Sparse online warped Gaussian process for wind power probabilistic forecasting

Abstract

Wind generation has experienced rapid growth around the world in the past decade. This highlights the importance of the short-term wind power forecasting. This paper focuses on the probabilistic short-term wind power forecasting. An online sparse Bayesian model is established. The key features of the proposed model are its non-Gaussian predictive distributions and its time-adaptiveness. This model based on the warped Gaussian process (WGP), which handles the non-Gaussian uncertainties in wind power series by automatically transforming it to a latent series. The transformed series is well-modeled by a Gaussian process (GP), then the non-Gaussian uncertainty associated with the wind power can be predicted in a standard GP framework. Wind generation is a process whose characteristics change with time, so a wind power forecasting model should exhibit adaptive features. To address this, we introduce an online learning algorithm to WGP, thus permitting WGP to track the time-varying characteristic of wind generation. Moreover, since the high computational costs of WGP hinder its practical application on large-scale problems such as wind power forecast, the proposed model also employs a sparsification method to reduce its computational costs, thus enhancing its practical applicability. The simulation on actual data validates the effectiveness of the proposed model. The data used in the simulation are obtained in the real operation of a wind farm in China.