Grid foundational models emerging
Image: IBM
AI-trained grid foundational models – GridFMs – are expected to help improve power grid operations, planning and control for the 21st century and beyond.
In a new project being led by IBM and including Canadian utility Hydro-Québec, such grid foundational models are being developed on an open source basis under LF Energy with the aim to address challenges emerging with the increasing complexity of grid transformation.
Foundation models are large AI models pre-trained on large data sets and adapted to a broad set of applications, with a key benefit being the ability for stakeholders to fine-tune a pre-trained model for specific applications using their own proprietary data in a scalable and economical way.
In a new paper in the publication Joule, IBM, which has worked with NASA on foundational models on the weather and climate among other areas, and partners conceptualise grid foundational models trained on power grid data providing a significant speed-up in computation of at least 3-4 orders of magnitude.
Have you read?
Winning strategies for risk mitigation for utilities in the GenAI era
Project ECLIPSE aims to transform Europe’s energy landscape
With this downstream modelling tasks could include contingency analysis, outage prediction, load forecasting, renewable forecasting, system adequacy, dynamic optimal power flow, system security, disaster recovery and dynamic state estimations.
In the paper the researchers also suggest a road map for GridFM starting with pre-training on more than 300,000 solved optimal power flow problems on grids of various sizes towards a first version of GridFM – GridFM-v0 – expected to be available in the second quarter of 2025.
“Foundation model technologies are a great fit for tackling the underlying complexity of the power systems,” said Juan Bernabé-Moreno, Climate and Sustainability strategy lead at IBM Research.
He adds that he has been working for many years in applying machine learning to drive the energy transition and for the first time, sees a fundamental step change in how AI can address these challenges.
“These include not just the integration of renewable sources but also supply security, electrification and more. GridFMs can capture the dependencies across all the data we find in modern grids in an AI representation and offer new possibilities.”
The work on GridFMs was initiated by IBM and Imperial College and joined LF Energy with its support for the open-source development of a common GridFM technology base.
With this it is envisaged that the power system and AI communities can collaborate to develop and harness emerging AI capabilities for the power grid.
Initially, the project is focussed on establishing a large collection of solved AC power flow for different grid topologies, parameters and load conditions.
In the second phase, suitable architectures are due to be evaluated by pre-training performance analysis and by adjusting models, training strategies and loss functions, with GridFM-v0 pre-trained to reconstruct masked power flow data.
Hydro-Québec plans to innovate on top of the open source model by validating and fine tuning to the utility’s specific downstream applications.
Other collaborators that have contributed to the work include ETH Zurich, Argonne National Laboratory, UK utility SSEN Transmission and a Swiss electricity system operator.
