Randomised learning-based hybrid ensemble model for probabilistic forecasting of PV power generation

Probabilistic forecasting of solar photovoltaic (PV) generation is critical for stochastic or robust optimisation-based power system dispatch. This study proposes a randomised learning-based hybrid ensemble (RLHE) model to construct the prediction intervals of probabilistic PV forecasting. Three different randomised learning algorithms, namely extreme learning machine, randomised vector functional link, and stochastic configuration network, are ensembled as a hybrid forecasting model. Besides, bootstrap is used as the ensemble learning framework to increase the diversity of training samples. For each algorithm, a decision-making rule is designed to evaluate the credibility of the individual outputs and the incredible ones are discarded at the output aggregation step. The weight coefficients of the aggregated outputs of the three algorithms are then optimised to compute the final point forecast results. Based on the point forecast results, the prediction intervals are constructed considering both model misspecification uncertainty and data noise uncertainty. The variance in model misspecification uncertainty is directly calculated with the individual outputs and the variance in data noise uncertainty is separately trained with an RLHE model. The proposed method is tested with an open dataset and compared with several benchmarking approaches.