Toward a Unified Graph Learning Framework for Power Grids: Progress on Argonne’s GridFM Research

We present recent progress on Argonne’s GridFM, a framework for training heterogeneous graph neural network (GNN) models across multiple power grid applications using a multi-task learning approach. GridFM integrates multimodal datasets from five key grid applications (ACOPF, SCUC, distribution system state estimation, reliability assessment, and transient stability analysis) to enable shared and task-specific representations. We highlight model architecture design, data pipeline development, and preliminary results showing cross-task generalization and scalability. This work aims to build a unified foundation for GNN-based surrogates in power system analysis and control.