Reproduce “CANOS: A Fast and Scalable Neural AC-OPF Solver Robust To N-1 Perturbations” on OpenResearch

About this paper

Luis Piloto, Sofia Liguori, Sephora Madjiheurem, Miha Zgubic, Sean Lovett, Hamish Tomlinson, Sophie Elster, Chris Apps and 1 more

Google DeepMind researchers developed CANOS, a Graph Neural Network (GNN)-based solver for Alternating Current Optimal Power Flow (AC-OPF) that achieves near-optimal and highly feasible solutions with processing times significantly faster than traditional approximations. The solver robustly handles N-1 topological perturbations and scales to 10,000-bus grids, predicting AC-OPF costs within 1% of the ground truth while running 20 to 750 times faster than DC-IPOPT.

Key insights

• Graph Neural Networks effectively learn complex AC-OPF relationships and inherently adapt to varying grid topologies, including N-1 perturbations, without requiring retraining for each new configuration.
• Incorporating constraint violation degrees into the loss function, alongside supervised L2 loss, guides the GNN to produce solutions that are both accurate and highly feasible.
• Raw GNN outputs serve as excellent initial points for power flow post-processing, enabling high convergence rates and ensuring bus power balance while maintaining overall solution speed.

Summary by alphaXiv.