196. Building ChatGPT for the Power Industry: EPRI Leads the Way

Hi, everyone. This is Aaron Larson, Executive Editor of Power Magazine, and you are listening to the Power Podcast. On today's episode, I'm joined by Jeremy Renshaw. Jeremy is Executive Director of AI and Quantum with EPRI. EPRI is many in our audience will know as an international independent non-profit R&D institute that works intimately with the utility industry. So, Jeremy, thanks for coming on the podcast with me and please tell a little bit about yourself, your background, and EPRI. All right. Well, thank you for having me on, Aaron. As you mentioned, I'm Jeremy Renshaw. I lead our AI and Quantum Research at EPRI. So, my work focuses on the intersection of artificial intelligence and quantum computing with the energy sector, specifically how we can responsibly and effectively apply these technologies to accelerate innovation, improve energy system reliability, and really support the transition to a clean energy future. So, one of the most exciting initiatives that I'm involved in today is the Open Power AI Consortium, which brings together global leaders in both the energy and technology spaces to help shape the future of artificial intelligence in the power sector. From what I understand, I was launched back in March during an event, and I guess I'd like you to talk a little bit more about it. I know Microsoft, AWS, Oracle, some of these other large companies were involved. What is the real purpose and mission? And how do you expect this to potentially benefit the power industry in the long term? Yeah. So, great question, Aaron. The Open Power AI Consortium was launched at NVIDIA's GTC event in March, as you mentioned, and we had founding members like Microsoft, Oracle, AWS, Google, and major utilities like Duke Energy, Southern Company, EDF, NL, Keppco, and many more energy and technology companies, universities, national labs, etc. The mission is really simple, but very bold. It's to accelerate the safe, secure, and scalable adoption of AI technologies across the power sector for a wide range of use cases. We're really looking at building an ecosystem to accelerate the development of deployment and recognizing that while AI is advancing rapidly, the energy industry has its own unique needs, especially around reliability, safety, regulatory compliance, and so forth. So the Consortium provides a collaborative platform to develop and maintain domain-specific AI models, thank a chat GPT tailored to the energy industry, as well as sharing best practices, testing innovative solutions in a secure environment. And long-term, we believe this will help modernize the grid, improve customer experiences, and support global, safe, affordable, and reliable energy for everyone. So, further more, for many years, the power industry has been somewhat siloed and there were not many touch points or communication between global utilities, technology companies, universities, and so forth. So this Consortium aims to facilitate making new connections between these important and impactful organizations to increase collaboration and information sharing that will benefit everyone. From what I understand, you are developing, you know, domain-specific gen AI models rather than using kind of the general AI tools that are available to the public. Can you explain why that's important and how a power sector-specific AI model might outperform some of these general AI solutions? Absolutely. So, we don't, in any way, want to negate the effectiveness of existing models today. Certainly, CHABGPT, Clawed, GROC, the other foundational models or frontier models that are available today, provides significant value. One of the reasons that we're looking at the domain-specific models is looking at improved contextual knowledge understanding and retrieval. So our industry deals with real-time systems, critical infrastructure, and strict regulatory requirements. So, a domain-specific gen AI model can be trained on utility-grade data, understanding utility terminology, and make contextually aware decisions. For example, a general AI model might misinterpret terms like islanding or blackstar, whereas a power-specific model can understand that these are critical grid operations. And this can help to lead to more accurate, safer, and actionable outputs where you can trust the output at a higher level, which is essential for dealing with mission-critical systems. So, in other words, if we look at AI models today, when they achieve an 80% or 90% on certain standardized benchmarks, that's celebrated widely around the world. So if an AI system were to get a 99% on a standardized cluster benchmark, that would be groundbreaking in many cases. But I think we can all agree that keeping the power on 99% of the time is completely unacceptable. Or if we take it to the extreme, making sure a nuclear power plant won't melt down 99% of the time, we would just never accept that. So there's a big difference in the reliability and safety needs of the power industry compared to general purpose applications in many other areas of the world. The press release I saw mentions creating a sandbox environment for testing AI applications. What does that actually look like in practice? And what kinds of breakthrough use cases are you most excited to validate in that setting? So great question, Aaron. I would say to take a step back, the sandbox is a secure cloud-based environment where utilities and technology partners can safely test AI applications using either synthetic or anonymized data. You could also look at it as a proving ground for innovation where we can validate models before they're deployed in the real world. So in reality, the sandbox is really a set of sandbox. It would be hosted and deployed in conjunction with technology partners with AWS, Google, Microsoft, WWT, SHI, Oracle, and many others. So there could be a range of sandboxes that would be hosted either by utility, by a technology provider or even by ePRI, to be able to test out and vet solutions, to allow for rapid development, deployment, and scaling once this solutions are vetted. In terms of use cases, it's really like asking me which is my favorite baby. You just don't have one, but a few use cases that I'm excited about would be things like predictive and condition-based maintenance looking at power generation or grid assets to do maintenance as needed, not necessarily on a scheduled basis. So case in point there, many systems that we perform maintenance on, or that the industry performs maintenance on, were working fine until they were brought down for scheduled maintenance that may not have been needed. And then that can result in foreign materials being introduced or different potential issues that happen as you perform maintenance. So in many cases, leaving equipment alone is often the best solution. But if you had additional data, you could understand how and when would be most effective to do that maintenance. Another use kit will get customer service automation. I don't know about you, but for me, whenever I get the computerized system at wherever I'm calling into, I'm often repeating agent or representative or whatever to get to a human. And imagine if we can make it so the AI agent was significantly more efficient than the human, so you didn't have to give your information three or five times to different people, though that in the future, we're asking for the AI agent instead of the human. So that would be another goal that we would like to get to. You could also imagine another use case for AI co-piles for system operators that could help to identify potential solutions for existing challenges faster, bring the data that you need to you immediately to make better decisions more rapidly. You can also look at AI assistant planning for renewable integration, setting switches in relays, running powerful simulation, short circuit simulations, load forecasts and weather forecasting, and so many more. We have already put together a library of over 250 use cases, and we're in the process of ranking and evaluating those to bring the most valuable solutions out for the conspiracy to look at and work on. And from what I understand, now, EPRI has already developed some of the first industry specific GNI models. They may not be available yet, but they're coming very soon from what I understand. What can these models do that wasn't possible before and how do you see them really benefiting utilities? That's another great question. So EPRI articulate and Nvidia have been working together on the first version of these domain specific models. I would call it the alpha version. So while they're not available yet, you're right. They will be available soon via an Nvidia NIM or Nvidia inference microservice, which means that they're easy to deploy in scale and can be brought to you in a containerized format. So some of the things that we're focusing on with the domain specific model and our partners and video articulate and others would be looking at deeper understanding and contextual knowledge retrieval for the comment I made earlier about islanding and Blackstar and various other industry specific terms that we have that may mean something different to us than someone else. Case in point, NDE for someone in the power industry means non-destructive evaluation, whereas outside of the power industry that often means near death experience. So we have very different acronyms, technology, terminology, and so forth. So using domain specific models could help with this to generate more accurate technical documentation, a whole generation for specific tailored applications, whether it's grid simulations, power plant operations, as well as improving the accuracy in relevance or summarizing complex documents like regulatory filings or outage reports. And finally, you could help to translate complex grid data into more plain language for decision makers to again make better decisions faster. While some of this could be possible to some degree with off-the-shelf models, we anticipate improvements in accuracy and relevance using customized fit-for-purpose models, and we're now seeing AI that speaks the language of the grid. So that's where we see the benefit for developing these domain specific models. I know you've got some massive utilities such as Duke Energy and Southern Company involved in the consortium, and I know there's also international players like Capco, Korea, Electric Power Company, and the Saudi Electricity Company that are members. How do you navigate the different regulatory environments, technologies, and priorities that each of these utilities have across such a diverse group? Yes, with over 125 organizations that have already joined the consortium, it's a great challenge for sure to navigate the needs of hundreds of voices from different parts of the world with different background experience or geographically specific needs. It also provides a very unique opportunity and highlights the importance of global collaboration. So we have built the consortium to be modular and inclusive, to be able to tailor to the needs of members around the world in different localities, different regions, to be able to collaborate on shared challenges, realizing that there may be differences in policy and regulation. One specific example with the congestion management, which is often done more based on local policy and regulation, than it is on necessarily technical challenges. So we are using working groups focused on specific themes like the domain specific model working group or tailored use case work groups that will focus on power plant or power grid type operations. It allows members to contribute where they have expertise as well as bringing that global experience because the way you operate a power plant even within a country in one region versus another may be different. You can imagine the western U.S. where it's drier versus the eastern U.S. where it's wetter. You can imagine a power plant or power grid in Saudi Arabia where it's very warm and dry versus a power plant in northern Canada where it may be cooler and wetter. So we need to bring in these global perspectives. And as we do that, we often find that many of the challenges are the same around the world and many challenges are different. But as we bring in the different voices, we help to tailor the research needs to a broader stakeholder group which generally provides more value to everyone. The energy sector has traditionally been very conservative about adopting new technologies. That's important because of safety and reliability issues that you definitely can't afford to ignore. How are you addressing the concerns that they might have about AI and the reliability and how do you build trust for mission critical applications using AI? That's a great question. So I would say trust is foundational and that's why we're really focused on having high reliability trust worthy systems and focusing on the AI sandbox to do rigorous testing with transparent documentation and human and the loop design. So we're not necessarily having AI make automated decisions that are on critical infrastructure. We're also working with stakeholders from around the world to define AI governance frameworks tailored to the energy sector. So we want to have predictable, explainable, and repeatable AI so that as we use it in challenging or highly complex environments, we understand what the performance will be and help to improve safety, reliability, and affordability of energy. And I guess I'll steal a quote from one of my favorite people Greg Selby who used to work here at EPRI and he would often say that power plants or energy companies line up to be fourth. So very rarely, as you mentioned, do they want to stick their neck out and be the first one to do something? They want a very robust, highly tuned and tailored and highly accurate solution before they get started. So that's why we want to make sure that as we develop these AI solutions, we keep that framework in mind. Again, going back to my comment of 99% isn't good enough and it's not even close to good enough. We have to understand the high threshold needed to be able to develop, to deploy these models, especially at scale with large companies, they're in this highly regulated and safety conscious environment. So we want to make sure that the solutions that we develop further support that mission and are able to be deployed in a way that they can be explainable and repeatable. But also, if anything were to go wrong, we can track down how and why and be able to address that and continue to improve the models in the future. One of the benefits that I heard coming from the consortium is that you're hoping to improve the energy customer experience. And from a customer's perspective, what tangible changes might people notice in their daily lives as these AI applications are rolled out? Being an energy customer myself, that is a topic that is of great interest to me as well, both from the open-power AI consortium standpoint, as well as, you know, paying my energy bill every month. And wanting to make sure that I have safe, affordable, and reliable electricity at all times. So one of the key things is faster outage response time. So none of us enjoy having the power out. We enjoy generally in most regions of the world very high reliability of energy. I have been to certain countries before where they didn't have reliable energy and there would be curtailment or periods of hours at a time on a daily basis where energy was cut off. So I think through improved AI tools, this could help not only in first world countries, but also in second and third world countries to reduce the frequency and duration of outages. You could also look at more accurate usage insights into your energy bill. One of the things my local energy company does is on a monthly basis. They'll give me a comparison of my home to other energy efficient homes. And it's kind of a competition that I have going on to beat all the other homes in my category. And they also give me insights into how I'm using energy, how much goes into heating and ventilation and cooling, how much goes into lighting, how much is always on, goes into cooking. So having these additional usage insights can help customers to understand their energy utilization and potentially work in the future with demand response programs to seamlessly integrate these in to save money for the customer as well as save money for the utility. So helping to find those win-win situations. You can also look at smarter EV charging and home energy management. So I think there are a lot of potential opportunities to benefit customers. In addition to just the traditional customer service angle of getting to your building statement and paying that faster and easier and addressing concerns or questions that you might have. So I think using AI in a responsible manner can help utilities to be more responsive, proactive, and customer-centric without compromising reliability, but also from the flip side help the customer to have a better experience, potentially pay less money and have more reliable energy. So I think there's the potential for everyone to get a win here. And I know for utilities, grid modernization and decarbonization are probably two of the most important challenges that they're facing. Can you talk about how AI might help address grid modernization and decarbonization? Yes, so I think there is a very real potential for AI to help us here. So if we look at the grid modernization angle first and then decarbonization, we can think of AI helping to optimize asset management as well as helping to enhance power flow in a grid. Every Google and RTE, the grid operator in France, had collaborated on the learning to run a power network challenge for a number of years looking at reinforcement learning techniques in simulated power grids and found that through these techniques, you could optimize specific power grids for responding to various types of challenges and looking at how we can effectively scale up those types of solutions could really help with modernizing our grid and helping them to be more efficient and more effective. You could also look at AI providing more real-time situational awareness if you have somebody who runs into a power line and knocks down a power pole or you have a tree that falls on a line or you have severe weather that comes and impacts your grid looking at that real-time situational awareness to be able to more rapidly respond to potential grid issues would help from that grid modernization standpoint. You could also support autonomous grid reconfiguration as needed with these types of tools so you know how and where to reroute power flows to adjust to certain types of perturbances. On the decarbonization side, AI can help to forecast renewable generation to be able to understand how and when you can optimally utilize these assets and potentially turn off other less green friendly technologies. You could also balance distributed energy resources and as I mentioned earlier have a more seam line approach to where instead of maybe turning someone's air conditioner off for 30 minutes or an hour you could turn it off for 5 or 10 minutes where it might not even be noticed or you could dim a television by 10 percent and then maybe delay a washing machine to turn on by 5 minutes or something like this to where if you stack up a very large number of very small changes you can have significant impacts on the grid to be able to balance supply and demand of energy. We could also look at the accelerating of permitting and planning for green energy projects and really in the at the end of the day it's all about making the grid more flexible, resilient and sustainable. So AI is one thing that can help in these different areas for grid modernization and decarbonization but certainly it's not the only thing so as we pair AI technologies with other solutions it will help us to be again more flexible, resilient and sustainable. And as AI has become more popular we continue to hear about the enormous amount of energy that it's expected to require for data centers and training requirements and all of the other computing that goes into it so how does the consortium think about some of those circular challenges? Are you thinking of ways to optimize energy usage in AI or isn't that really part of your mission? Yes so we're very aware of the irony of AI being able to help energy but also needing a significant amount of energy. So AI certainly has a very large energy footprint. AI models are growing in size and growing in the amount of energy they need to consume to be able to train and then deploy models. And on the flip side the chips that are being used are becoming more and more energy efficient. So there is a bit of an offset there but not enough to make up for the vast increases in model size and utilization. And there is a paradox, an interesting one called Jeven's paradox that as you make a technology more efficient, the utilization of that technology often goes up and overcomes those efficiencies. So there have been vast increases in the number of tokens generated. Tokens are on portions. You could think of this as the words generated by a GNI model. So vast improvements in the energy efficiency of token generation. But because AI models are so much more valuable today than they were two, three, four years ago, they're used that much more. So the energy requirement has continued to go up. So what we're focused on with the open power AI consortium is looking at these tailored fit for purpose model. So you can imagine not needing a one trillion parameter model if a model that's one percent of that size could get you an equally if not better response. So this is why we're focusing on very highly tuned models with just the data that they need going into that domain-specific models. So we can have efficient model architectures potentially utilizing edge computing in some cases. So we don't need to necessarily send data to a central processing unit, have a process, send the data back, but really utilize edge computing with smaller, highly perform models. And we're also looking at through parallel efforts, things like our DC Flex Initiative, where we can utilize data centers as a flexible grid resource to be able to replace diesel generators with clean energy generation, potentially even using the existing diesel generator with a clean fuel. So we're looking at various aspects of how the AI and power systems overlap and how we can utilize AI technologies to help the power system but also utilize the power system in a flexible way to support the power needs of AI. So we realize there's very much a circular economy going on here and a challenge, but we're tackling it head on with many of our partners from around the world. So we think bringing in that global perspective will help to bring the right insights and as we continue to bring the right insights together to be able to scale those not only within the United States but around the world. And maybe thought about the future and what that might be hold. I know EPRI is focused on research and development, not predicting the future, but what do you see and how do you envision AI transforming different aspects going forward and how the power sector might change as a result of it all? I think that's a great question and one that's really hard to predict with the vast scaling of AI models, what wasn't possible three months or six months ago is very possible today and even commonplace. And so being able to predict where we go in one, two, or five years, I would say there are many possibilities and the only thing I can guarantee is that my prediction will be wrong, but I'll make one anyway. So I'm excited about the concept of digital workers where every one of us could have two, three, four digital workers that report to us in terms of giving them tasks to do to be able to automate and accelerate certain portions of any of our jobs. So case in point, I don't really enjoy making LinkedIn posts, but it's something that can be important to get knowledge out into the world. So you can have a digital worker that creates posts for you. You can have one that reviews your email and says, hey, these are the next three tasks that you need to do that are most important. And here I've started on each of them for you. You could also have these digital workers go out and run simulations, collect data, do background research, summarize meetings and reports that you're able to attend and do various tasks for you. So I think it helps to utilize systems and technologies today to be able to build up our AI muscle, if you will, for the future because the technology's not going anywhere. It's moving quickly. It's evolving quickly and the value it's providing is expanding. So I think those of us who learn to embrace this technology will become so much more efficient and more effective than those who don't, that it will be a required skill in the future. If you'll indulge me for a second, I'll make a second prediction. There's more power sector specific. I would say something that I could see happening in the near future, maybe not in a year or two, maybe five plus years, would be something like self-optimizing micro grids, where you can use AI to manage a micro grid in real time to be able to island the micro grid, reroute power balance loads with or without human intervention to be able to improve resilience, enabling distributed energy as well as clean energy sources. And I would say it's, it sounds a little bit science fiction, but it's already starting in various pilot projects around the world. So I would say that this type of technology could provide a lot of value. And I think it's one that we need to pay attention to in the future. Yeah, and I think you may be right, and it may be coming sooner than we even realize, you know, five years in the future is probably too far for that type of thing to actually start becoming more important. So Jeremy, thank you so much. I really enjoyed this conversation. Is there anything that we haven't touched on that you think might be important to talk about before we wrap up the podcast? Well, I'll say two things. First, I like to look at where have we been and where have we come from. So when I started at every about 12 or 13 years ago, one of the first projects that I worked on was with robotics and drones. And we were saying, hey, this is a very useful technology. It's one that we really think the energy industry is going to need to use in the future and is going to use. And people kind of laughed at us and said, oh, you're just, you know, you're just dreaming over at every year. This isn't a real technology. It's never going to do that. And then about eight or nine years ago, we saw cloud computing really starting to take off in other industries. And we kind of came and said, hey, you know, the power industry really needs to pay attention to cloud computing. This is going to be a big and important area in the future to scale out what we're doing. And people kind of came back and said, ah, no, we don't think we'll ever put our data outside of our firewalls and on the cloud. It's just not feasible of practical. And then about five years ago, I took over our AI development efforts and started to see a lot of value in these technologies even before that. And so myself, along with many others, we're saying to the industry, hey, you need to pay attention to AI. This is something that's coming. And we often kind of got laughed at again saying, oh, no, this is just a pipe dream. And fast forward five years. Now we're flying drones, gathering data, storing it on the cloud and analyzing it with AI. And that's normal. And that's just ace of technology changes that we've seen. So before I wrap up, I want to say that the open power AI consortium is really more than just a technology initiative. It's a collaborative movement. It's an ecosystem that is continuing to advance and accelerate innovation for the power sector. Bringing together utilities, technology companies, researchers, regulators, and more to shape the future of AI to serve public good, to enhance grid reliability, efficiency, power system operation, and accelerate the clean energy transition. So I would say the power sector has always been about innovation. And with AI, we're really entering a new era. Now, I would say if you're interested, come join us. Look up the open power AI consortium. It's open for anyone in the power of industry or the technology industry to join. We are growing fast. And we're looking to make a big splash in the global scene. So we would welcome other organizations to join us as we embark on this journey. All right. Well, thank you, Jeremy. Again, for listeners. I've been speaking with Jeremy Renshaw. He is the executive director of AI in quantum with EPRI. It's been a pleasure having you on the program. And I look forward to what you are going to develop from the consortium and with AI in the future. So thanks so much for your work. Thank you for having me, Iron. This is a real pleasure.