorks by Jang (1993). From ANFIS archtecture (Fgure 1), t s observed that the gven values of the of premse parameters, the overall output can be expressed as a lnear combnaton of the consequent parameters. Based on ths observaton, a hybrd learnng rule s employed here, hch combnes a gradent decent and the least squares method to fnd a feasble of antecedent and consequent parameters. The detals of the hybrd rule are gven by Jang (1993) here t s also clamed to be sgnfcantly faster than the classcal back propagaton method. From the ANFIS archtecture shon n Fgure 1, e observe that hen the values of the premse parameters are fxed and the overall output can be expressed as a lnear combnaton. The output f can be rertten as: f= f + f = f 1+f = (x)p 1+(y)q 1+( 1)r 1+( x)p +( y)q +( )r (8) hch s lnear n the consequent parameters p 1, q 1, r 1, p, q, r.therefore, the hybrd learnng algorthm developed can be appled drectly. More specfcally, n the forard pass of the hybrd learnng algorthm, node outputs go forard untl layer 4 and the consequent parameters are dentfed by the least squares method. In the backard pass, the error sgnal propagates backard and the premse parameters are updated by gradent descent. (7) 99 Accordngly, the hybrd approach converges much faster snce t reduces the dmenson of the search space of the orgnal back-propagaton method. For ths netork created fxes the membershp functons and adapt only the consequent part; then ANFIS can be veed as a functonal-lnked netork here the enhanced representaton, hch take advantage of human knoledge and express more nsght. By fne-tunng the membershp functons, e actually make ths enhanced representaton. Predcton Model:- Fg -ANFIS predcton Model The data set s avalable from the A reneable energy resource laboratory,nasa.a complete data set of fve years of data ( ) are used for predcton of global solar radaton by usng dfferent Meteorologcal parameters. A total of sxty months ( ) datasets are used n ANFIS model. Forty months are consdered as tranng and tenty months are consdered under testng. Durng tranng, a fve layered ANFIS structure s constructed havng one nput, three hdden and one output. The Gaussan type of membershp functon (gaussmf) s used for nput and lnear type functon s used for output. The number of correct outputs s noted tll the error s mnmzed 3. RESULT AND DISCUSSIN:- To desgn and develop an ANFIS model n MATLAB 1.0 softare verson s used n ths study. The proposed ANFIS netork has adapted the tranng data groups to form best membershp functon so as to deduce the desred output for testng data th mnmum epochs.

4 The pattern of varaton of actual and predcted response s shon for tranng and testng dataset for proposed model Fgures 4 and 5 sho that actual (blue dot) and predcted (red dot) values are unformly dstrbuted respectvely for tranng and testng data Fgure3. ANFIS model structure for proposed ork The ANFIS archtecture for proposed model s shon n Fgure 3. Input membershp functon s descrbed th Gaussan membershp functon. Hybrd learnng algorthm s used and ANFIS model s run tll the error s mnmzed. Error s mnmzed n to epochs durng tranng. Then, testng of data s carred out. Fgure6. Surface plot for the proposed ork The surface plot shon n Fgure 6 ndcates that the total landscape of decson space s covered by the ANFIS model for proposed model. The resdual analyss s carred out for the predcted values of the model by calculatng the dfference of actual and predcted values for tranng and testng data. It s observed that the resduals are dstrbuted unformly along the center lne. The absolute percentage relatve error n tranng phase s and n testng phase Table -1 month Measured predcton usng nn predcton usng anfs predcton usng SVM Fgurer4.Tranng data for proposed ork jan feb mar apr may jun jul aug sep oct Fgure5.Dstrbuton of predcted and actual response durng proposed ork testng the Nov Dec

5 Estmated monthly global solar radaton n comparson Wth measured data for Bhubanesar from Usng dfferent model Fg 7.Measured & predcted values of dfferent model 4. CNCLUSINS: The proposed ANFIS model has successfully predcted the global solar radaton for varous domans and t becomes sutable for any desgn of Isolated solar energy converson applcaton. The ANFIS model shos better results n comparson th other models. The evaluaton results of solar radaton shos a sgnfcant mprovement n statstcal parameters and depcts better accuracy than other models. The comparatve results demonstrate the predctng capablty of ANFIS model and ts compatblty for any regon th varyng clmatc condtons. Ths predcton of solar radaton makes t sutable for nstallaton of a montorng staton for a remote place and t can be extended for the szng of standalone PV systems n future. REFERENCES [1]. Mellt A, Kalogrou SA. Artfcal ntellgence technques for photovoltac applcatons : a reve. prog Energy comb sc 008; 34: []. J. P. Duffe and W. A. Beckman, Solar Engneerng of Thermal Process, John Wley & Sons, Ne York, NY, USA, [3]. M. Iqbal, an Introducton to Solar Radaton, Academc Press, Toronto, Canada, [4]. Internatonal Energy Agency (IEA), Valdaton of models for estmatng solar radaton on horzontal surfaces, Atmosphere & Envronment servce, Canada, [5]. J. A. Prescott, Evaporaton from a ater surface n relaton to solar radaton, Phlosophcal Transactons of the Royal Socety, vol. 64, pp , [6]. Angstrom A. solar and terrestral radaton J R Meteorolog Soc 194; 50(10); [7]. Prescott JA. Evaporaton from a ater surface n relaton to solar radaton.trans R Soc Aug 1940; 64: [8]. Mellt A, Benghanem M, Salh H. An Adaptve Artfcal Neural Netork for Modellng and Smulaton of a Stand-Alone Photovoltac Poer System. In: Proceedngs of the 3th conference on systems, sgnals, and devces. IEEE 005; 4:57. [9]. Mellt A, Benghanem M. Modellng and smulaton of stand-alone poer system usng artfcal neural netork. Solar orld congress. ISES/ASES 005. [10]. R. Iqdour and A. Zeroual, A rule based fuzzy model for the predcton of solar radaton Revue des Energes Renouvelables Vol. 9 () [11]. Farzad Fathan BEPLS, Predctng Global Solar Radaton usng Genetc Algorthm,(GA) Vol. (6) May 013: [1]. E.. Falay, J.. Adeptan, and A.B. Rabu, Emprcal Models for the Correlaton of Global Solar Radaton th Meteorologcal Data for Iseyn, Ngera, Vol. 9.( ). Nov 008 [13]. V Svamadhav1, and R Samuel Selvaraj, Predcton of monthly mean daly global solar Radaton usng Artfcal Neural Netork. Earth Syst. Sc. 11. (6). Dec 01, pp [14]. J.. josu and L. K. Komolafe, Models for estmatng solar radaton avalablty n South Western Ngera, Solar Energy, vol. 16, pp , [15]. G. H.Hargreaves and Z. A. Saman, Reference crop evapotran spraton from temperature, Transactons of the ASAE, vol. 1, pp.96 99, [16]. Reddy K. Solar resource estmaton usng artfcal neural netorks and comparson th other correlaton models. Energy Converson and Management 003; 44(15):519e30. [17]. Jang Y. Computaton of monthly mean daly global solar radaton n Chna usng artfcal neural netorks and comparson th other emprcal models. Energy 009; 34(9):176e83. [18]. Kalogerrou SA. Artfcal neural netorks n reneable energy systems applcatons : a reve. Reneable and Sustanable Energy Reves 001; 5: 373e401. [19]. Kalogrou SA. Applcatons of artfcal neural netorks for energy systems. Appled Energy 000; 67:17e35. [0]. Mellt A, Kalogrou S. Artfcal ntellgence technques for photovoltac applcatons : a reve. Progress n Energy and Combuston Scence 008; 34(5):574e63. [1]. M.Rzan, M.jaml and D.P.Kothar, Generalzed Neural Netork Approach for Global solar energy estmaton n Inda, IEEE Transactons on Sustanable Energy, vol.3 (3), July 01. []. Benghanem M & Mellt. A, Radal bass functon netork-based predcton of Global solar radaton data: applcaton for szng of a stand-alone photovoltac system AI-Madnah, Saud Araba, Energy 010; 35: [3]. Dorvo ASS, Jervase, AI-laat A, Solar radaton estmaton usng artfcal neural netorks. APPL energy 00; 74: [4]. M. Jaml Ahmad and G.N. Tar, Study of Models for Predctng the Mean Hourly Global Radaton from Daly Summatons, pen Envronmental Scences, 008,, [5]. paul A.Lynn, Electrcty from sunlght: an ntroducton to photovoltacs.unted kngdom :john Wley & Sons, Ltd, 010. [6]. R. Iqdour and A. Zeroual, Modellng Daly Global Solar Radaton Usng Fuzzy Systems, Internatonal Conference on Modellng & Smulaton - ICMS 04, Valladold, Span, 004. [7]. R. Iqdour and A. Zeroual, Modellng Solar Data Usng the Takag- Sugeno Fuzzy Systems, Internatonal Conference on Modellng & Smulaton - MS 04, Lyon, France, 004.