Topology Conditioned Physics Informed Graph Surrogates for AC Optimal Power Flow Under Line Outages

Abstract

Security assessment requires repeated alternating current optimal power flow (AC OPF) solutions under changing injections and line outages. We propose a topology conditioned physics informed graph Fourier neural operator (GFNO PINO) that predicts voltage phasors and dispatch from network and operating data. Chebyshev filters adapt to each topology without eigenvector alignment. Output transformations enforce voltage and generation bounds and the reference angle; power balance, thermal limits, and angle differences remain soft constraints. Supervised initialization is followed by physics fine tuning with adaptive loss balancing. A topology disjoint benchmark covers four IEEE systems from 30 to 300 buses, three unseen outage topologies per system, and five training seeds. Physics fine tuning reduces mean sample maximum balance mismatch by 63.4--89.3\% relative to the data only GFNO\@. However, mismatch remains 0.281--5.959 p.u., with 0\% feasibility at the $10^{-3}$ p.u.\ threshold and no consistent advantage over topology blind or spatial models. CUDA inference medians are 8.3--138.1 times lower than stored IPOPT CPU solve medians, excluding verification and restoration. The model therefore requires independent feasibility checks and correction before operational use.