@aigclink: Microsoft open-sourced an AI foundation model for power systems: GridSFM, designed to accelerate research on AC optimal power flow in the power industry. GridSFM uses graph neural networks to approximate AC-OPF solving, treating the power grid as a graph, directly predicting near-optimal operating points, and then using them as warm-start initial values for traditional exact solvers to speed up convergence...

X AI KOLs Timeline Models
microsoft open-source power-system graph-neural-network optimal-power-flow ai-model foundation-model

Summary

Microsoft open-sourced GridSFM, an AI foundation model for power systems. It uses graph neural networks to approximate AC-OPF solving, is topology-agnostic, and can serve as a warm start for exact solvers achieving a 1.45x speedup, while also providing feasibility classification capabilities.

Microsoft recently open-sourced an AI foundation model for power systems: GridSFM, designed to accelerate research on AC optimal power flow in the power industry. GridSFM uses graph neural networks to approximate AC-OPF solving, treating the power grid as a graph. It directly predicts near-optimal operating points and then uses them as warm-start initial values for traditional exact solvers, accelerating convergence. A single model covers multiple power grids and is topology-agnostic. Previous approaches required training a separate model for each grid topology—changing the grid meant retraining. GridSFM is trained on dozens to hundreds of power grids of varying sizes and structures, achieving a single model that is topology-agnostic. This means that for new grids within the same structural range, the model can be used directly without needing to collect new data or retrain. Dual-head output: prediction + feasibility judgment—not only outputs the operating point but also determines whether the scenario is feasible. This is practical for large-scale scenario scanning, quickly filtering out infeasible scenarios and avoiding useless calls to the solver for impossible operating conditions. GridSFM outputs a full AC solution, including voltage magnitudes and angles, active/reactive power outputs, and active/reactive line power flows. This allows it to serve as a warm start for exact solvers, achieving a 1.45x speedup. A topology-agnostic AC-OPF neural network model with feasibility classification is essentially an AI accelerator for power system optimization. GridSFM comes in two versions: an Open version (open-source, approximately 15 million parameters) and a Premier version (not yet open-source). Application scenarios include power grid simulation and behavior study, rapid evaluation of large-scale configurations, etc. #GridSFM #AI电力模型
Original Article
View Cached Full Text

Cached at: 05/17/26, 03:28 AM

Microsoft has recently open-sourced an AI foundation model for power systems: GridSFM, designed to accelerate research on alternating current optimal power flow (AC-OPF) in the power industry.

GridSFM uses a graph neural network to approximate AC-OPF solutions, treating the power grid as a graph. It directly predicts near-optimal operating points, which are then used as warm-start initial values for traditional exact solvers, speeding up convergence.

A single model covers multiple power grids and is topology-agnostic.

Previously, each grid topology required training a separate model; switching to a different grid meant retraining from scratch. GridSFM is trained on tens to hundreds of power grids of varying sizes and structures, achieving cross-topology generality with a single model.

This means that within a structurally similar range, a new grid can be used directly without the need for new data collection or retraining.

Dual-head output: prediction + feasibility assessment – not only outputs operating points but also determines whether a scenario is feasible.

This is particularly useful in large-scale scenario scanning, quickly filtering out infeasible scenarios and avoiding unnecessary solver calls for impossible operating conditions.

GridSFM outputs a complete AC solution, including voltage magnitudes and angles, active/reactive power generation, active/reactive line power flows, etc. This enables it to serve as a warm start for exact solvers, achieving a 1.45x speedup.

An AC-OPF neural network model with cross-topology capability and feasibility classification essentially acts as an AI accelerator for power system optimization.

GridSFM comes in two versions: an open-source Open version with approximately 15 million parameters, and a Premier version that is not yet open-sourced.

Application scenarios include grid simulation and behavior studies, as well as rapid evaluation of large-scale configurations.

#GridSFM #AIPowerModel

Similar Articles