Scalable Heterogeneous Graph Foundation Models for Data-Driven Optimal Power Flow in Smart Grids

This article presents a scalable workflow for heterogeneous graph neural networks (GNN) aimed at data-driven optimal power flow (OPF) surrogate modeling and the development of foundational models. Leveraging HydraGNN, the workflow maintains the unique types of nodes and edges found in power grids—such as buses, generators, loads, shunts, AC lines, transformers, and device-to-bus connections—and facilitates distributed preprocessing, training, hyperparameter optimization (HPO), and subsequent fine-tuning on top-tier supercomputers. Utilizing three million heterogeneous graph instances across ten PGLib-OPF scenarios (ranging from 14 to 13,659 buses), HPO driven by DeepHyper was performed on the ORNL supercomputer. This methodology overcomes the challenges faced by current learning-based surrogates that simplify network structures or are constrained by limited topologies and scalability for graph model training, aiming to provide quick and dependable OPF approximations for smart-grid operations.