Towards GridFM: A Unified Graph Neural Solver Framework for Steady-State Grid Analysis

Thomas Brunschwiler, Etienne Vos, Alban Puech, Marcus Freitag – all IBM Research

At last year’s GridFM Workshop, we introduced an initial GridFM model together with supporting libraries for synthetic data generation and low-code model training (DataKit and GraphKit). Since then, we have substantially extended and hardened these assets to support the core steady-state grid analysis tasks, Power Flow (PF), Optimal Power Flow (OPF), and State Estimation (SE), within a single scalable framework.
In this talk, we present three major advances:
i) Unified Graph Neural Solver (uGNS)
We introduce the uGNS, a unified architecture composed of stacked Heterogeneous Graph Transformer layers combined with dedicated Constraint Enforcement Layers. These layers explicitly impose box constraints (e.g. generator and voltage limits). They also implicitly impose equality constraints (e.g., power balance equations) and inequality constraints (e.g., line loading limits). A physics-informed decoder ensures task-consistent outputs across PF, OPF, and SE, enabling a single model to generalize across traditionally separate problem classes.
ii) Task-Specific Benchmarking
We evaluate uGNS against classical AC and DC solvers (when applicable) as well as state-of-the-art learning-based approaches for PF, OPF, and SE. We report accuracy, runtime performance, and robustness metrics. In particular, we demonstrate strong performance in contingency analysis scenarios up to n-20, highlighting the model’s scalability and operational relevance.
iii) Synthetic Data Generation
We significantly extended the DataKit library to support large-scale synthetic grid data generation augmented with aggregated real-world load profiles, generator dispatch variations, and admittance perturbations. This establishes DataKit as a versatile foundation for training and benchmarking of GridFMs. Additionally, we introduce a multimodal disaggregation pipeline that produces bus-level datasets for the Texas grid, enabling realistic, fine-grained data sets. Together, these contributions push the frontiers of our GridFM vision for steady-state power system analysis, combining physics-aware modeling, scalability, and cross-task generalization within a single graph-based learning framework.