Brief

Researchers have developed a scalable heterogeneous graph neural network workflow, named HydraGNN, for optimal power flow (OPF) approximation in smart grids. This approach preserves the complex structure of power networks, including various node and edge types, and is designed for training on supercomputers. Experiments using millions of graph instances and distributed hyperparameter optimization on the ORNL Frontier supercomputer identified efficient models with approximately 1.6-1.7 million parameters. Fine-tuning these pre-trained graph foundation models demonstrated improved accuracy and stability for downstream tasks like feasibility classification and contingency regression, especially in low-data scenarios. AI