SiPhyR: An End-to-End Learning-Based Optimization Framework for Dynamic Grid Reconfiguration

Distribution grids are rapidly transforming with increasing penetration of distributed renewable generation, storage, and electric vehicles. These devices introduce new dynamic signatures including intermittency and changing demand patterns. To maintain a safe and reliable power grid, new operating paradigms are required for fast and accurate decision making. It is in this regard that we leverage machine learning for grid operations. We propose SiPhyR, a physics-informed machine learning framework that accomplishes end-to-end learning-based optimization for distribution grid reconfiguration. The reconfiguration problem optimizes the topology of the grid and power flows from distributed devices to reduce line losses, improve voltage profiles, and increase renewable energy utilization. To address the computational complexities of NP-hardness of binary decision variables, we propose a physics-informed rounding approach that explicitly embeds discrete decisions into an end-to-end differentiable framework. This enables Grid-SiPhyR to learn to simultaneously optimize grid topology and generator dispatch with certified satisfiability of safety-critical constraints. Our results are shown on three canonical distribution grids.