應用研究／AI 基礎設施
GENCO Uses a Single Graph Neural Network for Three Power-Grid Solvers, Boosting Large-Scale AC Power-Flow Throughput by About 29×
Through heterogeneous graph attention, physical-variable decoding, and iterative residual correction, GENCO unifies power flow, optimal power flow, and state estimation. Batched inference on large grids is substantially faster than conventional AC solvers, but its zero-shot cross-grid and real-world SCADA accuracy remains insufficient to directly replace existing tools.
Teams from IBM, ETH Zurich, and other institutions have introduced GENCO (GEometric Neural Corrective Optimizer) and the GridFM development framework, aiming to handle power flow (PF), optimal power flow (OPF), and state estimation (SE) with a single neural architecture. The model represents buses, generators, and transmission lines as a heterogeneous graph. Graph attention propagates messages according to different relationship types, while line admittance and thermal capacity constraints also modulate attention. In each correction round, the model first predicts voltage magnitudes, phase angles, and generation output. A physics decoder then reconstructs power flows and feeds power-balance and constraint-violation residuals into the next round, instead of directly producing an unverifiable black-box answer.
The performance figures need to be interpreted in light of the test conditions. The researchers ran batched inference on an H100 SXM5, while conventional solvers used a same-generation 84-core AMD EPYC 9634. Both measurements excluded loading, compilation, and warm-up, and each side used its optimal batch size or worker count. On PF cases with 2,000 and 10,000 buses, GENCO Tiny was 28.2× and 29.1× faster, respectively, than Newton–Raphson AC-PF, while also producing voltage and reactive-power outputs unavailable from DC-PF. Its speed, however, was only about half that of DC-PF. On the 2,000-bus OPF case, GENCO Small was 84.5× faster than AC-OPF.
The supporting tooling also adds engineering value. The Apache-2.0-licensed `gridfm-graphkit` already provides YAML-based training, fine-tuning, evaluation, and inference CLIs, while `gridfm-datakit` can generate data containing N-k topology variations and perturbations to admittance and generation costs. The datasets are available on Hugging Face. However, the paper reports a zero-shot PF residual of 11.07 MW on unseen grids, worse than DC-PF's 2.30 MW. On Hydro-Québec SCADA data, the model also required about 15,000 fine-tuning samples to approach DC-PF accuracy. Paper-specific configurations, scripts, and complete models are still several weeks away. The next question is whether third parties can reproduce the reported throughput, tail-end constraint-violation rates, and cross-topology generalization—not merely the average error.