Day-Ahead Photovoltaic Power Forecasting Using Multi-Source Data Fusion

This study proposes a day-ahead forecasting framework for photovoltaic power generation, integrating numerical weather prediction and deep learning to enhance prediction accuracy and optimize grid scheduling. Using 15-minute resolution data from August 2018 to June 2019, a comprehensive model was developed to analyze PV power characteristics, extracting seasonal and diurnal patterns via solar irradiance theory and Fast Fourier Transform. An Attention-based Long Short-Term Memory model, combined with NWP data, predicts 24-hour power curves with a mean absolute error of 5.2 kW. Further enhancements using Convolutional Neural Networks and Bidirectional LSTM, alongside spatial downscaling with XGBoost and Kriging interpolation, improve accuracy in complex terrains, reducing wind speed prediction MAE by 69.93%. The framework performs robustly across sunny, cloudy, and extreme weather scenarios, offering reliable support for grid operations and renewable energy integration.