A Novel Graph Reinforcement Learning-Based Approach for Dynamic Reconfiguration of Active Distribution Networks with Integrated Renewable Energy

The dynamic reconfiguration of active distribution networks (ADNDR) essentially belongs to a complex high-dimensional mixed-integer nonlinear stochastic optimization problem. Traditional mathematical optimization algorithms tend to encounter issues like slow computational speed and difficulties in solving large-scale models, while heuristic algorithms are prone to fall into local optima. Furthermore, few scholars in the existing research on distribution network (DN) reconfiguration have considered the graph structure information, resulting in the loss of critical topological information and limiting the effect of optimization. Therefore, this paper proposes an ADNDR approach based on the graph convolutional network deep deterministic policy gradient (GCNDDPG). Firstly, a nonlinear stochastic optimization mathematical model for the ADNDR is constructed, taking into account the uncertainty of sources and loads. Secondly, a loop-based encoding method is employed to reduce the action space and complexity of the ADNDR. Then, based on graph theory, the DN structure is transformed into a dynamic network graph model, and a GCNDDPG-based ADNDR approach is proposed for the solution. In this method, graph convolutional networks are used to extract features from the graph structure information, and the state of the DN, and the deep deterministic policy gradient is utilized to optimize the ADNDR decision-making process to achieve the safe, stable, and economic operation of the DN. Finally, the effectiveness of the proposed approach is verified on an improved IEEE 33-bus power system. The simulation results demonstrate that the method can effectively enhance the economy and stability of the DN, thus validating the effectiveness of the proposed approach.