Generating realistic synthetic power system data and training GNN models using GridFM-graphkit and GridFM-datakit

Alban Puech and Tamara Rosemary Govindasamy – both IBM Research

This tutorial introduces attendees to gridfm-datakit and gridfm-graphkit, two Python libraries developed by IBM, Hydro-Québec, and Stony Brook University, already adopted by several research groups for synthetic data generation and training AI models for power system analysis. These libraries consolidate scattered state-of-the-art practices in data generation and modeling for steady-state power system analysis, supporting tasks such as optimal power flow, power flow, and contingency analysis. By providing standardized datasets and model implementations, they overcome limitations in comparability, scalability, and the reliance on overly simplified grid settings.

Participants will be guided through hands-on notebooks to generate power flow and optimal power flow datasets using datakit, and to train and evaluate graph neural network–based solvers with graphkit, building models from scratch.