Day-Ahead Photovoltaic Power Forecasting Based on Multi-Source Data Fusion and the SARIMA-LSTM Hybrid Model

To address the high variability and uncertainty of photovoltaic (PV) power output caused by solar irradiance, ambient temperature, and cloud cover, this study proposes a day-ahead forecasting model based on a SARIMA-LSTM hybrid framework with multi-source data fusion. SARIMA captures the linear components of historical power data, while LSTM models nonlinear fluctuations. The model integrates 48-hour numerical weather prediction (NWP) data to improve responsiveness to abrupt weather changes. To enhance spatial resolution, a spatial downscaling technique is applied to refine NWP inputs from 1 km to 100 m, allowing better representation of local terrain and microclimates. Experimental results show that the model reduces the root mean square error (RMSE) during daytime from 0.098 kW/kWp (ARIMA baseline) to 0.080 kW/kWp, achieving an 18% improvement. The forecasting accuracy reaches 91.62%, with a qualification rate of 98.15%. Incorporating downscaled NWP data further reduces prediction error by 36.4%, with improvements up to 40% in mountainous regions. The model demonstrates strong stability and generalization, making it suitable for fine-grained PV forecasting across diverse geographic and climatic conditions.