The Exchanges, every show

Q What is, in, in the modern day, but prior to, you know, revolutionization by AI, what is utility resource planning actually look like inside the utility?

A Yeah, there, there's many parts of planning. When you say resource, it might be integrated resource on the bulk power system, or with, um, centralized generation, there's transmission planning , there's distribution planning, then there is DER planning. Each one of these are on silos, um, and recently just talking to a very large East Coast IOU, even transmission planning, there are like 12 departments doing their own thing. And each time a study is done, it's completely separate and siloed, so each use case is like a study. You want to connect a generator, it's a study. If you want to connect a load, it's a different type of study, but the underlying model is still the same. So studies today, I would say half if not more of the time, especially for distribution, is on cleaning up data. So data is in a ton of different places. Data quality is not too par, and a lot of manual effort is required to pull data together. Then you run the analysis, which is really tuned for a worst-case scenario planning. What's the five hours of the year that's going to be worst case for the next 10 years, and you plan to that level of standard. Now, people are improving, like looking at 96 hours per year, two 88 hours per year, five 76 hours per year, so each of those are like high, best case and low cases on weeks of the season or month over the years, and I would say the gold standard today is pr…

AI assessment note: “Each one of these are on silos... half if not more of the time is on cleaning up data.”

Q training, or using the, the model that has been trained to To, um, execute tasks. Is it a similar thing in the sense that what you have to do is basically train a model on the physics of the grid, and then once it has been trained, then the idea is you can query it like you would query an LLM and get something that is driven by that physics?

A That's right. I think, one, we're not an LLM, but I think the training inferencing process is similar. An LLM is more meant for generalization. It learns a lot of stuff, and you can have natural language type of interfaces, but it is prone to hallucination. So for us, since we're trying to really bring the world of engineering and AI together, We really lean into machine learning, so deep machine learning models, but we are really playing at the intersection of physics equations and the AI models, the machine learning models themselves. But what that drives towards is a very deterministic model. So our models actually can't hallucinate. It's actually deterministic. You ask it the same thing, it will give you the same answer all the time. The other thing is you have that type of efficiency. We look like the, the unit economics of AI models. Like we're very familiar now with frontier, uh, LLMs that will cost billions of dollars to train and months, if not years to train as well. It learns a lot of stuff, but it is a huge hefty investment and you have to generalize it towards everybody. So it is a jack of all trades and getting smarter and smarter all the way. For us, there's a few things. One is we need to establish security. So we do models that are exclusive for that utility. We're not getting one utility's data and trying to cross train with another utility's data because that…

AI assessment note: “That's right. I think, one, we're not an LLM, but I think the training inferencing process is similar.”

Q Yeah, I mean, just based on your description of it, it sounds like, um, Maybe there is less to be gained from AI in operations versus in planning. In other words, like, if it's already more automated today, setting aside alarm management and extreme events, maybe there's, there's less of a, the ceiling is lower for improvement. Is that true in your mind, or is it just that it's different?

A I think it is the, the complete opposite. I think the potential is even higher. And when you align, like I think today we care so much about connecting data centers and large loads and even generators that we, we put a lot of emphasis on the business case of planning because you are increasing, uh, you are, it's the growth engine. But, uh, before this whole large, big electrification, low growth process, I think the biggest value has been reliability. And there's a lot more to do with that. And then workforce efficiency, workflow efficiencies. So in the, in the operational side, the role of AI is really to bring a lot of those system studies in the past. That's one time, ad hoc, or episodic, or periodic, or reactive, or responsive to customer interconnection request into a continuous, real-time analysis engine. So what is the true, like, autopilot? What is the true copilot? I would expect that from a car, of course, a plane, and a spaceship, but can I now expect that from The grid. Can the grid study, read the roads, analyze scenarios, assess risks all the time? And then that's one layer. The one, the first layer I would say, the foundational layer would be the grid analysis and simulation intelligence layer. But the layer on top is there's so much to be gained by having AI learn how the human processes can be improved. How dispatch can be improved, how alarm management can be …

AI assessment note: “I think it is the, the complete opposite. I think the potential is even higher.”

Q it's not automated, that there are, you know, there's like a bunch of people in a, in a knock with a bunch of big screens who are like watching the screen light up red when there's an outage somewhere, and then they have to go dispatch somebody to go solve it. But the, generally, the grid is already fairly automated from an operations standpoint. What is it actually like today?

A I think that's a fair statement. It's, um, um, I think the automation today is, uh, is there thanks to Field automation. So I think we, traditionally we have been very concerned about automation inside the control room, but the grid itself in a substation at a generator site, it's pretty much completely autonomous. And that's, that's why it has been so reliable and so fast and it balances itself so well. And it's, it's too fast for human in the loop control. So the human in the loop side for operations, it's really monitoring and responding To the abnormalities. So it's alarm management. What happens when things go wrong and the, the operator has to diagnose it. Uh, it's a lot of rules as well. So the, the, so there's an operational planning phase. That's, that's traversing timelines. So operations typically look at, look up maybe, uh, a week ahead, two weeks ahead to look at, uh, uh, switching procedures. So which lines do I have to take out, out of service to do plan work? If we can't do planned work, we cannot upgrade the grid. And now here's the sort of the constraint during that time period. The grid is getting so congested that if we do worst case scenario all the time, um, we, we are, the utilities are beginning to struggle to find time to take the grid into contingency for planned work. So if we too, if we overload the grid too much, we, we can't work on it. That's a bi…

AI assessment note: “the grid itself in a substation at a generator site, it's pretty much completely autonomous.”

Q like, is that harder than it seems, basically? Um, is it going to be operationally very difficult to run one of these very large loads that probably doesn't, that has like a spiky load profile off of this gigantic microgrid? From what you know about operations of a macrogrid, Do you think it's going to be easy or hard? I mean, or, you know, what's the, what's the nuance here?

A Well, it's a bit of both. But, uh, I think we were building microgrids on the distribution side since the, well, nineties and 2000, right? We, we understand that they can add a lot of, a bit of local resiliency, but let's remember why we have a grid in the first place, which is so that we don't become individual islands. And individual islands are always more expensive. Than sharing the resources that we have. And so, yeah, I think they can use it to protect themselves and to advance their time, speed to power, their own capacity requirements, etc. But it's not going to sort of serve the greater good by sharing, having centralized generators plan appropriately and serving a variety. So it is really being an island. So technically, I think the good thing is, yes, you can mitigate some of the immediate capacity constraints, but I think there are Existing latent capacity. So let me just make a statement here, which is, I believe the grid has enough existing latent capacity to connect the majority, if not all of the data centers today. We just need to find where they are and improve the overall processes of planning and operations for the utility. So if we are to build all these generators behind the meter, why not just use those to Really upgrade the infrastructure itself to plan better, to operate better. But in the meantime, while the utility is modernizing, yes, generators can …

AI assessment note: “All your transients, all your EMT, like electromagnetic transients, will get way more complicated”

Q a study Um, is it just that utilities are cycling through studies of individual feeders all the time, or is it like, hey, we noticed that there's a bunch of new load on this feeder, that, that neighborhood has a bunch of EVs that just suddenly showed up, we should do a study of whether we need to upgrade a substation. Like, what, what does the distribution side look like?

A Well, I think it's actually quite common across both, but the characteristics are different. So transmission system, uh, Naturally, it's one big thing, so every single time you study an entire region, I think for a cluster study or large load interconnection study, the grid impact and solution side, the base, I think most utilities are talking about six to nine months to perform one of those studies, and it's a one-time, right, so risk of restudies will compound that, and, um, and, uh, it does cost internal utility, typically a hundred, A couple of 100,000 dollars. The typical benchmark is about quarter million dollars per study, so it is a long time, and it is very expensive, but you study the whole system at once, and it's completely reactive and ad hoc, so you basically respond to an interconnection request, and maybe now we can cluster it, but I think clustering, it's sort of a shortcut because that the queue changes all the time, so even if you cluster it, I think it doesn't add any more certainty to the queue. Distribution, you have simpler studies, But there are places where it is actually more complicated. As I mentioned, data quality suffers so much more because the level of coordination between distribution utilities and transmission and bulk power and the market participants are not there. So you have less checks and balances and audit trails around distribution data…

AI assessment note: “I think it's actually quite common across both, but the characteristics are different.”

Q these studies that take a long time. That's kind of a, an outcome. The, the challenges that lead to that are data quality, particularly at the distribution level, the, you know, all the, um, different parts of the system being siloed and all the studies being siloed and so on. Like, what does it look like in a, Native AI world, and what's the fundamental challenge to get to that?

A So AI can be ambiguous. So I think let's look at general AI and where we have been focusing on, but I think the origins of Think Labs is to ask a big question, right? Can we automate the grid, whether it's planning or operations similar to how we can drive towards autonomous driving? So can AI help drive towards that efficiency and autonomy to handle the complexities, um, uh, with human in the loop? Uh, so human-centered AI is what we're looking at, um, So, but AI can be, has been historically used for things like data cleansing. Um, pattern recognition, right? Really good for data cleansing. Uh, forecasting is probably the dominant use case in AI. And most recently, there's a lot of general AI around, like, there's massive amounts of data. Can we retrieve all the records? Can we learn from manuals? Can we learn from previous rate filings? That's great. For us, we specifically ask, can AI help us plan the grid? Meaning actually to run a system study. How does generators and loads and battery storage impact the grid? If we need to invest in lines or connections, reinforcements, where do we do it? How do we do it? Uh, how much flexibility versus how much storage versus how much wires? That's a really complex problem. So can AI help us do that? And so for us, what AI is, is physics-informed AI. So can we teach the AI how the grid works? How power flows? How voltage, uh, traverses …

AI assessment note: “For us, we specifically ask, can AI help us plan the grid?”

Q we have the reliability that we do have, right? Like, we actually have quite good reliability, right? And it's worked for quite a long time. Do you think that that is because, As you said, basically the planning is built around, let's make sure we solve for the worst case scenario, and so is the net result of that that we have high reliability, but a lot of overbuild, basically?

A I, I do think the grid is the, as we call it, the human's most complex and largest machine is an act of miracle. Like, it is impressive how, how much reliability gives us, and how much we all depend on it for all Like livelihood, social, economic development, security, etc. And it's been keeping up doing well for what we need it to be. And I think redundancy is one thing, but it is an engineering marvel for its time. I just think it sort of failed to keep up with the digitalization and the modernization process, especially for the last, like, two, two decades or so. But build, we're very good at building infrastructure, like bare metal, Um, coppers and wires, oil in the, in the ground, et cetera. But running a digital grid as a software company, as a modern company, I think we are just in the beginnings of that phase.

AI assessment note: “redundancy is one thing, but it is an engineering marvel for its time.”