Research on Short-Term Probability Forecasting Methods for Photovoltaic Power Generation Based on Multi-Source Data Fusion and Feature Engineering

Abstract

The intermittent and uncertain nature of photovoltaic power generation poses significant challenges to grid stability. To enhance short-term forecasting accuracy and quantify uncertainty, this study proposes a probabilistic forecasting method for photovoltaic power generation that integrates multi-source data with feature engineering. Utilising operational data from a 50 kW actual photovoltaic power station combined with NASA meteorological records, a 29-dimensional feature system was constructed encompassing temporal characteristics, solar geometric features, physical statistical attributes, and meteorologically derived features. Employing a LightGBM quantile regression model, we achieve probabilistic forecasting of photovoltaic output for the next hour. Experimental results demonstrate excellent point prediction performance (RMSE = 7.323 kW, nRMSE = 0.085), with a P50 quantile loss of 2.4615 in probability forecasting, indicating effective uncertainty quantification capability. Feature importance analysis indicates that direct solar radiation, historical power lag terms, and solar elevation angle are key influencing factors. Although the P10–P90 forecast interval coverage (59.0%) reveals room for improvement in capturing extreme fluctuation events, this study provides a reliable data-driven solution for intelligent power system dispatch under high renewable energy penetration.