Short-Term Wind Power Forecasting Using Artificial Neural Networks for Resource Scheduling in Microgrids

Transcription

1 Sort-Term Wind ower Forecasting Using Artificial eural etworks for Resource Sceduling in Microgrids Abinet Tesfaye Scool of Electrical and Electronic Engineering, ort Cina Electric ower University, and Goldwind and Etecwin Electric Co., Ltd Beijing, Cina J. H. Zang Scool of Electrical and Electronic Engineering, ort Cina Electric ower University Beijing, Cina D. H. Zeng Goldwind and Etecwin Electric Co., Ltd Beijing, Cina Dereje Siferaw Department of Electrical and Computer Engineering, Adama Science and Tecnology University Adama, Etiopia Abstract: Use of wind power as one of renewable resources of energy as been growing quickly all over te world. Wind power generation is significantly vacillating due to te wind speed alteration. Terefore, assessment of te output power of tis type of generators is always associated wit some uncertainties. A precise wind power prediction aims to support te operation of large power systems or microgrids in te scope of te intraday resources sceduling model, namely wit a time orizon of 5-10 minutes, and tis can efficiently upold transmission and distribution system operators to improve te power network control and management. A Sortterm A based wind power prediction model, for day-aead energy management and sceduling in microgrids, is developed utilizing a real database of 12 monts wit 10 minutes time interval data from measured information of online supervisory control and data acquisition (SCADA) of Goldwind microgrid wind turbine system (Beijing, Cina) as well as umerical Weater rediction (W). Te A predicted wind power as been compared wit te actual power of Goldwind wind turbine system. Te predicted wind power as sown acceptable agreement wit te actual power output tat as been recorded by te SCADA wic confirms te robustness and accuracy of te A wind power forecasting model developed in tis researc paper. Keywords: Artificial neural network, microgrid, numerical weater predictions, resource sceduling, supervisory control and data acquisition, wind power prediction. 1. ITRODUCTIO Kyoto rotocol wit te aim of acieving te stabilization of greenouse gas concentrations in te atmospere at a level tat would prevent dangerous antropogenic interference wit te climate system is an environmental convention [43]. By adopting tis protocol, due to environmental benefits of renewable resources particularly wind power generation, utilization of tese types of energy as acquired noticeable consideration in eye-catcing number of countries, recently. In comparison wit te environmental damages of traditional sources of energy, te ecological influences of wind power is relatively negligible [43]. In fact, wind power fuel usage is incomparable wit fossil power plants as well as fuel emission. Furtermore, wind power enjoys infinitesimal progress expenses, in addition to an average cost of investigation. However, despite remarkable environmental benefits, te continuous and caotic fluctuations of wind speed make te output power of wind farms completely stocastic and different from tose of conventional units. However, instant electrical generation must be equal to te grid consumption to fulfill te network stability requirements and te spare capacity may be reduced in te power system [43]. Due to tis uncertainty, it may bestow ample callenges to connect large quantities of wind power into a power system. However, tis callenge is not insuperable. In order to increase te economic efficiency and acceptability of te wind power and to permit a diminising in te punisment of an instantaneous spot market coming from extra estimation or underrating of te production, te exact prediction of wind power as well as wind velocity is required. Surely, a reliable prediction system can elp distribution system operators and 144

2 power marketers to make a better decision on critical situation [43]. owadays, several metods ave been developed to predict te wind power and speed. Existing metods can be arranged as statistical, pysical and time series modeling metods based on te used prediction models [1]. Te pysical metods are based on local meteorological service or W model data of te lower atmospere (i.e. in relation wit atmosperic pressure degree, at 2m or 10m eigts above te ground) associated wit topological data like obstacles, rougness and orograpy. Te core idea of pysical approaces is to estimate generation power of te wind turbine wit outsourcing tese obtained data up or down to exact eigt of te wind turbine ub and ten, utilizing te manufacturer's power curve for te logaritmic power law or te specific wind turbine. Statistical forecasting tecniques are only based on one or more models. Tese approaces found coordination between predict and istorical numerical quantity of meteorological parameters and wind power measurements associated wit te power quantities coming from istorical data [43]. Moreover, statistical models enjoy freedom of using W models; wereas, pysical models must utilize W models [2]-[5]. Recently, it is prevalent tat autors use a combination of a progressive statistical model and pysical metods besides eac oter to reac an optimal approac tat is applicable for longer orizons of forecasting system. In tese metods statistical model play supplemental role to data gatered by pysical metods. In fact, models not utilizing numerical wind prediction suffer from lack of preciseness in time orizons more tan 3 or 4 ours [6]. Altoug two main classes of tecniques ave been recognized for te wind prediction, (in [7] and [8], compreensive reviews of tese metods are prepared), as aforementioned, combination of statistical and pysical metods are more prevalent tan te oters [9], [10]. Furtermore, several oter spatial correlation tecniques are proposed for sort term wind power forecasting wit te goal of acieving iger prediction accuracy [11]. However, by te passage of time, more advanced metods ave been proposed. To tis end, Artificial eural etwork (A) in [12], [13], A wit adaptive Bayesian learning and Gaussian process approximation in [14], combination of A wit wavelet transform in [15], fuzzy logic metods in [10], [16], Kalman filter in [17], support vector macine in [18] and some ybrid metods in [6] ave been proposed for wind power prediction. In tis researc paper, a new effective wind power forecasting metod based on combination of measured data from SCADA and W model is proposed. Te proposed metod employs feed forward A wit back propagation learning algoritm. In te course of te training process, te A utilized one year (on ten minutes interval basis) SCADA records of wind speed (as training input variable) and wind power (as training target variable). Forecasted meteorological parameter (i.e. wind speed) from a W model is ten applied as input to te developed (trained) A wind power prediction model in order to estimate te next day wind power output of an actual wind power system ( Goldwind microgrid wind turbine system, Beijing, Cina). Te forecasting is for te next 24 ours wit 10 minutes intervals. Simulation results sow tat te presented metod can effectively enance te exactness of te wind power forecasting. Te paper is organized as follows: Section 2 describes an overview of te wind power prediction. Section 3 introduces proposed model in wic tecnique of employing SCADA system and W model is described. Descriptions of A as prediction system and te prediction results for te casestudies considered are provided in sections 4 and 5, respectively. Te paper is concluded in section LITERATURE REVIEV OF SHORT- TERM WID OWER FORECASTIG AROACHES Te very sort-term wind power prediction tecniques utilize statistical models suc as ARX, ARMA and Kalman Filters; wile, tey are mostly based on time series approac. In tis category, te input data is past values of te forecasted variable like wind generation and wind speed; wereas, explanatory variables like temperature, wind direction and umidity can be employed in order to enance precision and applicability of te forecasting metods. Tis model as te capability of prediction for orizons between 3 6 ours according to tis consideration tat tey merely use te past production data [43]. In a ierarcical classification of sort-term wind power prediction approaces, two major groups are consisting: 1) tose tat use manufacturer s power curve or empirical curves of wind turbines to map predicted wind speed to te generation output power, and 2) tecniques in wic wind generation power can be calculated directly witout interfering te turbines caracteristics and role of W model is intensively igligted for te sort-term wind power forecasting [43]. Te second approac is adopted in tis researc wic combines measured data from SCADA records and W model witout referring manufacturer s power curve or empirical curves of wind turbines. Table 1. An overview of metods for sort-term wind power prediction Sort-Term Wind ower rediction Models Adaptive eural Fuzzy System [30] Local olynomial Regression [31] Bayesian Clustering by Dynamics (BCD) [32] Support Vector Macines [33] Mixture of Experts [33] Locally Recurrent eural etworks [3] Autoregressive wit Exogenous input (ARX) [34] Autoregressive wit Exogenous Input and Multi-timescale arameter (ARXM) [35] Random Forests [36] eural etworks [37] 145

3 3. WID OWER REDICTIO MODEL 3.1 roposed Wind ower rediction Metod In tis researc paper, a sort term wind power forecasting using feed forward back propagation neural network (FF) is presented. Te basic data sources for wind power prediction are istorical measurement records of wind turbine SCADA system database and numerical weater prediction (W). To construct te model te forecasting system used istorical measurement records of SCADA system database from a real wind power system ( Goldwind microgrid wind turbine system, Beijing, Cina). International Journal of Science and Engineering Applications Figure 1. Block diagram illustration of A training Te A wind forecasting model as been trained utilizing 12 monts data wit 10 minutes interval provided from Goldwind wind turbine SCADA system. SCADA recorded wind speed and wind power data ave been used as te network training input and target respectively as sown in Figure 1. In te process of modeling, istorical data provided from SCADA are used to train an A tat effectively can estimate a transfer function between te specific patterns of input and output vectors. Finally, te forecasted wind speed data provided from W model (WRF) projected around te vicinity of te wind farm as been preprocessed and applied to te developed (trained) A forecasting model in order to estimate wind power for te next 24 ours of te next day on 10 minutes interval basis as sown in Figure Online SCADA System SCADA system as te nerve center and inseparable element of te wind farm plays a vital role for te prediction system. Usage of an online SCADA gives te operator freedom to oversee wind farm by supervising all of te wind turbines. Tis opportunity is provided for operator to set proper actions in critical situations by a 10 minutes record of te wind park turbines information. Moreover, tis supervision system provides a compreensive record of te wind speed and output powers as well as availability of turbines, wic acts as a foundation for te sort term wind power forecasting. Figure 2. Block diagram illustration of A wind power forecaster for next day 3.3 umerical Weater rediction (W) Model owadays, wind data as a non-negligible impact on wind power prediction. Tere are several approaces to obtain te wind data: observations, data mining and numerical weater simulations. Te most straigtforward and reliable way to obtain wind data is troug on-site observations. However, tey are not usually available. Data mining is flexible, but its ability to downscale te weater data is limited. Te W models use pysical conservation of energy equations and tis allows a more realistic downscaling of te data. In fact, ig-resolution W of wind plays te key role for power forecasting. In recent years, regarding availability of enanced computational systems, many wind power estimation researces are directed utilizing W models wind data. Tis researcers use several W models like WRF, COSMO, MM5, and RAMS [30] - [33]. Also, various metods of extrapolation suc as logaritmic law and wind sear power law ave proposed by autors to provide appropriate wind information at te eigt of wind turbine ub (i.e. approximately 50 m) using meteorological data tat are gatered at 10 m above of te ground (According to World Meteorological Organization (WMO) approval)[19]. 4. ROOSED FRAMEWORK OF A 4.1 Overview of A eural network is a powerful data modeling tool tat is able to capture and represent te complex input/output relationsips. An Artificial eural etwork (A) is an information processing paradigm tat is inspired by te way 146

4 biological nervous systems, suc as te brain, process information. Te key element of tis paradigm is te novel structure of te information processing system. It is composed of a large number of igly interconnected processing elements (neurons) working in unison to solve specific problems [45]. As, like people, learn by example. An A is configured for a specific application, suc as pattern recognition or data classification, troug a learning process. Learning in biological systems involves adjustments to te synaptic connections tat exist between te neurons. Tis is true of As as well. And for te validation process A is followed, te uman brain provides proof of te existence of massive neural networks tat can succeed at tose cognitive, perceptual, and control tasks in wic umans are successful [45]. An illustrative representation of a multi-layer neural network is sown in Figure 3. Figure 3. Multi-layer neural network Were Xi te i t input to te node (neuron) and Yi is te output of te node. Te matematical equation wic sows te relationsip between te inputs Xi to te network and te output Yi of te network is given by equations (1) below. n Y i = fi ( W. X j + bi ) (1) ij j = 1 Were Xj is te j t input to te node, Yi is te output of te node, Wij is te connection weigt between te input node and output node, bi is te bias of te node, and fi is known as activation function tat determines te property of te neural network. Te mean squared error (MSE) of te network is defined by equations (2) below. MSE = 1 ( T Y ) 2 i i i = 1 Were Ti is te target at i t pattern, Yi is te prediction of te network s output at i t pattern and is te number of training set samples. 4.2 A Learning: Back ropagation Algoritm Tere are different types of training algoritms used for A learning; tis paper used te Levenberg-Marquardz back propagation supervised training algoritm wic is described in Figure 4 below. (2) Figure 4. Back propagation supervised training algoritm Te back propagation algoritm cycles troug two distinct passes, a forward pass followed by a backward pass troug te layers of te network. Te algoritm alternates between tese passes several times as it scans te training data as described below [44]: Forward ass: Computation of outputs of all te neurons in te network Te algoritm starts wit te first idden layer using as input values te independent variables of a case from te training data set. Te neuron outputs are computed for all neurons in te first idden layer by performing te relevant sum and activation function evaluations. Tese outputs are te inputs for neurons in te second idden layer. Again te relevant sum and activation function calculations are performed to compute te outputs of second layer neurons. Backward pass: ropagation of error and adjustment of weigts Tis pase begins wit te computation of error at eac neuron in te output layer. A popular error function is te squared difference between Yi te output of node i and Ti te target value for tat node. Tis error value is calculated continuously for eac data point in te training data set, and te new value of te weigt Wij of te connection from node i to node j is adjusted continuously until te error value reaces zero or falls below a certain tresold value. Te backward propagation of weigt adjustments along tese lines continues until we reac te input layer. At tis time we ave a new set of weigts on wic we can make a new forward pass wen presented wit a training data observation. 4.3 Wind ower Forecasting Accuracy Measures In order to evaluate te accuracy of te wind power prediction, te mean absolute error (MAE) and te mean absolute error percentage (MAE) can be used. Tese criterions are defined as follows: MAE = 1 = 1 a f (3) 147

5 MAE = 100% = 1 a f a (4) Were a and f are te actual and forecasted wind power, respectively, at period, and corresponds to te number of forecasted periods. Te MAE is an average of te absolute errors. In MAE te absolute values of all te percentage errors are added and te average is computed, producing a measure of relative overall fit. Anoter criterion tat may be used is te standard deviation (SD) tat represents a measure of ow spread out te numbers is. SD = 1 ( f ) -1 = 1 (5) Were is te wind power average. Te root mean square error (RMSE) is anoter frequently used measure of te differences between forecasted and te actual observed values. Te RMSE is te square root of te variance. RMSE gives a relatively ig weigt to large errors. RMSE = 1 ( = 1 a f 2 ) Te RMSE is most useful wen large errors are particularly undesirable. Te MAE and te RMSE can be used togeter to diagnose te variation in te errors in a set of forecasts [46]. Te RMSE will always be larger or equal to te MAE; te greater difference between tem, te greater te variance in te individual errors in te sample [46]. In tis case study te MAE, RMSE and SD ave been used. (6) Figure 5. One Day-Aead Wind ower Forecast for 26/3/2015 Simulation results ave sown tat te predicted Goldwind microgrid wind turbine output power is almost similar wit te actual wind power recorded by te SCADA wic indicates te effective performance of te A wind power forecaster model tat as been developed in te paper. Table 2 presents te forecasting errors tat ave been calculated for eac day s forecasting simulation. In order to evaluate te accuracy of te wind power prediction model developed in tis paper, it was considered te mean absolute error (MAE), te root mean square error (RMSE), and te mean absolute percentage error (MAE) as reported in section CASE STUDY AD RESULTS In order to evaluate performance of te proposed wind power forecasting sceme, te prediction model was built for te Goldwind microgrid wind farm, Beijing, Cina. Tis wind farm as a single wind turbine wit generating capacity of 2500KW. Te wind speed and power time series of tis wind farm are recorded from te 26 t Marc 2014 to te 25 t Marc Te forecasting information is given for te next 4 days (26 t Marc 2015 to te 29 t Marc 2015) for every 10 minutes intervals. Te following figures present te system performance (testing of te wind power forecasting model) in te period of 26 t to 29 t Marc Figures 5 8 sow te comparison of te actual Goldwind wind turbine output power wit tose of te predicted output power for Goldwind microgrid wind turbine system for eac of te four days aead (26/03/ /03/2015) of te last training date (25/03/2015) on 10 minutes time-interval basis. Figure 6. One Day-Aead Wind ower Forecast for 27/3/

6 Figure 7. One Day-Aead Wind ower Forecast for 28/3/2015 Figure 8. One Day-Aead Wind ower Forecast for 29/3/2015 Table 2. Wind power forecasting errors of eac day rediction Day Marc 26, 2015 Marc 27, 2015 Marc 28, 2015 Marc 29, 2015 rediction Accuracy Measures MAE RMSE MAE (KW) (KW) (%) rediction values of error obtained by metod proposed in [6] is 17% for 24 aead (one day aead or 144 ten minute intervals aead) from a specific day. Wile, tis error values for te presented paper are % for Marc 26, 2015, % for Marc 27, 2015, % for Marc 28, 2015 and % for Marc 29, 2015 tat are severely lower tan [6]. Moreover, te MAE and RMSE values presented in Table 2 ave sown less error values tan most of te literatures reported in tis paper. Results indicate tat proposed metod enjoys significantly effective performance for 24 (144 ten minute intervals) aead wind power forecasting. 6. COCLUSIOS In tis researc paper, a sort term wind power prediction model, for day-aead resource sceduling in microgrids, is proposed based on feed forward A for eac 24 ours of te next day. Te proposed A prediction model is developed utilizing SCADA records of an actual wind farm ( Goldwind microgrid wind turbine system, Beijing, Cina) wit a total installed power of 2500KW on Matlab neural network fitting tool using one year ten minutes interval data (26/03/ /03/2015) taken from Goldwind wind turbine SCADA system. Tis model was tested to predict Goldwind wind turbine output power for four days aead of te last training date (26/03/ /03/2015) and te results of te prediction ave been compared wit te actual output power of Goldwind wind turbine system. Finally, te forecasted wind speed data provided from W model (WRF) projected around te vicinity of te wind farm as been preprocessed and applied to te developed (trained) A forecasting model in order to estimate wind power for te 24 ours of te next day on 10 minutes interval. As demonstrated by te prediction results and te prediction error measures above, te predicted wind power as insignificant error and acceptable agreement wit te actual wind power taken from Goldwind wind turbine SCADA system wic sows te robustness and accuracy of te A wind power forecasting model developed in tis paper. Te model as resulted in desirable prediction accuracy and terefore, te forecasted wind power can be used as one of te input quantities for day-aead resource sceduling in microgrids wit ig wind energy penetration

7 7. REFERECES [1] J. Juban,. Siebert and G.. Kariniotakis "robabilistic sort-term wind power forecasting for te optimal management of wind generation", in roc IEEE Lausanne ower tec, vol. 15, pp , [2] M. C. Alexiadis,. S. Dokopoulos, and H. S. Sasamanoglou, Wind speed and power forecasting based on spatial correlation models, IEEE Transactions on Energy Conversion, vol. 14, no. 3, pp , Sep [3] T. G. Barbounis, J. B. Teocaris, M. C. Alexiadis, and. S. Dokopoulos, Long-term wind speed and power forecasting using local recurrent neural network models, IEEE Transactions on Energy Conversion, vol. 21, no. 1, pp , [4] L. Landberg, G. Giebel, H. A. ielsen, T. ielsen, and H. Madsen, "Sort-term prediction-an overview," Wind Energy, vol. 6, no. 3, pp , [5] G. Giebel, L. Landberg, G. Kariniotakis and R. Brownsword "State-of-te-art on metods and software tools for sort-term prediction of wind energy production", in roc. EWEC, [6] G. Sideratos and. D. Hatziargyriou, "An advanced statistical metod for wind power forecasting," IEEE Transactions on ower Systems, vol. 22, pp , [7] A. Costa, A. Crespo, J. avarro, G. Lizcano, H. Madsen, and E. Feitosa, "A review on te young istory of te wind power sort-term prediction," Renewable & Sustainable Energy Reviews, vol. 12, pp , [8] L. Ma, S. Y. Luan, C. W. Jiang, H. L. Liu, and Y. Zang, "A review on te forecasting of wind speed and generated power," Renewable & Sustainable Energy Reviews, vol. 13, pp , [9] L. Landberg, Sort-term prediction of te power production from wind farms, Journal of Wind Engineering and Industrial Aerodynamics, vol. 80, pp , [10] I. G. Damousis, M. C. Alexiadis, J. B. Teocaris, and. S. Dokopoulos, "A fuzzy model for wind speed prediction and power generation in wind parks using spatial correlation," IEEE Transactions on Energy Conversion, vol. 19, no. 2, pp , June [11] T.G. Barbounis, J.B. Teocaris, "A locally recurrent fuzzy neural network wit application to te wind speed prediction using spatial correlation," euro computing, vol. 70, pp , [12] alomares-salas, J.C., de l a Rosa, J.J.G.,Ramiro, J.G., Melgar, J., et.al, "ARIMA vs. eural networks for wind speed forecasting", in roc. IEEE International Conference on Computational Intelligence for Measurement Systems and Applications, pp , [13] L. Suui, D. C. Wunsc, E. A. O'Hair, M. G. Giesselmann, Using neural networks to estimate wind turbine power generation, IEEE Transactions on Energy Conversion, vol. 16, no.3, pp , Sept [14] R. Blonbou, Very sort-term wind power forecasting wit neural networks and adaptive Bayesian learning, Renewable Energy, pp , [15] J..S. Catalao, H.M.I. ousino, V.M.F. Mendes, Sortterm wind power forecasting in ortugal by neural networks and wavelet transform, Renewable Energy, pp , 2011 [16] G. Sideratos and. Hatziargyriou, "Using radial basis neural networks to estimate wind power production," in roc. IEEE ower and Energy Soc. General Meeting, pp. 1-7, [17]. Louka, G. Galanis,. Siebert, G. Kariniotakis,. Katsafados, I. ytaroulis and G. Kallos, "Improvements in wind speed forecasts for wind power prediction purposes using Kalman filtering," Journal of Wind Engineering and Industrial Aerodynamics, vol.96, pp , [18] D. Ying, J. Lu, Q. Li, Sort-term wind speed forecasting of wind farm based on least square-support vector macine, ower Systems Tecnology, vol.32, no. 15, pp , [19] Y. Carabi Arabian summer monsoon variability: teleconexion to ESO and IOD, Atmosperic Researc, pp , [20]. inson, L.E.A. Cristensen, H. Madsen,. Sørensen, M.H. Donovan, and L.E. Jensen, Regime-switcing modelling of te fluctuations of offsore wind generation, Journal of Wind Engineering & Industrial Aerodynamics, vol. 96, no. 12, pp , [21] G. Kariniotakis, E. ogaret, and G. Stavrakakis, Advanced Sort-Term Forecasting of Wind ower roduction, in roc. European Wind Energy Conference EWEC 97, Ireland, pp , October [22] I.G. Damousis and. Dokopoulos, A fuzzy model expert system for te forecasting of wind speed and power generation in wind farms, in roc. IEEE International Conference on ower Industry Computer Applications ICA 01, pp , [23] I.J. Ramírez-Rosado and L.A. Fernández-Jiménez, ext-day wind farm electric energy generation forecasting using fuzzy time-series, in roc. International Conference on Modeling, Identification, and Control, Innsbruck, Austria, pp , [24] G. Kariniotakis, E. ogaret, A.G. Dutton, J.A. Halliday, and A. Androutsos, Evaluation of Advanced Wind ower and Load Forecasting Metods for te Optimal Management of Isolated ower Systems, in roc. European Wind Energy Conference EWEC 99, pp , ice, France, Marc 1 5, [25] E.A. Bossanyi, Sort-Term Wind rediction Using Kalman Filters, Wind Engineering, vol. 9, no. 1, pp. 1 8, [26] S. Alpay, L. Bilir, S. Ozdemirny, and B. Ozerdem, Wind speed time series caracterization by Hilbert transform, International Journal of Energy Researc, vol. 30, pp , [27] R.E. Abdel-Aal, M.A. Eladidy, and S.M. Saaid, Modeling and forecasting te mean ourly wind speed 150

8 time series using GMDH-based abductive networks, Renewable Energy, vol. 34, no. 7, pp , July [28] C.W. otter and M. egnevistky, Very sort-term wind forecasting for Tasmanian power generation, IEEE Transactions on ower Systems, vol. 21, no. 2, pp , [29] T.H.M. El-Fouly, E.F. El-Saadany, and M.M.A. Salama, Grey redictor for Wind Energy Conversion Systems Output ower rediction, IEEE Transactions on ower Systems, vol. 21, no. 3, pp , [30] M. egnevitsky,. Jonson, and S. Santoso, Sort term Wind ower Forecasting using ybrid intelligent systems, in roc. IEEE ower Engineering Society General Meeting, pp. 1 4, June 24 28, [31] Ismael Sancez, Sort-term prediction of wind energy production, International Journal of Forecasting, vol. 22, no. 1, pp , [32] Su Fan, James R. Liao, Ryuici Yokoyama, and Luonan Cen, Forecasting te Wind Generation Using A Twostage Hybrid etwork Based on Meteorological Information, Information and Communications Engineering, Osaka Sangyo University, [33] R. Jursa, Wind power prediction wit different artificial intelligence models, in roc. European Wind Energy Conference EWEC 07, Milan, Italy, May [34] Mario J. Duran, Daniel Cros, and Jesus Riquelme, Sort-Term Wind ower Forecast Based on ARX Models, Journal of Energy Engineering, vol. 133, no. 3, pp , Sept [35] A. Yamaguci, T. Isiara, K. Sakai, T. Ogawa, and Y. Fujino, A ysical-statistical Approac for te Regional Wind ower Forecasting, in roc. European Wind Energy Conference EWEC 07, Milan, Italy, [36] Lionel Fugon, Jérémie Juban, and G. Kariniotakis, Data mining for Wind ower Forecasting, in roc. European Wind Energy Conference EWEC 08, Brussels, Belgium, April [37] A. Kusiak, H.-Y. Zeng, and Z. Song, Wind Farm ower rediction: A Data-Mining Approac, Wind Energy, vol. 12, no. 3, pp , [38] Kariniotakis, G.., inson,., "Uncertainty of sort-term wind power forecasts a metodology for on-line assessment", in roc International Conference on robabilistic Metods Applied to ower Systems, pp , [39] J. Connors, D. Martin, and L. Atlas, "Recurrent neural networks and robust time series prediction", IEEE Transactions on eural etworks, vol. 5, pp , [40] E. Ataspaz-Gargari and C. Lucas, "Imperialist Competitive Algoritm: An Algoritm for Optimization Inspired by Imperialistic Competition", in roc. IEEE Congress on Evolutionary Computation (CEC 2007), pp , [41] S. Mollaiy Berneti, M. Sabazian, An Imperialist Competitive Algoritm Artificial eural etwork Metod to redict Oil Flow Rate of te Wells, International Journal of Computer Applications, [42] ational Climate Data and Information Arcive, [Online]. Available: ttp:// [43] Gadi, M. Jabbari, S. Hakimi Gilani, H. Afrakte, and A. Bagramian, A novel euristic metod for wind farm power prediction: A case study, International Journal of Electrical ower & Energy Systems, [44] ocw.mit.edu [45] ijecs.in [46] [47] stonito.com [48] M. Jabbari, Gadi, S. Hakimi Gilani, H. Afrakte and A. Bagramian, Sort-Term and Very Sort-Term Wind ower Forecasting Using a Hybrid ICA- Metod, International Journal of Computing and Digital Systems, vol. 3, o. 1, (2014)