Sparse Heteroscedastic Gaussian Process for Short-term Wind Speed Forecasting

With increasing penetration levels of wind energy in power systems, wind speed forecasting becomes critical for the scheduling and planning of the grids. However, the accuracy of wind speed forecasting is highly variable due to the stochastic nature of wind, so probability prediction of future wind speed is important for assessing the uncertainties in the forecast results. This paper focuses on the short-term probability prediction of the wind speed. A sparse heteroscedastic Gaussian process forecasting model is formulated. With this model, we can provide predictive distributions that capture the heteroscedasticity of wind speed. This model employs l(1/2) regularization to reduce its computational costs, thus make it practical for large-scale wind speed forecast problems. The explanatory variables of the model are extracted from the historical wind speed observations at spatially correlated wind monitoring sites. Simulation results demonstrate the effectiveness of the proposed model.