PROBABILISTIC FORECASTING OF PHOTOVOLTAIC POWER OUTPUT USING RVM-QR

PROBABILISTIC FORECASTING OF PHOTOVOLTAIC POWER OUTPUT USING RVM-QR

This study addresses the significant challenges in forecasting the photovoltaic (PV) power output under non-clear sky conditions, characterized by high randomness and volatility which traditionally result in low prediction accuracy. We propose a probabilistic forecasting method for photovoltaic power output that integrates Relevance Vector Machines (RVM) with quantile regression (QR). By constructing a hybrid kernel function that combines polynomial and Gaussian kernels, and optimizing the kernel parameters and weights, our model adapts to different weather conditions, thereby improving the accuracy of point forecasts for PV power output under non-clear skies. Quantile regression is utilized to calculate the power fluctuation intervals at different confidence levels, quantitatively capturing the uncertainty and variability of photovoltaic power output. Experiments conducted using real data from the Ningxia Sun Mountain Photovoltaic Station, and comparative analyses with other forecasting methods, reveal that our approach achieves the highest accuracy in probabilistic forecasting of PV power output under non-clear sky conditions, with prediction errors ranging between 2.62% and 4.35%. Furthermore, the forecasting model demonstrates excellent interpretability.