Shared Foundations, Better Outcomes: A Collaborative Roadmap for Grid Foundation Models

Hendrik F. Hamann – Brookhaven National Laboratory and Stony Brook University

Grid Foundation Models (GridFM) are an emerging capability in which large foundation models learn from multi-modal, multi-utility power-system data to support a broad range of applications for operating, managing, and planning the electric grid. Beyond technical advances, GridFM also represents a new collaboration paradigm: stakeholders jointly develop shared foundation models and then fine-tune them for utility- and use-case–specific needs, improving outcomes while reducing duplicated effort.

In this presentation, we outline a roadmap for collaborative GridFM development across three pillars: (i) AI-ready grid data, (ii) foundation-model architectures, and (iii) enabled applications. For data, we highlight priorities such as high-quality time-series and distribution-system measurements, realistic and diverse load profiles, integration of alternative data sources (e.g., assets, events, weather), and ensuring real-world representativeness for training and evaluation. For model development, we discuss design choices and pretraining strategies that govern generalizability, scalability, accuracy, and performance, and how these tradeoffs shape deployable GridFM capabilities. Finally, we describe how a shared GridFM foundation can unlock a portfolio of differentiated applications—moving beyond today’s point solutions toward more robust, transferable, and decision-relevant AI for the grid.