High dimensional very short-term solar power forecasting based on a data-driven heuristic method

Abstract

Improving the accuracy of solar power forecasting has become crucial for dealing with the negative effects of the integration of continually increasing solar power into power systems. This is a more challenging task when historical solar radiation data has not been recorded and no specific sky imaging equipment is available. This paper proposes a univariate data-driven method to improve the accuracy of very short-term electrical solar power forecasting. This approach includes defining new features that efficiently tackle the nonlinear characteristics of electrical solar power. An instance-based variable selection is also used to identify the best relevant features. Three state-of-the-art learning algorithms (i.e. neural networks, support vector regression, and random forest) have been used and compared as prediction algorithms of the proposed method. The effectiveness of the proposed approach is evaluated in a 15-min ahead prediction trial using a real solar power dataset and against three evaluation measures. The results show the proposed method significantly enhances the performance of very short-term solar power forecasting.