GridDiffuser: Constraint-Guided Graph Diffusion for AC Optimal Power Flow

Modern power systems are undergoing a significant transition from deterministic, steady-state dispatch to operation under high-dimensional uncertainty and rapid fluctuations driven by renewables, electrification, and topology perturbations. This shift requires Optimal Power Flow (OPF) to be solved frequently under diverse operating conditions. Classical AC-OPF solvers provide high physical fidelity but are slow for high-frequency scheduling, while linear approximations such as DC-OPF sacrifice accuracy for computational speed. To address these challenges, we propose GridDiffuser, a graph diffusion solver for AC-OPF. Unlike deterministic learning methods that predict a single solution and struggle to represent the non-convex landscape, GridDiffuser models a multi-modal distribution over operating points and generates multiple candidate solutions. Our method further applies an efficient Levenberg--Marquardt correction to the clean-space estimate, iteratively steering samples toward the constraint-feasible set. Experiments on thousand-bus grids with instance-level N-1 contingencies show that GridDiffuser substantially reduces pre- and post-power flow (PF) violations compared with deterministic baselines while generating solutions much faster than AC-IPOPT.