As AI data centers expand across the country and demand from electric vehicles continues to grow, the challenges to the nation’s power grid seem to become more vexing by the day.
A 2025 report from the U.S. Department of Energy called the current trajectory of the country’s energy infrastructure “unsustainable,” with new generation capacity falling behind what will be needed by 2030 and the risk of outages skyrocketing.
Adding new generation capacity will be an important part of maintaining a reliable supply of power. But making the power grid more efficient and resilient will also be key to getting the most out of existing generation sources and handling increasing fluctuations in supply and demand.
“In the past, electricity demand was fairly stable and predictable, but that’s changing,” says Professor of Electrical and Computer Engineering Wencong Su. He explains that the rise of AI data centers has layered a new challenge onto the grid: While traditional data centers, used for applications like cloud storage, are very consistent in their power draw, AI data centers can see swings in their power consumption of as much as 40%. Su and others are building tools to develop new grid management strategies. And, as counterintuitive as it might sound, they’re increasingly turning to AI for solutions.
Microgrids are a mix of virtual and real-world test beds that private companies and public researchers use to try out new equipment and grid management strategies, as well as to plan for future energy scenarios. They have existed for years, but their use has grown as grids and the demands placed on them have become more complex.
Su explains that the use of AI as a management tool introduces a completely new paradigm for energy researchers who work with microgrids. More computing power could enable them to game out increasingly intricate scenarios and consider variables that were simply too difficult to work with in the past.
“Traditionally, we’ve used very complex mathematical models to understand what’s happening in our microgrids," he says.” The computations for those models are very time-consuming, and when you want to try something new, you have to reconfigure the whole system. AI enables something completely different—what we call a “model-free” system that’s constantly gathering new information, trying out different scenarios and adapting as it goes forward. So an experiment that took weeks before might only take a few minutes.”
Ultimately, he sees AI-powered management tools as a game changer for the growing number of experimental grids around the country, including the one that’s currently operating on UM-Dearborn’s campus.
In addition to improving testing capabilities in existing microgrids, Su is developing an AI tool that will help electrical engineers design and deploy new real-world grids using a chatbot similar to the ones most of us are familiar with. Acting on plain-language prompts, it will enable AI and power engineers to work together to design and optimize new microgrid infrastructure, taking into account the many variables and tools available.
“We’re developing an interface where an engineer can say ‘I want to deploy a microgrid in California that generates X amount of power in X amount of space,’” Su says. “And the bot could recommend a combination of strategies to make it work.”
Su believes that AI tools will also make it easier to integrate new strategies into the power grid, like batteries that can store and release energy to smooth out fluctuations in power demand. While battery storage holds promise as a way to improve grid resilience and reliability, it also adds complexity, and Su says better software could be a way to manage that complexity.
Su is optimistic about the power of AI to improve the power grid, while acknowledging the irony in using it to manage power demand challenges that it largely created. And he believes that energy engineers and AI developers will need to find new ways to work together if lasting solutions are to be found.
“There’s a big gap right now between the AI space and the energy space,” he says. “The people who are developing large language models only care about performance, they don’t think about power issues. And power people like me don’t know how the GPUs that power AI work. I think there’s a lot of potential to build systems that work more efficiently and are easier on the grid.”
Story by Gabe Cherry.
