GridSFM
An AI foundation model for power grids introduced by Microsoft Research. It treats the electrical grid as a graph structure and is designed to solve the AC Optimal Power Flow (ACOPF) problem. The model is topology-agnostic and supports both zero-shot and fine-tuned deployment.
Timeline (last 90 days)
Microsoft Research has introduced GridSFM, aimed at simultaneously addressing two challenges in power systems research: limited data access and slow solving speed.
Microsoft has built an open data pipeline that uses only publicly available data sources to reconstruct physically consistent power grid models for any region of the United States, with the process divided into five steps.
GridSFM models the power grid as a graph, treating buses as vertices and transmission lines as edges, using an attention module to handle grid structure, and possesses topology-agnostic capability.
GridSFM trains a single model on more than 150 base grid topologies and approximately 500,000 scenarios, in a training approach similar to ImageNet pretraining.
About half of GridSFM's training data contains infeasible scenarios, treating infeasibility as a first-class output the model needs to learn to detect rather than discarding it as noise.
GridSFM returns four physical snapshots and a feasibility judgment with confidence in a single forward pass, running approximately one thousand times faster than an AC solver.
GridSFM can triage contingency scenarios for feasibility with over 95% accuracy, routing only borderline cases to the solver.
When GridSFM serves as a warm start to provide an initial point for traditional solvers, it achieves an average speedup of approximately 1.7×, with speedups exceeding 7× on some larger meshed grids.
In zero-shot testing on an unseen grid of 6,470 buses (approximately 40% larger than any grid in training), the cost error increases by about 10%, but the model can still rank scenarios by cost with near-perfect accuracy.
Fine-tuning GridSFM with just 10 scenarios can reduce cost error to below 2% and bring feasibility detection above 90%; using 1,000 scenarios can reach in-sample accuracy levels.
2 more timeline events
All Facts (12)
GridSFM's data, models, and code are all openly released, including open data, open code, and open weights.
50%UnverifiedMicrosoft has built an open data pipeline that uses only publicly available data sources to reconstruct physically consistent power grid models for any region of the United States, with the process divided into five steps.
50%UnverifiedGridSFM models the power grid as a graph, treating buses as vertices and transmission lines as edges, using an attention module to handle grid structure, and possesses topology-agnostic capability.
50%UnverifiedGridSFM trains a single model on more than 150 base grid topologies and approximately 500,000 scenarios, in a training approach similar to ImageNet pretraining.
50%UnverifiedAbout half of GridSFM's training data contains infeasible scenarios, treating infeasibility as a first-class output the model needs to learn to detect rather than discarding it as noise.
50%UnverifiedGridSFM returns four physical snapshots and a feasibility judgment with confidence in a single forward pass, running approximately one thousand times faster than an AC solver.
50%UnverifiedGridSFM can triage contingency scenarios for feasibility with over 95% accuracy, routing only borderline cases to the solver.
50%UnverifiedWhen GridSFM serves as a warm start to provide an initial point for traditional solvers, it achieves an average speedup of approximately 1.7×, with speedups exceeding 7× on some larger meshed grids.
50%UnverifiedIn zero-shot testing on an unseen grid of 6,470 buses (approximately 40% larger than any grid in training), the cost error increases by about 10%, but the model can still rank scenarios by cost with near-perfect accuracy.
50%UnverifiedFine-tuning GridSFM with just 10 scenarios can reduce cost error to below 2% and bring feasibility detection above 90%; using 1,000 scenarios can reach in-sample accuracy levels.
50%UnverifiedGridSFM incorporates Kirchhoff's law violation penalty terms and thermal limit penalty terms into the loss function, constraining the model by the laws of physics while pursuing prediction accuracy.
50%UnverifiedMicrosoft Research has introduced GridSFM, aimed at simultaneously addressing two challenges in power systems research: limited data access and slow solving speed.
50%