IEEE TRANSACTIONS ON SMART GRID, VOL. 5, NO. 1, JANUARY 2014 501 Short-Term Wind Power Ensemble Prediction Based on Gaussian Processes and Neural Networks Duehee Lee, Student Member, IEEE, and Ross Baldick, Fellow, IEEE Abstract—We propose an ensemble short-term wind power forecasting model that is based on our novel approaches and advanced forecasting techniques in the modern literature. The performance of our model has been veriﬁed by forecasting wind power up to 48 hours ahead at seven wind farms for one and a half years. Our model ranked fourth in the Power and En- ergy Society (PES) wind power forecasting competition. The fore- casting model uses 52 Neural Network (NN) sub-models and ﬁve Gaussian Process (GP) sub-models in parallel. For 48 hours, the NN sub-models forecast the future wind power based on historical wind power data and forecasted wind information. In parallel, for the ﬁrst ﬁve hours, ﬁve GP sub-models are used to forecast wind power using only historical wind power in order to provide accurate wind power forecasts to NN sub-models. These models provide various forecasts for the same hour, so the optimal fore- cast should be decided from overlapped forecasts by the decision process. Index Terms—Ensemble forecasting, Gaussian process, neural network, wind power forecasting competition. I. INTRODUCTION W IND POWER FORECASTING (WPF) starts from un- derstanding the physical mechanisms underlying wind, wind turbines, and by extension, wind farm conﬁgurations. The next step is to select observation data having high correlations with wind power from the physical mechanisms [1]. Then, in- ference models are selected to model those correlations con- necting observation data and wind power forecasts. According to the dependence of observation data on successive uniform time intervals, inference models can be classiﬁed into two rep- resentative models. The ﬁrst one is a time-series model where wind power forecasts depend on time-shifted wind power. Other time-shifted observation data, such as weather information, can be used as exogenous factors. The time-series model recursively predicts wind power by analyzing the linear stochastic structure of time-shifted observation data [2]. This model forecasts wind power directly, or it forecasts wind speed and then converts it into wind power. If more than two kinds of outputs depending on each other are forecasted, it is called the multivariate time-se- ries model. Manuscript received February 15, 2013; revised June 04, 2013, August 05, 2013, August 13, 2013, and August 14, 2013; accepted August 14, 2013. Date of publication September 23, 2013; date of current version December 24, 2013. This work was supported in part by National Science Foundation under Grant ECCS-1065224. Paper no. TSG-00130-2013. The authors are with Electrical and Computer Engineering, University of Texas at Austin, Austin, TX 78712 USA. Color versions of one or more of the ﬁgures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identiﬁer 10.1109/TSG.2013.2280649 The second model does not depend on the time sequence of observation data, and it is usually based on data mining or ma- chine learning algorithms. Therefore, this model can be trained by randomly enumerated pairs of input data and target data. This model focuses on ﬁnding patterns of input and output data. Both models can receive forecasted weather informa- tion—such as wind speed, wind direction, temperature, or air density—from the Numerical Weather Prediction (NWP) models. Data from NWP models is converted to wind power in three steps: downscaling global weather information to the local one, converting local weather information to the wind power of a single wind turbine through a power curve, and upscaling it to total wind power of that wind farm [3]. In these steps, two models mentioned above are used in the second step, and we research this second step. The time-series model has been widely used since it is simple and has shown high performance when forecasting the wind speed and wind power within a short time. Since the time-se- ries model is more suitable to forecast wind speed than wind power [4], much literature, starting with [5], forecasted the wind speed using the time-series model. For example, [6] forecasted the mean hourly wind speed through the time-series model up to an hour ahead, and [7] forecasted it up to 10 hours ahead. However, the time-series model cannot forecast more than a day ahead [8] for three reasons. First, since its outputs depend on previous outputs, if it should depend on forecasted outputs be- cause of short historical data, errors in previous forecasts easily propagate. Second, its forecasts eventually converge to the mean value as the forecast horizon increases. Third, the time-series model is based on the linear relationship between observations and forecasts, but wind turbines follow a nonlinear power curve which causes non-stationary characteristics of wind power. One of the methods for handling the non-stationarity is to divide wind power into categories that represent the varying status of wind power in order to keep the stationarity [9]. However, the number of regimes is limited and discrete, and regimes are ar- bitrarily deﬁned. Furthermore, if regimes are classiﬁed based on training data, models can not be trained by training data in other regimes. The other method is to transform input data using a nonlinear function, but it is hard to ﬁnd a suitable transforma- tion function. On the contrary, the second model is specialized to capture the nonlinear relationship between weather information and wind power by means of machine learning algorithms. For example, the support vector machine is used in [4] to forecast the wind power, the genetic algorithm is used to train the fuzzy logic model to forecast wind speed in [10], and the particle swarm op- timization is used to improve the performance of the fuzzy logic model in [11]. Moreover, [12] uses the nearest neighbor to de- 1949-3053 © 2013 IEEE