Physics-guided residual learning for day-ahead forecasting and scheduling in active distribution networks

Abstract

With the high penetration of distributed photovoltaic (PV) generation and wind power in active distribution networks, dayahead scheduling has become increasingly dependent on accurate source-load forecasting. Traditional mechanism-based models offer a certain degree of physical interpretability, yet they are often inadequate for capturing random fluctuations. In contrast, purely data-driven models possess strong nonlinear fitting capability, but may lack physical plausibility and engineering reliability. To address this issue, this paper proposes a physics-guided residual learning method for day-ahead scheduling in active distribution networks. Specifically, baseline forecasting models for PV generation, wind power, and load are first constructed based on irradiance, wind speed, and historical load patterns to provide trend priors. Then, a BP neural network is employed to learn the residual errors of the mechanism-based models and refine the baseline forecasts. Finally, the forecasting results are embedded into a day-ahead scheduling model with energy storage systems and demand response, and the proposed method is comprehensively evaluated from both forecasting and scheduling perspectives. Experimental results under three simulated test scenarios show that the proposed method achieves relatively low forecasting error for PV prediction and yields the lowest scheduling cost in all scenarios. Compared with the conventional mechanism-based method and the pure BP method, the average operating cost is reduced by approximately 15.50% and 4.70%, respectively. These results indicate that combining physical priors with residual learning can provide more effective forecasting inputs for source-load-storage coordinated scheduling in active distribution networks.