A Photovoltaic Power Forecasting Method Integrating Physical Mechanisms and Deep Learning

With the rapid expansion of distributed photovoltaic installations, the intermittent and fluctuating nature of their output power poses challenges to grid dispatch and operational stability. Physical models offer good interpretability but suffer from limited accuracy under complex meteorological conditions. Pure data-driven methods lack sufficient generalization under extreme operating conditions, and existing hybrid approaches often employ static weights that struggle to adapt to environmental changes. To address these issues, this paper proposes a photovoltaic power forecasting method that integrates physical mechanisms with multi-scale deep learning. The key innovations of this method include: (1) A non-uniform error compensation strategy based on irradiance interval segmentation to enhance the accuracy of the physical branch across different light intensity ranges; (2) Construction of a 37-dimensional physically enhanced feature system incorporating meteorological, temporal, physically derived, and statistical interaction features, utilizing a parallel Multi-scale CNN (kernel lengths 3/7/15) combined with BiLSTM architecture to extract multi-scale temporal characteristics; (3) A dynamic weighted fusion mechanism based on four-dimensional confidence levels (irradiance, temperature, time, and weather stability) to achieve complementary environmental perception between physical and data-driven models. Validation using 15-minute resolution annual measurement data from a prefecture-level city photovoltaic power station demonstrates significant improvements over pure physical and pure neural network approaches in metrics including MAE, RMSE, and R$^2$ (example results: MAE = 3.571 kW, RMSE = 6.384 kW, R$^2$= 0.9781). Ablation analysis further validated the effectiveness of each module. By balancing physical consistency with data-driven expressiveness, this approach provides practical value for enhancing the accuracy of photovoltaic power forecasting and improving engineering applicability.