GridSFM: A Foundation Model for Solving AC Optimal Power Flow
GridSFM is presented as a 15 million-parameter graph model that can warm-start AC-OPF solvers on grids as large as 10,000 buses.
The paper says the model was pretrained across 54 grid topologies ranging from 500 to 4,000 buses. On a held-out 10,000-bus case, it reports a 2.45% zero-shot generation-cost error without degradation as system size grows. With physics-informed fine-tuning using only 100 solved instances, the authors say it adapts to unseen grids up to 10,000 buses. They also claim the approach beats single-topology neural models on cost and solver iterations, and release the models, data, and code. ArXiv · AI/CL/LG's note
The paper says the model was pretrained across 54 grid topologies ranging from 500 to 4,000 buses. On a held-out 10,000-bus case, it reports a 2.45% zero-shot generation-cost error without degradation as system size grows. With physics-informed fine-tuning using only 100 solved instances, the authors say it adapts to unseen grids up to 10,000 buses. They also claim the approach beats single-topology neural models on cost and solver iterations, and release the models, data, and code. ArXiv · AI/CL/LG's note
score 5