An IBM-led team is exploring how AI can prepare the electrical grid for the low-carbon era

At the start of the new millennium, the National Academy of Engineering named the power grid the 20th century’s greatest engineering achievement. A quarter-century later, this backbone of modern civilization is showing its age.

Electrification still makes our food, health care, communication, and transportation systems possible, not to mention powering the businesses, computing, and entertainment that drive modern life. But it’s also no secret that the electrical grid in the US and in other countries around the world is under strain. Extreme weather events have disabled local power grids, aging hardware has contributed to some of the most destructive wildfires in recent memory, and power grid attacks have become a major national security concern. As we enter the low-carbon energy era, new power generation sources and distribution schedules will push the power grid in ways it wasn’t initially built for. That’s where IBM researchers come in.

In a new paper, featured on the cover of Joule this month, IBM and its partners conceptualized the role foundation of models in transforming power grids as humanity moves into a new energy era.¹ These models, called GridFMs, have the potential to help improve power grid operations, planning, and control for the 21st century and beyond. The paper also describes a practical road map for GridFM starting with pre-training on more than 300,000 solved optimal power flow problems on grids of various sizes. A first version of GridFM is expected to be available in the second quarter of 2025. Researchers envision this model using multi-modal spatial, temporal, and text data to support a more resilient grid. Downstream GridFM tasks could include contingency analysis, outage prediction, load forecasting, renewable forecasting, system adequacy, dynamic optimal power flow (OPF), system security, disaster recovery, and dynamic state estimations.

For the average person, this means a power grid that experiences fewer outages, costs less for users, and makes better use of renewable energy resources, according to Hendrik Hamann, chief science officer for climate and sustainability at IBM Research. “This massive network of cables only works if you understand exactly how much of the power flows through every little piece of it,” Hamann said. “It is vastly complex, and to improve that intelligence is exactly where GridFM comes in.”

To address climate change, the only practical way we will decarbonize is through clean energy, said Hamann. And clean energy requires energy transition in the electric grid. For that, you need a massive infusion of intelligence.

“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. 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,” he said. And IBM isn’t doing it alone.

This seminal paper started with a working group that IBM convened along with Imperial College London this year at its global research headquarters in Yorktown Heights, New York. And to move this vision forward, IBM is working with partners in the energy industry to build what was designed in the paper. These partners include Linux Foundation for Energy (LF Energy), which is supporting the open-source development of a common GridFM technology base; Québec electrical utility Hydro-Québec, which will innovate on top of the open-source model by validating and fine tuning to the utility’s specific downstream applications. Other collaborators that contributed to this work include ETH Zurich, Argonne National Laboratory, UK electrical company SSEN Transmission, and a Swiss electricity system operator.

The evolution of our grid

One of the growing problems the world faces is that renewable energy sources are more variable than conventional fossil fuel ones, making it harder to predict and match consumer demand with what utilities can supply. For example, a coal power plant’s output is predictable and reliable, whereas the power produced by wind and solar facilities is subject to shifting weather patterns. The power inverters employed by renewables pose unprecedented problems for the grid, too. Software-controlled inverters at the edge of the grid are crucial for making renewable power production affordable, but the existing power grid wasn’t designed to handle their phase, voltage, and frequency fluctuations — changes whose occurrences and consequences are hard to predict.

“Decarbonizing the grid has been one of the most difficult quests in the last years,” said Bernabé-Moreno. “Not only do we have to account for the volatile nature of renewable energy sources, but there is the added complexity created by decentralization and new challenges introduced by digitalization.”

Even the way people use electricity is changing, as home solar panels feed power back into the grid and electrical vehicle owners charge their cars at home, at work, and everywhere in between. The drastic increase in energy demand that support AI workloads have also added new variables to the power grid, said Bryan Sacks, IBM's energy CTO. “As the increase in grid complexity accelerates, our ability to effectively model the grid is increasingly constrained, leading to compromise and inefficiencies,” Sacks said.

Building on a solid foundation of foundation models

The modern energy system brings with it many new challenges. It requires mechanisms for dealing with the uncertainty of renewable energy generation sources; anticipating the effects of extreme weather events such as hail, extreme storms, landslides, and heat waves on the physical infrastructure; and accounting for the short- to long-term impact of climate change on energy generation and consumption patterns.

We’ve been partnering with NASA to create foundation models capable of understanding changes on the surface of the Earth or understanding the weather and the climate. Our foundation models can, for example, downscale climate models to understand how the weather is going to be in the future, identify areas impacted by natural catastrophes, or predict the formation of heat islands.

When it comes to the GridFM effort, these ongoing foundation model projects can be incorporated as inputs to achieve superior performance. For example, modeling perturbations of the ocean’s surface temperature can aid in understanding effects on offshore wind turbines, not just in terms of energy generation but also in identifying which curtailment strategies will need to be applied to make sure the grid can cope with the generated energy.

New opportunities

Foundation models have the unique potential to analyze and manage the modern power grid. They can sort through massive volumes of information to extract and make sense of data whose full patterns and potential elude the human observer — and which conventional computation can’t crack. This is why IBM, along with academic and industry partners, is betting on modernizing the grid with foundation models. “We see the energy grid as the cornerstone of the energy transition,” said Bernabé-Moreno. And GridFM’s role will be enabling the insights needed to meet the challenges that accompany this transformational era.

Foundation models could provide an orders of magnitude increase in performance for simulating power flow, compared to existing modeling software energy companies rely on. New AI models could also unlock new methods for optimized power flow as renewables are incorporated into the grid, said Sacks. With so many new ways to generate and draw down energy, balancing demand and supply in real time is getting more complex every day.

Another crucial area is what’s called “n-x contingency” planning. “Utilities perform what is often referred to as n-1 contingency planning, where they simulate what happens if they lose one major ‘bus’ or important part of the network,” Sacks explained. “They don’t have enough compute to simulate what happens if they lose more than one, though, so ‘n-x’ would allow them to simulate more scenarios, leading to better resiliency.”

GridFM may be new, but it’s not coming out of the blue. IBM has been at the forefront building foundation models that solve real problems facing businesses. And now, with the scientific community working on new ways to address AI's energy demands, the time is right to use AI to optimize our energy systems for the low-carbon era. “I’ve been working for many years in applying machine learning to drive the energy transition,” said Bernabé-Moreno, “and for the first time, I see a fundamental step change in how AI can address these challenges.”

Publishing this paper is just the beginning. Another GridFM conference will be held at Argonne National Laboratory in Lemont, Illinois, in February, which will be an opportunity for academic and industry partners to come together and engage on the role that GridFMs will play in our collective future.