The physical-encoded Photovoltaic forecasting method combined with continuous learning and multi-digital twins mechanisms

End-to-end neural network models, often seen as black boxes, have been widely used in photovoltaic (PV) power forecasting. However, they face challenges regarding poor model adaptability, transferability, and interpretability. To address these issues, this paper proposes a physical-encoded PV forecasting model, which decomposes the end-to-end network into a data-driven external parameter forecasting model and a physics-driven power calculation model. The power calculation model, with explicit physical meanings, enhances the model's interpretability. A continual learning mechanism is designed to enable the model to quickly adapt to environmental changes, mitigating the impact of model drift and improving adaptability and transferability. A multi-digital twins synergistic operation mechanism is introduced to incorporate the strengths of other models, further enhancing forecasting accuracy. Model drift can be categorized into concept drift and data drift. This paper designs two scenario experiments to test these drifts. Scenario 1 focuses on concept drift, and the experimental results show that the proposed method in this paper achieves improvements of 30.5 %, 16.5 %, and 1.9 % in the nMAE, nRMSE, and R² metrics, respectively, compared to the best results of the comparison models. In Scenario 2, the model is transferred to other power plants for data drift tests. Results show that when transferred to Plant 4, its accuracy improves by 45.8 %, 21 %, and 2.1 % compared to the best comparison method; for Plant 5, the improvements are 34.1 %, 18.3 %, and 2.5 %.