Extreme Probabilistic Solar Power Prediction via Localized Sample Structure Recognition and Generalized Error Estimation

The fluctuations and uncertainty of solar power constantly threaten the secure operation and economic dispatch of power systems. Existing end-to-end point or probabilistic solar power prediction methods mostly lack effective integration of the two approaches, and the latent error caused by machine learning (ML) techniques is rarely taken into consideration. Hence in this paper, a combined extreme probabilistic solar power prediction (EPSPP) scheme is proposed, by integrating point forecasting with extreme error estimation. Firstly, the localized sample structure recognition (LSSR) is conducted to determine the neighborhood of meteorological conditions, where feature weights of Euclidean distance measurement are allocated with respect to the valid mutual information (MI) derived by two-dimensional diffusion kernel density estimation (2D-DKDE). Secondly, with the neighborhood generated by LSSR, an improved localized generalization error estimation (ILGEE) algorithm is put forward to infer the real-time maximal second-order origin moment of solar power point forecasting error corresponding to designated confidence levels. Finally, the solar power at each temporal moment is deduced as distinct Gaussian distributions, by modifying the mean value and variance according to statistical principles. For the sake of the so-called “extreme”, the proposed scheme could maintain reliability even under circumstances of the worst ML model precision. Cases from a real-world solar power station in Oregon, USA, are used to validate its effectiveness.