An ensemble prediction intervals approach for short-term PV power forecasting

Prediction intervals (PIs) estimation is a powerful statistical tool used for quantifying the uncertainty of PV power generation in power systems. The lower upper bound estimation (LUBE) approach, when combined with extreme learning machines (ELM), is effective for constructing PIs. ELM is an efficient but unstable machine-learning method in generating reliable and informative PIs. To overcome this instability, a novel ensemble approach based on ELM and LUBE (ELUBE) is proposed for short-term PV power forecasting. To optimize quality of PIs, the sigmoid, radial basis and sine functions are used to train three groups of ELUBE models, and the models with higher performance are selected as ensemble members. Furthermore, a weighted average method is developed to aggregate the selected individuals. An improved differential evolution algorithm is used to perform the search for the optimal combination weight values of PIs. The feasibility and effectiveness of the proposed approach are evaluated by using PV datasets, obtained from a lab-scale DC micro-grid system.