An ensemble machine learning based approach for constructing probabilistic PV generation forecasting

Abstract

Photovoltaic (PV) generation forecasting plays an important role in accommodating more distributed PV sites into power systems. However, due to the stochastic nature of PV generation, conventional point forecast methods can hardly quantify the uncertainties of PV generation. Being capable of quantifying uncertainties, probabilistic forecasting tools, like prediction intervals (PIs), are receiving increasing attention. This paper proposes a new framework to construct PIs and make point forecasts. In the proposed framework, an efficient and robust algorithm is employed to perform quantile regression. Based on the quantile regression results, PIs for multiple confidence levels are constructed utilizing different quantiles. Simulation results on a PV generation system reveal that the proposed framework is more reliable and accurate, compared with state-of-the-art methods, as measured by multiple performance indices.