EPRI’s Jeremy Renshaw on the benefits and limitations of artificial intelligence in the power sector
It may seem as though artificial intelligence (AI) was born with the launch of ChatGPT in late 2022. However, the reality is that the development of AI and the study of its many potential applications has been underway for several decades, dating back to shortly after the invention of the computer. Indeed, EPRI has researched the impacts of AI on the energy industry for over a decade. EPRI has completed or is currently involved with over 70 projects using AI to improve everything from wildfire detection and response to grid management to cybersecurity.
Policymakers, regulators, utilities, and the public have many questions about the appropriate and responsible use of AI. EPRI Senior Technical Executive Jeremy Renshaw testified before a U.S. House Energy and Commerce Subcommittee last year about the challenges and opportunities of leveraging AI to build a more reliable, resilient, clean, and efficient power system. Renshaw recently sat down with EPRI Journal to discuss the many current uses of AI in the energy industry, how AI can deliver more value and the critical importance of quality data in delivering on the promise of AI.
EJ: Talk of AI is everywhere these days. At a high level, how does EPRI view AI, and how is EPRI helping with its effective application in the electric power industry?
Renshaw: AI is one of the biggest buzzwords today. From startups to tech companies, whenever you turn around, it seems like somebody is trying to sell some new product with AI. At EPRI, we’re trying to help people understand the value of AI without overhyping it. It’s a difficult balancing act to ensure that we’re appropriately telling people what AI can do and what it might be able to do in the future. It’s exciting, but we want to ensure that what we’re saying is backed up by facts without overselling or overhyping it.
EJ: Is most of the AI activity in the electric power industry just research right now?
Renshaw: There is much talk and many R&D projects, but there is also increasing adoption. We have seen energy companies using AI for wildfire detection, improvements in online monitoring and predictive maintenance, and more, but adoption has lagged in some other industries. Part of the reason the industry has moved a little slower than others is because of its focus on safety and reliability. For example, if you’re doing marketing or sales and advertising and have trained an AI model that reaches 99% accuracy on the people you want to reach, that’s a huge win. If you’re in the electric power industry and hit 99% electricity availability, that’s a huge miss since it means 14 minutes per day without power. People expect more. There are different expectations of safety and reliability for electric power versus just about any other industry.
EJ: You recently had the opportunity to testify before Congress about AI in the electric power industry. One of your points was that humans still need to be relied on to do the tasks humans are best at, while AI can help expand the value that computers and technology can deliver. What were you trying to get across?
Renshaw: For example, my kids were on a swim team for several years. And you do four different strokes in these relay races. If you allow each person to do the stroke that they want to do, you may not end up with as good of a result as when each kid is doing their best stroke. It’s similar to humans doing what humans do best and computers doing what they do best. If you love doing math and are great at performing calculations, you’ll still never be able to perform calculations as fast or accurate as computers. However, there are things that humans can do far better than computers in terms of creative thinking, developing new ideas and concepts, or responding to new scenarios they’ve either seen once or never been exposed to—what we call single-shot or zero-shot learning in the AI community. By helping humans and computers work together, doing what they do best, the result is greater than the sum of its parts.
EJ: What is an AI use case that benefits the industry today?
Renshaw: Several areas are delivering value now, like predictive maintenance. Our Generation Sector uses AI to look at wind turbine gearbox analysis. They have machine learning models that monitor the different operational parameters. Using that data, they can predict gearbox failures months ahead of the failure. When you do that, a utility can schedule maintenance and repair components at an earlier stage of degradation where they’re only replacing a $50,000 component versus a $350,000 system. The analysis that they’ve done shows that an average wind turbine farm, in aggregate, can save about $1 million per year by planning its maintenance around what they see with the data they already have.
EJ: What are some promising new applications of AI?
Renshaw: One that will continue to improve for many years is grid management. The U.S. power grid is often described as the most complex machine humans have ever built. There are several opportunities for AI to help today in improving forecasting for both loads and weather, evaluating control center alarm data, and analyzing the large numbers of images utilities capture of their T&D (transmission and distribution) infrastructure to perform evaluations. Despite these and other current benefits, some things will take quite a while, like solving optimal power flow. That’s one of the most complex, if not the most complicated, mathematical challenges in the world because the complexity scales exponentially with each grid-connected component and is far beyond what even today’s supercomputers could do in any reasonable amount of time. And so, while the power grid is operated safely, reliably, and reasonably efficiently, it’s essentially impossible with today’s technology to fully optimize power flow in a power network because it’s exceptionally complex.
EJ: Can AI help with the complexity of distributed energy resources and variable renewable generation?
Renshaw: That’s a big challenge for sure. With solar, the sun goes behind a cloud, and your power output goes down by 30%, 50%, or 70% almost instantaneously, and the grid must adjust. Fortunately, there is much inertia on the grid currently that helps. However, with increasing renewable penetration, these swings can become more severe and unpredictable, so inertia alone won’t be enough. For example, I was recently in Texas, and in certain areas, they’re starting to see power peaking at 7 pm or 8 pm when the sun goes down instead of the traditional peak of 3 or 4 in the afternoon. AI can help you understand how and when to turn on your other power plants or turn off smart resources so that you have both more flexibility and more lead time to ramp up generation in response to changes in generation or consumption.
EJ: What is another impactful use case?
Renshaw: One of the applications I’m most excited about is AI for wildfire detection. We worked with a company using cameras on mountaintops, looking for evidence of smoke and fire throughout the day. This potentially allows utilities to respond significantly faster to wildfires and extinguish them at earlier stages when they’re easier to handle. It also can save money because you don’t need an army of people looking for fires. AI can analyze images all day, every day. It never gets bored, distracted, or takes a break while looking for fires. Regardless of the source of a fire, this solution may offer a low-cost, reliable, and safe way to help with a real threat to people, property, and the environment that can have drastic consequences.
EJ: How much of the value of AI depends on the quality and volume of data?
Renshaw: Data is at the heart of AI. Data quality and quantity are central to anything you will do with AI. Everyone likes to think their data is clean and wonderful, just waiting to provide insights. However, a large percentage of a data scientist’s job is cleaning up messy data. Let’s say you’re monitoring a power plant. If you periodically shut the power plant down for an outage, now you have sensors gathering data that’s not very useful because the plant isn’t operating. Temperatures, pressures, and flows are all very different but generally do not provide any operational insight. You may need to manually remove that data from the analysis, or you’ll have a bunch of zeros or weird numbers that are not indicative of anything useful. It’s just unusable data.
Similarly, many things can happen during operations, such as interference, instrument noise, noise from a loose wire or poor physical connection, poor communication, or simply a dropped data point that may cause you to need to go and evaluate the data before you can just upload it into an AI model. If you plug unconditioned data into a model and assume it will work, you will often come out with something that wastes time and is useless. This often results in people thinking that AI can’t do things it can. It’s just more complicated than people realize to do the pre-work before training and using an AI model.
EJ: What are data quality implications for utilities and others in the industry pursuing AI?
Renshaw: You’ll generally get more bang for your buck by increasing your data quality and quantity than you will from using the latest and greatest AI model. People are always coming up with new algorithms and models. What inexperienced users often forget to focus on is feeding them very high-quality data as well. The biggest value will almost always focus on the data first, then the model second. Of course, we want to use the best models available, but cleaning and preparing the data should come first.
EJ: EPRI is currently working with Google, Nvidia, and other stakeholders to develop AI capabilities to improve grid management. What are some other AI projects EPRI is pursuing?
Renshaw: As mentioned earlier, EPRI has about 70 AI projects that are active or recently completed. We have a few that are rising to the top that we need to execute, including generative AI, which is the buzzword of buzzwords right now, though I should point out we were doing generative AI before it was cool. We have worked with generative adversarial networks or GANs to generate images of insulators and other transmission and distribution infrastructure and create nondestructive evaluation (NDE) data to train inspectors on a wide range of virtual flaws. We have built on that work to look at how we can use large language models to augment what we’re doing with our EPRI.com search. Similar to what Bing, ChatGPT, or some of the other tools will do, we want to augment our EPRI.com search to ask questions and get accurate answers quickly and easily. In parallel, we’re also working on program-specific expert systems so that each EPRI.com search will give you the best possible answer to your question based on your access. In that way, we’re hoping to be able to provide essentially expert-level answers to questions that people ask, as well as the underlying reference materials. This allows people to double-check answers because, at the end of the day, the large language models are essentially a very fancy autocomplete and have a tendency to hallucinate. We want to ground models to minimize the chance that the model is making up an answer and have the model provide the reference the answer is based on so it can be verified. It’ll take a while before we get there. As I mentioned, this is much harder than most people think. But if we could bring all that knowledge together, it would be an extremely valuable tool. All of this takes a long time to accomplish and is difficult, but we believe it’s worth the effort.
EPRI Technical Expert:
For more information, contact firstname.lastname@example.org.