IBM Research, alongside partners including Brookhaven National Laboratory and Hydro-Québec, has released GENCO, an open-source geometric neural corrective solver for steady-state grid analysis. The release is part of the OpenGridFM project hosted by Linux Foundation Energy (LF Energy) and includes the GridFM Development Framework to facilitate building and benchmarking neural grid solvers.
- GENCO unifies power flow, optimal power flow, and state estimation tasks into a single model using a heterogeneous graph transformer backbone.
- The solver delivers speedups up to 30 times faster than Newton-Raphson AC solvers for power flow and up to 85 times faster than interior-point solvers for optimal power flow.
- GENCO maintains accuracy comparable to DC approximations while recovering full grid variables like voltage magnitudes, with median residuals about 3 times lower in high-order contingency scenarios.
- The GridFM Development Framework provides open datasets of 4 million instances, data generation tools, and low-code environments for training and benchmarking.
- The model was validated on real-world SCADA data from Hydro-Québec’s 1,200-bus transmission network over a full year.
This release aims to provide grid operators and planners with a tool capable of evaluating thousands or millions of chronological power-flow cases at unprecedented speed, addressing the computational bottlenecks caused by increasing renewable integration and AI-driven load.