Robust functional regression for wind speed forecasting based on Sparse Bayesian learning

Accurate wind speed forecasting is helpful for reducing the instantaneous fluctuation of voltage, and also has great practical significance on power dispatching and plan. There are two problems in wind speed forecasting: loss of details when transforming the high-resolution data into the low-resolution data and outliers existing in our data. So, a sparse Bayesian-based robust functional regression model is proposed in this paper. First, both the low-resolution and high-resolution data are considered as inputs to forecast future wind speed. Specifically, besides the historical 10-min mean wind speed, the corresponding functional variables, constructed by wind speed data recorded every 5 s in each 10-min interval, are also taken as inputs to make 10-min-ahead wind speed forecasting. But, not all functional variables contribute to the accurate wind speed forecasts. So, a multi-Laplace prior is given to the corresponding coefficient functions to get sparse solutions, which can reduce the adverse effects of the redundant functional variables on the final forecasting results. Second, a multi-mixture of Gaussians prior is assumed for the forecasting error to enhance the robustness of the forecasting model. Results of spatial-temporal and multi-step-ahead wind speed forecasting show that the proposed model provides more accurate forecasts than the compared models.