Distribution network optimization based on topology security-constrained integrated reinforcement learning

With the increasing penetration of large-scale renewable energy sources into the power grid, distribution networks are facing significant challenges, including intensified voltage fluctuations and increased network losses. Although deep reinforcement learning has made considerable advancements in addressing optimization problems compared to traditional algorithms, there has been limited focus on enhancing convergence and safety in cooperative optimization scenarios, particularly those involving topological reconstruction. To overcome these challenges, this paper proposes a distribution network optimization model that incorporates topological security-constrained integrated reinforcement learning. The model improves the encoding of topologies by representing them in a multi-dimensional discrete space and introduces a topological masking mechanism to achieve high safety and computational efficiency. Additionally, an ensemble strategy is utilized to develop an action network group, improving action prediction and screening, thereby achieving better training stability. Experiments conducted on an enhanced IEEE 33-node distribution network system indicate that the proposed improvements significantly enhance training stability and support the safe and efficient operation of the system.