A Graph Diffusion Model For Active Distribution Network Reconfiguration

Distribution network reconfiguration holds potential for the integration of distributed energy resources and supporting economic operation. However, it is a combinatorial optimization problem with numerous integer variables, making fast and accurate solutions challenging. This paper proposes a graph diffusion model for distribution network reconfiguration. By integrating physical feasibility constraints directly into the reverse denoising process, the proposed method is able to generate only feasible radial topologies without relying on post-generation filtering. The diffusion model further enables global exploration of the combinatorial space and produces diverse candidate topologies with low online computational cost. Based on the generated candidates, power flow optimization determines the topology and corresponding power flow distribution. Numerical studies on multiple benchmark systems demonstrate that the proposed model achieves a superior trade-off between solution optimality and computational efficiency compared with existing learning based methods.