You’re drowning in documents. So are the people who can make our energy grid cleaner

Back in early 2021, a PhD candidate in electrical engineering at UT Dallas named Roshni Anna Jacob was developing a proposal with her advisor to study the resilience of Texas’ power grid. Her special approach involved using AI to look at the incredibly complex questions around how energy gets delivered to millions of homes and businesses. Then, in February of that year, a polar vortex hit, sending the temperature to record lows, overwhelming the grid, and throwing most of Texas—including Dr. Jacob—into darkness.

Dr. Jacob, who completed her PhD this summer, has a warm presence and can amiably explain graph neural networks to a reporter without a computer science degree. She chuckled at the irony of spending a few days without electricity or water while studying grid resiliency. But many in Texas had a worse time of it. Some 11 million people lost power for as long as three days. Temperatures went down to 0 degrees Fahrenheit, and officials put the estimated death toll at 246—two thirds from hypothermia—though some experts put the total as high as 700.

Unusually frigid temperatures took a number of power plants offline precisely when consumers were turning their space heaters on full blast. Some were so cold they used hair dryers to keep warm. The Texas grid operator—which, unlike in every other state, functions independently of the Federal Energy Regulatory Commission (FERC)—began massive load shedding, creating intentional blackouts to prevent the grid from failing.

Most people don’t think too hard about where their power comes from. Decades of consistent, reliable electricity have created the illusion that when you plug something in it’s basically guaranteed to turn on. This is not the case globally. In 2022, Pakistan averaged 22 power outages a month. Operating the grid is, on the one hand, incredibly complex—involving hundreds of power plants, half a million miles of high-voltage transmission lines, transformers that bring the voltage down so that it can be distributed to 336 million people, with the added twist of increasingly popular solar panels and batteries creating and storing power locally. On the other hand, the grid is also pretty analog, at least technologically speaking. It is, at some level, about switching power on and off such that the supply matches demand.

People refer to the US power grid as the largest machine in the world, which is awe-inspiring. But much of it is comparatively ancient, built decades ago. And, as innumerable sources report, it’s in desperate need of upgrades. Yet with so many of our decarbonization goals linked to switching from fossil fuels to electricity—from oil burning furnaces to heat pumps, and from gas-powered cars to EVs—the grid is experiencing a surge in demand, or load growth, unlike anything we’ve seen in recent history. 

One factor is the proliferation of data centers used for AI, with Goldman Sachs predicting that AI could drive a 160% increase in data center power demand. But this is one of the core contradictions of the energy transition. While AI is part of the problem, more and more people are asking an interesting question: Can it also be part of the solution?

~ ~ ~

Like other deeply-entrenched, highly-regulated industries, the utilities that deliver our power are resistant to change. Theirs is a high-stakes job with little room for error. With tragedies like the Texas blackout always back of mind, they prefer the reliability of good old gas plants—which, unlike wind and solar, aren’t subject to changes in the weather. That means they’re cautious any time an ambitious group of people attempt to build a new wind farm, like the one led by Matt Perkins outside Anchorage, Alaska. Before Perkins’ power can come online, the regional grid operator must study the impact of the new project on the rest of the grid. This process prioritizes safety, and it functioned better back when new projects came along at a trickle. Now there are hundreds waiting in line.

Each planning study includes other proposed projects in the interconnection queue in order to evaluate what their collective impact would be. If one project drops out, they have to redo the study. This leads us to the second big contradiction at the heart of the energy transition: The wattage of renewable energy waiting to be plugged into the grid is twice our current capacity. 

It is the bottleneck of all bottlenecks—a bureaucratic morass so behind schedule that it now takes an average of four years for a project to get approval, and can take as long as seven. That’s double the time it took ten years ago. Fewer than one in five proposals actually make it through. Former FERC Commissioner Allison Clements called it an “irrational barrier to entry” on the Volts podcast.

Alp Kucukelbir, the chief scientist at Fero Labs and an adjunct professor at Columbia University, thinks AI can expedite the process. “The business of working within the grid is largely documents and text, communication between people and bodies and entities and regulators,” Kucukelbir said on a Zoom call. “All that is up for significant acceleration with AI.”

Kucukelbir—pronounced “cue-chew-kell-beer”—co-authored a 334-page report compiling use cases for AI in mitigating climate change. Among its recommendations, it suggests using language models and AI-based simulations to expedite the planning for new projects. “Can we make permitting faster?” Kucukelbir asked with a hint of exasperation. “Can we go through regulatory steps faster? How do we accelerate the entire process?”

Interconnection planning involves both complex engineering challenges and regular old workflow challenges. The process for reviewing interconnection applications is manual and labor-intensive. AI tools could streamline this by automating data integration, generating preliminary study models, and identifying grid constraints more efficiently. Some people are already working on this. Pearl Street Technologies developed a software tool to accelerate modeling and simulation studies. “This type of analysis is very complex and can take weeks-to-months of engineering time to get to a solved state,” explained David Bromberg, Pearl Street’s CEO, during a recent interview. “Our software was developed to automate these types of time-consuming processes.”

As if to validate the importance of this work, the Department of Energy just announced $30 million in funding for projects using AI to speed up the interconnection queue. With the help of AI, grid operators might be able to race through the worst of the planning process speed bumps. That would mean less busywork, and more time to focus on the real, pressing engineering questions, potentially cutting the wait time down by a matter of years.

In case this is starting to feel like a fusty municipal planning meeting, let’s add some table stakes: For the US to meet its climate goals, the country aims to generate 44% of its power from renewable resources by 2050. That means doubling the amount of clean energy currently hooked up to the grid. Author David Wallace-Wells only underscores the urgency in his book The Uninhabitable Earth; it puts the number of climate refugees somewhere between 200 million to 1 billion if we don’t limit warming to 2 degrees Celsius in the next 25 years.

~ ~ ~

One company—Utilidata—found a way to earn the trust of the utilities, who often get a bad rap in climate circles. But it didn’t happen for them right away. At first, they tried to offer a software solution to help operators optimize the grid. “But then we were asking utilities to become really great at deploying software, and they're not good at that,” Utilidata CEO Josh Brumberger said over Zoom. “But you know what utilities are really good at? Deploying hardware.”

So Utilidata partnered with NVIDIA to develop a chip that could be installed directly on the meter. With AI hardware deployed on site, Utilidata was able to create a platform that gave utilities insight into what is actually happening on the grid—something they’d never had before. Once power is behind the meter—that is, charging iPhones and nuking frozen burritos—the utility is riding blind, at least traditionally. But as the grid becomes more decentralized and distributed, knowing what’s going on becomes increasingly essential. That’s especially true when it comes to big power draws like EV charging.

“How smart is the grid?” Brumberger said. “The answer is, it's not. Period. So what do we do about it? How do you bring the grid to life? Imagine taking a system that's pretty blind today to one 5-10 years from now that can see everything and control everything.”

It’s a vision of a future grid in which AI chips are distributed throughout our energy infrastructure, constantly evaluating how to maximize efficiency. 

~ ~ ~

Reducing emissions in our power system is essential—but so is grid resiliency. What the National Oceanic and Atmospheric Administration (NOAA) calls “billion-dollar disaster events” are increasingly common as global temperatures rise. And the Texas outage made all these wonky policy questions real for people on the ground.

Not long after her power came back on and the water started running from the taps again, Dr. Roshni Anna Jacob began working on a new project with a team of researchers at UT Dallas. To prepare for the next outage, they had a vision of connecting an advanced machine learning model to their clunky, sprawling grid, and teaching the grid to do something new: learn how to heal itself.

“We represent the power grid as a graph,” Dr. Yulia Gel, one of the co-authors and now a professor at Virginia Tech, explained on a Zoom call. “And we use graph theory tools to better understand what the optimal action will be.”

The model can “detect outages or faulty events and then automatically take action to reduce the impact,” Dr. Jacob said. “There will still be outages but it can reduce that with emergency load shedding. Once the outage has been somewhat mitigated, it can take other actions to restore power. […] This requires good decision making on what switches to open and what switches to close, and has to be done very quickly—almost at the speed of light.”

The grid already has some of these capabilities. The issue is that existing algorithms aren’t able to solve problems in real time—especially as the proliferation of distributed energy resources (DERs), such as residential rooftop solar panels and batteries, increase the complexity of the system. During outages, this new model creates microgrids powered by DERs. “From a resilience perspective, the smaller the unit, the easier it is to control. You contain the outages. It’s a more local solution,” Dr. Jacob said.

It takes time, however, for experimental new models like this to migrate from academia to the daily work of operating a power grid—which raises the issue of trust. Kucukelbir, the chief scientist at Fero Labs, is passionate about this subject. AI, he emphasized, is something assistive—a tool that you work with. Take the impact of AI on NOAA’s advanced weather modeling: “We're not throwing our national labs away anytime soon,” Kucukelbir says. “NOAA's here to stay. But if you can accelerate that with AI, that becomes a parallel tool.”

It can be an uphill battle getting large, traditional industrial and institutional players to integrate new technologies such as AI into their workflows—which makes the trust factor all the more essential.

“As we encourage the adoption of AI in industry, we need to recognize that there's someone there whose job is to make sure that there isn't a blackout, that this pot of steel doesn't explode molten metal, that this chemical doesn't leak into Mexico,” Kucukelbir said. “All of these things are someone's job, and they will only trust technology that they understand.”