Method Of Nonparametric Wind Power Generation Forecast

As a result of the existence of wind randomness and volatility, the large-scale integration of wind power into grid will cause great effect on power system, such as the system stability and the power quality. Thus wind power dispatch is considered to be difficult. Therefore, forecasting the uncertainty of wind generation is a great tool to solve those problems.Typical wind generation probability forecast approaches are single predictive approaches, which commonly assume one of the forecasting models to be the best one. Single predictive approaches can be classified into two categories, i.e. parametric and nonparametric ones. Parametric approaches assume that predictive distribution follows the pre-defined shape, such as Gaussian distribution, Warped Gaussian distribution, Beta distribution, Versatile distribution and Logit-normal distribution. Parameters of the models can be achieved analytically and used to predict the wind generation. Since different wind farms may have distinct statistical properties, the predetermination of the probability distribution type restricts the universality of these approaches. The nonparametric approaches do not suffer from this restriction and avoid the model error introduced by the misjudgement of the probability distribution type of wind generation.However, each single approach involves inherent uncertainty and is applied to specific wind farms. At present, there is no forecasting method is suitable for each wind farms. Thus, the single best forecasting model is not really the optimal one. Ensemble forecasting approach is able to reduce the effect of random factors in the individual forecasting model and get a better nonparametric prediction result by combining superiority of different individual distribution forecasting methods.Currently, there are several following issues in short-term probabilistic wind power generation:â‘  the need to further improve the prediction accuracy, such as expectation prediction accuracy and distribution prediction accuracy;â‘¡ the existing methods are mostly parametric ones and the predetermination of the probability distribution type restricts the universality of these methods;â‘¢ the existing methods are mostly individual forecasting methods and the existing ensemble methods are either for spot forecast or parametric probabilistic forecast, which is not practical in real application.In this paper, both individual and ensemble nonparametric methods for probabilistic wind generation forecast are proposed. Firstly, a nonparametric approach for short-term probabilistic wind generation forecast based on the sparse Bayesian classification (SBC) and Dempster-Shafer (DS) theory is proposed. This approach is composed by the following steps:(1) Support vector machine (SVM) is applied for spot forecast of wind generation; (2) The range of forecast error is discretized into multiple intervals, and the conditional probability of each interval is estimated by a two-class sparse Bayesian classifier; (3)D-S theory is applied to integrate the probabilities forecasted by the classifiers to get an overall probability distribution of the SVM forecast error; (4) the forecasted probability distribution of error and the SVM forecast result are combined to estimate the distribution of wind generation. The proposed approach has good generalization capability by using the sparse learning mechanism. The range constraint that wind generation should be within [0, GN] can be systematically considered in the approach by applying D-S theory, which makes the forecast result more practical. Then, an ensemble method for probabilistic wind generation forecast is proposed. Instead of only selecting same individual distribution forecasts in traditional BMA, the proposed approach is able to get a better nonparametric prediction density by combining different individual distribution forecasts, such as Gaussian distribution, Weibull distribution and kernel density estimation function. Optimal model parameters are achieved by minimizing the distribution forecast accuracy indicator continuous ranked probability score. The ensemble model is so adaptive that we can achieve arbitrary distribution results by using this ensemble model based on sample data from different wind farms. Test results illustrate the efficiency of the methods.