AI and the Future of the U.S. Electric Grid

The best data the researchers could find came from Europe. So, they modeled one bitterly cold week there, when the grid strained to keep homes warm. They found that the cruise-control AI, known as load reduction, more than doubled energy reserves and reduced average costs by around 10 percent. The other AI, known as load shifting, had little effect on prices, but it did improve energy reserves and prevent sharp fluctuations in demand. A third AI application that the team looked at, automatic control of wind-farm turbines, produced almost no savings.

But then the researchers looked at the risks. Any new technology presents a new target for cyberattacks. But AI could also make confusing or harmful decisions on its own. And when it does, there might be no way to understand its reasoning or how to fix the problem. AI systems also have no human ethics, so if diverting energy away from older, less energy-efficient homes could maximize grid performance, it might not hesitate. An AI also could make erratic decisions in the middle of a natural disaster or other emergency not well-represented in its training data.

RAND researchers identified one other cause for concern in a recent working paper. They ran thousands of simulations of a working power grid. As they increased the number of grid operators with AI capabilities, something unexpected happened. The performance of the overall grid began to slip. Operators with AI started to make decisions from a much wider menu of options. Those without AI struggled to make sense of their often-unexpected moves and to respond as grid performance started to swing. The system “performs well,” researchers wrote, “until a critical mass of AI-enabled operators begins to take hold.”

AI can make parts of the grid run more efficiently, researchers concluded. But it still needs active and capable human oversight. Energy companies will be under tremendous pressure in the years to come to deploy AI to reduce costs. But they need to fully understand what they’re rolling out before it goes live.

Regulators should develop so-called sandboxes where companies can test AI applications before deploying them to the grid. They should require power companies to report any use of AI in the electric system—and those companies should be open about their plans. After all, researchers wrote, the AI transition, with all of its potential benefits, will only succeed if policymakers and the public buy into it.

“AI can do a lot of good,” Arciniegas Rueda said. “It can predict problems before they happen; it can identify equipment that needs repair before it fails. It allows companies to integrate more kinds of energy, like wind and solar. But at the same time, we’ve seen what happens when regulation falls short, and things get out of hand. You get rolling blackouts in California. You get Enron.

“We cannot afford that kind of whiplash, especially not with all of these new data centers coming online, all of these new demands. The grid is the network of all networks. The key over the next few years is going to be fixing it without breaking anything.”

Doug Irving is a communications analyst at RAND. This article is published courtesy of RAND.