Physics-Informed Graph Neural Network for Dynamic Reconfiguration of Power Systems

TL;DR

GraPhyR addresses the challenge of dynamic reconfiguration in distribution grids by learning a physics-informed Graph Neural Network that co-optimizes topology and dispatch. The approach embeds switch behavior as gates, uses local predictors for scalable predictions, and employs a physics-informed rounding layer to produce feasible topologies, all while training in an unsupervised manner with a loss that enforces power-flow physics. Key contributions include gated message passing for switches, a scalable local prediction scheme, a PhyR rounding mechanism, and topology-aware input handling that generalizes across grid topologies. Empirical results on BW-33, G1, and TCP-94 show substantial speed-ups over traditional MIP solvers and competitive accuracy, with demonstrated adaptability to changing grid conditions and multiple topologies, indicating practical utility for real-time grid management.

Abstract

To maintain a reliable grid we need fast decision-making algorithms for complex problems like Dynamic Reconfiguration (DyR). DyR optimizes distribution grid switch settings in real-time to minimize grid losses and dispatches resources to supply loads with available generation. DyR is a mixed-integer problem and can be computationally intractable to solve for large grids and at fast timescales. We propose GraPhyR, a Physics-Informed Graph Neural Network (GNNs) framework tailored for DyR. We incorporate essential operational and connectivity constraints directly within the GNN framework and train it end-to-end. Our results show that GraPhyR is able to learn to optimize the DyR task.

Figures (5)

• Figure 1: GraPhyR: proposed framework to solve the DyR problem.
• Figure 2: Message passing layers where switches are denoted by red-dashed lines. The node and switch embeddings are represented by blue and red colored blocks respectively, where the number of squares per-block indicates the dimension of the embeddings $h$.
• Figure 3: Local predictions made by the switch and line predictors use the node and switch embeddings extracted after $\mathcal{L}$ message passing layers.
• Figure 4: Grid topology of BW-33 (left) and the synthetic $\mathcal{G}_1$ (right). Switches indicated with green dashed lines. Solar generator locations indicated with yellow nodes.
• Figure 5: Magnitude of the inequality constraint violations for GraPhyR. The constraint sets on nodal generation ($\mathbf{p^G} \text{ and } \mathbf{q^G}$, as in Eq. \ref{['eq:var_lims']}), voltage limits ($\mathbf{v}$, as in Eq. \ref{['eq:v_pcc']}), and connectivity constraints (Eq. \ref{['eq:connectivity']}) are separated by black vertical lines.