Probability density forecasting of wind power using quantile regression neural network and kernel density estimation

Owing to the increasingly serious social energy crisis nowadays, wind power and other renewable energy are paid more attention. However, penetration of wind power prominently enhances the degree of complexity and difficulty in planning and dispatching of electric power systems. High-precision and more-information short term wind power forecasting (STWPF) results can effectively alleviate the uncertainly of wind power and balance the electrical power. Kernel function and bandwidth selection method have significant impact on the results of STWPF. A hybrid wind power probability density prediction method based on quantile regression neural network and Epanechnikov kernel function using Unbiased cross-validation (QRNNE-UCV) is presented. The wind power predicting results at different conditional quantiles are used as the input of kernel density estimation (KDE), which is capable of estimating the comprehensive wind power probability density forecasting information at any time in the future. In order to evaluate the wind power prediction results, the paper constructs two evaluation criteria, including evaluation metrics of point prediction results and evaluation metrics of prediction interval (PI). As a point prediction result, the probability mean is first constructed in the paper. Two real datasets of wind power from Ontario, Canada, are used to verify the QRNNE-UCV method. Moreover, by comparing with the probability density results at various confidence levels, the influence of confidence level on STWPF is investigated in this article. Experiment results show that the QRNNE-UCV method can construct more accurate PI and probability density curves, and the calculated probability mean is superior to the other point predictions. Meanwhile, the quality of PICP and PINAW improves with the increase of confidence level. The above prediction results have the ability to validly quantify the indeterminacy of wind power generation in contrast to existing support vector quantile regression (SVQR) and quantile regression neural network and triangle kernel function (QRNNT) probability density forecasting methods.