A physically-informed interpretable ensemble learning method for distributed photovoltaic power forecasting

• Physical-knowledge surface is constructed to enhance model interpretability. • A generalizable ensemble framework improves performance of diverse base models. • The integration of the SHAP method enables interpretation of the ensemble model. • Tests on 55 PV stations verify the method’s robust forecasting performance. Accurate forecasting of distributed photovoltaic (PV) power generation is vital for grid stability and efficient energy management. While recent advances in machine learning have improved PV power prediction, most models overlook physical mechanism features, resulting in black-box behavior with limited interpretability and only marginal gains in forecasting accuracy. To address these limitations, this paper proposes an interpretable ensemble learning framework augmented with physical information. A physically constrained surface for PV power generation is first constructed based on astronomical angles and irradiance-related features. This surface is incorporated into the deep learning training process by modifying the loss function, allowing physical priors to guide model learning and enhance interpretability. Building on this, an interpretable ensemble learning framework is developed by integrating a Kolmogorov–Arnold theory-based physical predictor with conventional machine learning models, leveraging the complementary strengths of these models to improve prediction accuracy and generalizability. To further enhance model transparency, the SHapley Additive exPlanations method is employed for feature attribution, providing insights into the contribution of input variables to the model output. Case studies conducted across 55 distributed PV stations demonstrate that incorporating physical-angle features enhances forecasting accuracy by 0.4%–1.37% on average. In data-scarce scenarios, models augmented with physical priors exhibit clear performance advantages. Overall, the proposed physically-informed ensemble model achieves superior generalization capability and consistently outperforms baseline methods such as LightGBM and MLP, yielding an average accuracy improvement of 0.38%–2.03%.