Вы находитесь на странице: 1из 6

GA-ANN for Short-Term Wind Energy Prediction

Mohan Kolhe, Tzu Chao Lin, and Jussi Maunuksela


satisfy global energy demands and unlock a new era of economic growth, technological progress and environmental protection [1]. For a long time, wind forecasting has been needed for different type of applications in the land and offshore region, like for planning of construction and operation-related works as well as for prediction of power output from wind turbines located in offshore / on shore. As wind is a variable generation resource and weather changes can cause large and rapid changes in output, system operators will need accurate and robust wind energy forecasting systems in the future. Rapid changes of wind generation relative to load require rapid dispatching of generation and transmission resources to balance generation versus load, regulate voltage and frequency, and maintain system performance within the limits established by National Grid. Accurate wind energy forecasts can help the system operator to anticipate rapid changes of wind energy generation versus load and in making the decisions. Usually, such forecasting has been done by using statistical time-series analysis or by adopting complex atmospheric models. Milligan et al. [2] presented several statistical wind forecasting models and investigated the time-series analysis models known as autoregressive moving average (ARMA) models for both wind speed and wind power output. In several cases, it was found that these ARMA models did a good job forecasting over the testing time frame. The Risoe National Laboratory for Sustainable Energy at the Technical University of Denmark (DTU) has used detailed, area-specific three dimensional weather models and has worked with numerical weather prediction models, such as High Resolution Limited Area Model, the UK Meteorological Office Mesoscale model or the Lokal-Modell of the German Weather Service [3]. These systems work like weather forecast models, predicting wind speeds and directions for all wind farms in a given area. Based on these models, the wind output power has been calculated by using physical equations such as in WAsP (Wind, Atlas Analysis and Application Program) [3]. These models have been designed primarily to provide information for grid network companies to be able to operate their power systems. Wind power forecasting based on stochastic processes has been used for wind energy trading [4]. In ref [5], the neural networks method was used to forecast the wind speed for coastal locations and it was found to provide much more accurate forecasts compared to the traditional stochastic timeseries model of ARMA. A simple probabilistic method has

Abstract -- Wind turbine power output is totally intermittent in the nature. For grid connected wind turbine generators, power system operators (transmission system operators) need reliable and robust wind power forecasting system. Rapid changes in the wind generation relative to the load require proper energy management system to maintain the power system stability and of course to balance the power generation, frequency, voltage regulation within the statutory limits. Accurate wind energy forecasting helps the power system transmission system operators in anticipating rapid changes in wind turbine power output with respect to load and helps in making decision not only for optimum energy management but also for energy trading in the open electricity market. In this article, wind energy prediction for short term has been done by using artificial network in combination with genetic algorithms. The developed model has tested and analyzed for Taiwan Wind Power companys real operational results. The results show that the combination of ANN and GA model gives wind power output prediction very well except during the occurrences of gust. It has been observed that ANN performs well in non-linear mapping, but the combination of ANN-GA gives more accurate prediction. This model has been implemented in different time scales, which will also be useful for wind energy trading in the open electricity market. Index Terms-- Wind Energy Forecasting, Artificial Neural Network, Genetic Algorithm.

I. INTRODUCTION

global energy challenge of our time is to tackle the threat of climate change, meet the rising demand for energy and to safeguard security of energy supplies. The wind industry is one of the world's fastest growing energy sectors and offers the best opportunity to begin the transition to a global economy based on sustainable energy. Wind Force 12 is a global industry blueprint which demonstrates that there are no technical, economic or resource barriers to supplying 12% of the world's electricity needs with wind power alone by 2020 and this against the challenging backdrop of a projected two thirds increase of electricity demand by that date [1]. Nowadays wind industry is capable of becoming a dynamic, innovative 80 billion annual business by 2020, helping to
Mohan Kolhe is with the University College London, School of Energy & Resources and International Energy Policy Institute. www.ucl.ac.uk/australia (e-mail: m.kolhe@ucl.ac.uk). Tzu Chao Lin is with the Taiwan Power Company, Taiwan (e-mail: u772020@taipower.com.tw). Jussi Maunuksela is with Renewable Energy Programme, University of Jyvaskyla, PO Box 35, FI-40014, Jyvaskyla, Finland (E-mail: jussi.o.maunuksela@jyu.fi)

978-1-4244-6255-1/11/$26.00 2011 IEEE

been developed to predict the ability of energy storage to increase the penetration of intermittent renewable generation such as wind and to enhance the value of the electricity generated by time-shifting delivery to the network [6]. Due to large variability of wind generation, power quality and stability of the power system will become poor as well as the system monitoring, control and protection [7,8,9]. Hence wind power forecasting technique is going to assist in maintaining power quality and stability of the grid as well as system control and protection. It is indicating that generalizing capacities are needed in wind speed forecasting over different time scales. Predictability of wind in nature is extremely uncertain, so no single technique can be entirely satisfactory. This leaves room for alternative approaches. Time series modeling and forecasting require simulating the system behavior in the time domain. It is easy for a linear system by adjusting parameters of curve fitting, but it is more complex for nonlinear system and needs application of identification methods to simulate or resolve the mapping of system in order to forecast the future behavior of the power output model [3]. The forecasting techniques used for mapping nonlinear systems include methods such as linear regression [10-12], time series model [13-16], and ARMA method [17]. Linear regression method utilizes the common relationships among all variables in the model to predict the relative change of one variable by the change in other variables. When the relationship among the reference control variation is unclear, it will cause forecasting error. In time series model, it is assumed that the power output is a stationary time series and has normal distribution characteristics. Studies show that the error of forecast increases when the historical data does not support this condition. Finally, the ARMA model is known to be not efficient in modeling discrepancies in data, i.e., weekends, holidays, and seasonal change periods [18]. The advantage of these models is the easy physical interpretation of model parameters. Many studies are focusing on forecasting techniques based on artificial intelligence (AI) and artificial neural network (ANN) and have received a lot of attention [19-28]. In most of these works ANN has been utilized as a black box tool to identify the power output behavior. Due to good mapping ability of ANN in nonlinear function, it performs the forecasting. In order to speed up the ANN process and to promote its accuracy many optimization techniques have been proposed in combination with the ANN. In a field of optimization methods, genetic algorithm (GA) is the most popular topic. The GA is a directed random search technique for optimization problems where the numbers of parameters are large and analytical solutions are difficult to obtain [18, 29, 30]. The GA can assist in finding the optimal solution globally over a domain. It has been utilized in different areas such as fuzzy control, path planning, modelling and classification etc. The GA optimization process consists of searching a group of feasible solutions. This search goes along the orientation of general optimum avoiding the local minimum. The ability of GA in training the ANN for short-

term forecasting has been observed in Ref. [29] and is a key point in implementing GA with ANN for short-term wind energy forecasting. In this work, a hybrid model for short term wind energy forecasting has been developed by utilizing the nonlinear mapping ability of ANN and the ability of GA for finding the optimum solution. The results of this study will be useful in forecasting wind energy output of operational wind power stations. II. METHODS AND PROCESSES The wind power is variable in nature and a properly developed ANN can be used to forecast the wind power output. By combining the GA optimization with ANN, the forecasting can be done more efficiently and precisely. In this article, the hybrid forecasting model by combining GA and ANN has been developed and used for wind power forecasting. A. Artificial Neural Network The basic element of an ANN is neuron. The weighted summation function of inputs xi and weights wi is given by [20, 32],

uk = x1w1 + x2 w2 +

+ xn wn = xi wi
i =1

The activation function of neuron is used as a transfer function to stimulate the response in the same manner as human neuron does. Normally, different types of activation functions such as threshold function, piecewise-linear function, and sigmoid function in accordance with the problem have been used. The neuron output as a mathematical operation is

(1)

yk = (uk k ), (2) where is the activation function, is the weighted sum of


the inputs, and is the threshold. When many groups of neurons has been collected in the structure, the whole framework performs like a human learning operation [18-20]. Multilayer feed-forward neural network or back propagation ANN (BPANN) is one of the most commonly used networks in different applications. By inputting the training data, the experience values in the network are going to change. Such a learning experience is stored in thresholds and weights. Normally, these inputs are often critical parameters that affect the system like, e.g., in wind forecasting, where wind speed and wind direction are the critical parameters. In this work, a three-layer feed-forward network has been used in developing the model. The learning method used during the learning process is a gradient descent with moment weight change and historical learning experience. The calculation of the weight adjustment for a given neuron (i, j) at layer k is denoted as:

Wijk (t + 1) = Wijk (t ) + Wijk (t 1),


where Wijk (t + 1) = Wijk (t + 1) Wijk (t ) , and

is the learning rate

(3)

is the momentum term and is the coefficient of the last

modification. The number of neurons in the hidden layer is a parameter which affects the performance of ANN. There are batch and sequential modification methods for the training process of ANN. In batch learning, training process is done after accumulating all the training datasets and sequential learning means that the network performs learning as each single data is inputted. ANN can work with a highly nonlinear model very well after a training process with enough connections and learning epochs [20]. B. Genetic Algorithm In general, the GA is a method for solving optimization problems based on natural selection and imitation of the course of evolution. The overall effect of the GA is that a new population moves toward better solutions with fitness values of objective function. The GA is often applied to solve various optimization problems that are not well suited for standard optimization algorithms, including problems in which the objective function is discontinuous, non-differentiable, stochastic or highly nonlinear. When we form the optimization problem, the space of variables should be defined. A value of one variable is often looked as a gene and the composition of these genes will form a chromosome as a point in hypercube space. Each chromosome represents a solution and goal is to find the best solution that is given by the optimum of the objective function (lowest/highest). There will be a group of chromosomes, or an initial population, randomly generated within the predefined range in the beginning. After the population is generated, the values of the objective function of these chromosomes are evaluated and the GA process proceeds [18, 29, 30]. Some of the chromosomes with a good objective solution will be selected into the mating pool that is completed with others selected according to some rules like Roulette wheel selection or ranked rule. In this work, GANN is used to denote the ANN model in which the parameters are optimized by the GA. In this model, a training dataset is first pre-processed and then inputted into the GANN. According to the desired output, the square error between output and target is calculated and the total sums of these square errors have been used as objective fitness function. For searching the optimum, the GA has been applied in the solution. C. Process GANN An adaptive artificial neural network (ANN) has been combined with GA in developing a model for mapping wind speed and wind turbine output power. The learning step and the momentum coefficient in ANN have been decided based on the steepest descent direction in the optimization. ANN has learnt the mapping relation by adjusting internal connecting weights and threshold and then searched the least square error. Even though the ability of ANN has been proven in nonlinear mapping, the experience shows that it performs forecasting more accurately if as many characteristic features as possible are extracted from raw data for training. The parameters of ANN have been optimized by GA with Roulette wheel selection, Gaussian mutation and random mutation. The flow chart given in Fig. 1 demonstrates the process of GANN

forecasting model. The trained GANN model is directly implemented to forecast wind power output from training and test raw data.

Fig.1 GANN wind power forecasting process

D. Data Processing When the operation records from wind power stations have been taken, the raw data has been pre-processed to get rid of abundant or un-required data. E.g. in our case, the SCADA (Supervisory Control and Data Acquisition) system has recorded detailed data sequentially even in the maintenance state or shutdown by sudden gust and, thus, the raw data has not been helpful as such for the training of GANN. The raw data is therefore pre-processed and the characteristic data has extracted. The characteristic data (v, p) used in characterizing the wind turbine power production has been sampled from the raw data using the following criteria:

1. 0 v vcutout and 0 p Pmax AND 2. vnom v < vcutout and p 0.9 Pnom AND 3. vcutin v < vnom and p 0.
(4)

where vcutin, vnom, vcutout and Pmax are the operational parameters of the studied wind turbine type, i.e., the cut-in, nominal, and cut-out wind speeds and the rated power, respectively. This prevents the noise signal from disturbing the learning of GANN model. However, the extracted data cannot be fed directly into the model, because the range of the data has been very large while considering different operation parameters of the wind turbine. This problem has been solved by normalizing the data to uniform space mapping. Therefore, the operational parameters of the specific wind turbine should be known in advance. The extracted and normalized operational data of wind turbine are obtained so that they are ready to be used in the training process. It has been proposed in Ref. [31] that curve fitting can be utilized to present the characteristic behavior of wind turbine. This idea of curve fitting was adopted and applied as the Sigmoidal fitting of the Boltzmann function (Eq. 5) (rather than the more common fitting of the polynomial function): A1 A2 f ( x) = A2 + , 1 + exp[( x x0 ) / x] (5)

where xo is the centre value of the exponential distribution, x is the width of the exponential distribution, A1 is the initial value of y(-), and A2 is the final value of y(+). E. Training Processing In training process, the GANN model has been trained by the extracted dataset. At first, the training data have been fed into the model and the total square error of all training data has been calculated as batch learning. The goal of the GA has been in searching the optimum parameters of ANN that minimize the total square error of the training data. Second, the GA utilizes the Roullete wheel selection, real valued intermediate crossover, and the mutation of Gaussian and random. In Gaussian mutation, GA performs mutation by the Gaussian probability distribution function and it enforces the accuracy of each gene in chromosome by shortening the range. In random mutation, each gene mutates with a random probability which enforces the optimization. The number of chromosomes has been set as 150 to search the optimum and it searched 100 epochs. Theoretically, the smaller chromosomes save more running time, but it affects the searching dimension of optimum solution and it prevents the searching from getting stuck into a local minimum. After the training process of GANN, the final trained values of the connecting weights and thresholds of ANN are saved and it is used in checking the final performance of the training. F. Forecasting Process When the GANN model has been trained, it is ready to be used for forecasting. Just as the GANN has been trained by the training dataset, the forecasting dataset has been facilitated to perform the forecasting process [32]. It should be noticed that the output of GANN has managed in a reasonable range. Because of the prior knowledge of the maximum power output of wind turbine, the forecasting output of GANN maximum power output is limited. For these possibilities and realization, the training output or forecasting output has limited to the prescribed range in which minimum is zero and maximum is set as the real maximum power. The result of forecasting has been evaluated to observe the performance of GANN. III. WIND POWER STATION In this work, one wind power stations of Taiwan Power Company is selected for short-term wind power forecasting. The details of the wind turbines of this commercial wind power plant are as: Chung Tung wind power plant has eight Enercons 600 kW wind turbines (E40). Rated power of each turbine is 600 kW, transmission gearless, cut-in wind speed 2 m/s, nominal wind speed 12 m/s, cut-out wind speed 28 m/s. The wind operation records from this wind power station have been selected to demonstrate the forecasting characteristics of the developed GANN model.

speed (average, maximum and minimum), rotational speed of rotor (average, maximum and minimum), power output (average, maximum and minimum), nacelle position, etc. For this work, the dataset for November 2006 has been first extracted, and then normalized by the nominal wind speed (12 m/s) and the nominal power output (600 kW). For November 2006, the power output and wind speed of 10 minutes intervals are shown in Fig.2.

Fig.2 Chung Tung WPS wind speed and power output (November 2006)

Also the scatter plot of power output versus wind speed has been shown in Fig.3. A curve was fitted into the normalized data with a sigmoidal fitting of the Boltzmann function and the coefficients for the Boltzmann function were A1=0.0046, A2=1.04911, x0=0.79169, and x=0.13095.

Fig.3 Scatter plot of power output vs. wind speed (November 2006)

After pre-processing the training dataset for the GANN, the input data has been stored in desired format to wait for feeding into the model. After 100 epochs with 150 offspring in searching the optimum and the final objective fitness of GA has been 0.35765. By plotting the objective fitness values of each epoch (Fig.4), one can observe the good optimization ability of the GA.

IV. RESULTS AND DISCUSSION The acquired operation records of the Chung Tung wind power station have been taken between October 24th and December 20th 2006 (total of 58 days). The records have been collected by SCADA every 10 minute and include time, wind

see that deviation is located between the acceptable range if there is a rare power drop and a sudden drop of wind speed. In normal sense, this is reasonable for GANN to forecast the power output with respect to the wind speed. If the gust can be measured as an input of GANN, then it will enforce the forecasting accuracy of this model.

Fig.4 Optimization process of GA

The calculated output using the optimum solution of the GANN architecture has been given in Fig.5, with training dataset. The power output of the trained GANN fits quite well with training data. To demonstrate the forecasting of this GANN model, the operational records for December 2006 have been taken into account (Fig.6). The dataset shows that there werent records with wind speed equaling zero, but there are still some zero power output values. These values are because of the sudden shutdown during the wind gust that exceeded the cut-off limit. Some sudden drops in power output are apparent, and there are only three periods with changing power output. Generally speaking, the power energy is abundant and stable in December 2006 at this site.

Fig.7 Forecasting and real power output of Chung Tung WPS

Fig.8 Forecasting deviation of Chung Tung WPS

Fig.5 Power output and training result of Chung Tung WPS

A. Discussions This work uses short-term operational records of commercial wind turbines in Taiwan to forecast the future power output. The results show that this GANN model can handle very well the variation behavior between wind speed and wind power output. However, the error due to sudden wind gust seriously affects the accuracy of forecasting which is not good in open electricity market. The accuracy of forecasting can be increased if one can consider the gust as an effective input. It has been observed that the WTG power output is zero when the wind speed (10 min average) is still high. By consulting the site engineers, it has been found that such shutdown (zero power output) states are caused due to sudden gust and they are difficult to avoid in the forecasting since they are due to design regulations (protective shutdown) for protecting the wind turbine from such wind gusts. V. CONCLUSION In this work, the methodology for short-term wind power forecasting of wind power plants has been developed by combining ANN and GA. The forecasting results obtained through the developed GANN methodology have been compared with the real operational results of Taiwan wind

Fig.6 Chung Tung WPS wind speed and power output (December 2006)

When the collected test dataset is fed into the trained GANN for forecasting, one can see that the forecasting power output follows the curvature with good extent (Fig.7). The deviation of each point shows that the forecasting error is around 6.52% except when the gust exceeds the cutout limit (Fig.8). We can

power plant. It has been observed that ANN performs well in non-linear mapping of the wind energy, but the combined ANN and GA gives more efficient and accurate results. The results show that the combination of ANN and GA model does wind power output forecasting very well except during the occurrences of gust. The forecasting accuracy of GANN is well while the power output is stable during steady wind speeds. However, the forecasting error becomes large at the time of gust during the operation. The accuracy of forecasting can be increased if one can consider the gust as an effective input. This forecasting model can also be implemented in different time-scales, which will not only help for wind energy trading in liberalized electricity markets, but also will help the system operator to anticipate rapid changes of wind energy generation versus load and in making the decisions for effective network power management. VI. ACKNOWLEDGEMENT Tzu Lin Chao is very much thankful to the Taiwan Power Company and his colleagues for providing the data and permitting to use for this study. VII. REFERENCES
[1] [2] [3] [4] [5] [6] [7] Global Wind Energy Council (GWEC), Wind Force 12: A blueprint to achieve 12% of the world's electricity from wind power by 2020, Belgium, Jun. 2005. M. Milligan, M. Schwartz, Y. Wan, Statistical Wind Power Forecasting Models: Results for U.S. Wind Farms, Preprint, NREL/CP-500-35087, National Renewable Energy Laboratory, Colorado, 2003. Thomas Ackermann, Wind Power in Power Systems, John Wiley and Sons Ltd, 2005. J. Matevosyan, L. Soder, Minimization of imbalance cost trading wind power on the short-term power market, IEEE Transactions on Power Systems, vol. 21, 2006, pp. 1396-1404. Anurag More and M.C. Deo, Forecasting wind with neural networks, Marine Structures, vol. 6, 2003, pp. 35-49. J.P. Barton, D.G. Infield, Energy storage and its use with intermittent renewable energy, IEEE Transactions on Energy Conversion, vol. 19, 2004, pp. 441-448. A. Arulampalam, M. Barnes, N. Jenkins, J.B. Ekanayake, Power quality and stability improvement of a wind farm using STATCOM supported with hybrid battery energy storage, IEE Proceedings Generation, Transmission and Distribution, vol. 153, 2006, p. 701-710 D. Karlsson, M. Hemmingsson, S. Lindahl, Wide area system monitoring and control - terminology, phenomena, and solution implementation strategies, IEEE Power and Energy Magazine, vol. 2, 2004, pp. 68-76. M.R. Patel, Wind and Solar Power Systems - Design Analysis and Operation, CRC Press, 2005. S. Kyung-Bin, H. Seong-Kwan, P. Jung-Wook, K. Dong-Jin, K. KyuHo, Hybrid load forecasting method with analysis of temperature sensitivities, IEEE Transactions on Power Systems, vol. 21, 2006, pp. 869-876. S. Kyung-Bin, B. Young-Sik, H. Dug Hun, G. Jang, Short-term load forecasting for the holidays using fuzzy linear regression method, IEEE Transactions on Power Systems, vol. 20, 2005, pp. 96-101. J. Guajardo, J. Miranda, R. Weber, A hybrid forecasting methodology using feature selection and support vector regression, in Fifth International Conference on Hybrid Intelligent Systems, 2005. M. Espinoza, C. Joye, R. Belmans, B. DeMoor, Short-Term Load Forecasting, Profile Identification, and Customer Segmentation: A Methodology Based on Periodic Time Series, IEEE Transactions on Power Systems, vol. 20, 2005, pp. 1622-1630. A. Sfetsos, C. Siriopoulos, Time Series Forecasting of Averaged Data With Efficient Use of Information, IEEE Transactions on Systems, Man and Cybernetics Part A, vol. 35, 2005, pp. 738-745.

[8]

[9] [10]

[11] [12] [13]

[15] A. Sfetsos, C. Siriopoulos, Time series forecasting with a hybrid clustering scheme and pattern recognition, IEEE Transactions on Systems, Man and Cybernetics, Part A, vol. 34, 2004, pp. 399-405. [16] A. M. Gonzalez, A. M. S. Roque, J. Garcia-Gonzalez,, Modeling and forecasting electricity prices with input/output hidden Markov models, IEEE Transactions on Power Systems, vol. 20, 2005, pp. 13-24. [17] H. Shyh-Jier, S. Kuang-Rong, Short-term load forecasting via ARMA model identification including non-Gaussian process considerations, IEEE Transactions on Power Systems, vol. 18, 2003, pp. 673-679. [18] L. Gwo-Ching, T. Ta-Peng, Application of a fuzzy neural network combined with a chaos genetic algorithm and simulated annealing to short-term load forecasting, IEEE Transactions on Evolutionary Computation, vol. 10, 2006, pp. 330-340. [19] G. J. Tsekouras, N. D. Hatziargyriou, E. N. Dialynas, An optimized adaptive neural network for annual midterm energy forecasting, IEEE Transactions on Power Systems, vol. 21, 2006, pp. 385-391. [20] S. Kelouwani, K. Agbossou, Nonlinear model identification of wind turbine with a neural network, IEEE Transactions on Energy Conversion, vol. 19, 2004, pp. 607-612. [21] D. Baczynski, M. Parol, Influence of artificial neural network structure on quality of short-term electric energy consumption forecast, IEE Proceedings-Generation, Transmission and Distribution, vol. 151, 2004, pp. 241-245. [22] T. Senjyu, H. Takara, K. Uezato, T. Funabashi, One-hour-ahead load forecasting using neural network, IEEE Transactions on Power Systems, vol. 17, 2002, pp. 113-118. [23] H. S. Hippert, C. E. Pedreira, R. C. Souza, Neural networks for shortterm load forecasting: a review and evaluation, IEEE Transactions on Power Systems, vol. 16, 2001, pp. 44-55. [24] N. Amjady, M. Ehsan, Evaluation of power systems reliability by an artificial neural network, IEEE Transactions on Power Systems, vol. 14, 1999, pp. 287-292. [25] H. S. Hippert, C. E. Pedreira, Estimating temperature profiles for shortterm load forecasting: neural networks compared to linear models, IEE Proceedings Generation, Transmission and Distribution, vol. 151, 2004, pp. 543-547. [26] T. Senjyu, P. Mandal, K. Uezato, T. Funabashi, Next day load curve forecasting using recurrent neural network structure, IEE ProceedingsGeneration, Transmission and Distribution, vol. 151, 2004, pp. 388-394. [27] T. Saksornchai, L. Wei-Jen, K. Methaprayoon, J. R. Liao, R. J. Ross, Improve the unit commitment scheduling by using the neural-networkbased short-term load forecasting, IEEE Transactions on Industry Applications, vol. 41, 2005, pp. 169-171. [28] T. G. Barbounis, J. B. Theocharis, M. C. Alexiadis, P. S. Dokopoulos, Long-term wind speed and power forecasting using local recurrent neural network models, IEEE Transaction on Energy Conversion, vol. 21, 2006, pp. 273-284. [29] S. H. Ling, F. H. F. Leung, H. K. Lam, L. Yim-Shu, P. K. S. Tam, A novel genetic-algorithm-based neural network for short-term load forecasting, IEEE Transactions on Industrial Electronics, vol. 50, 2003, pp. 793-799. [30] K. Daijin, K. Chulhyun, Forecasting time series with genetic fuzzy predictor ensemble, IEEE Transactions on Fuzzy Systems, vol. 5, 1997, pp. 523-535. [31] M.-S. Kang, Generation Cost Assessment of an Isolated Power System With a Fuzzy Wind Power Generation Model, IEEE Transactions on Energy Conversion, vol. 22, 2007, pp. 397-404. [32] M. Kolhe, T.C. Lin, J. Maunuksela, Wind energy forecasting by using artificial neural network genetic algorithm, ISBN 978-3-639-11297-9, VDM Publishing, 2008.

[14]

Вам также может понравиться