End-to-End Collaborative Optimization for Active Distribution Network Power Dispatch Based on Sparse Model-Ensemble Learning Policy

With the higher penetration of distributed renewable power sources, novel active distribution networks are increasingly implementing flexible adjustment strategies. Currently, the dual uncertainties from both sources and demand significantly affect power dispatch in distribution networks. Typically, power dispatch is performed using a predict-then-optimize approach, making it challenging to quantify the gap between the real-time and theoretically optimal dispatch performances due to inaccuracies in power predictions. Hence, this study introduces a novel end-to-end policy to solve a collaborative optimization between prediction and dispatch. The policy directly utilizes all available information, such as gridded numerical weather forecasts, for dispatch decision-making, which eliminates the need for power predictions as intermediate variables for dispatch. To address the challenges of high-dimensional and open-scenario model training in end-to-end policies, sparse model-ensemble learning is proposed to formulate the dispatch policy model. The model is solved using constrained policy optimization. Comparative studies show that the proposed end-to-end policy outperforms the predict-then-optimize policy in real-time dispatch cases involving photovoltaic reactive power ancillary service and demand response within distribution networks.