Nov. 13, 2025

AI Energy Forecasting: How Machine Learning is Transforming the Grid

Show Notes
Transcript

#319

The modern power grid faces an existential challenge: integrating massive amounts of renewable energy while keeping the lights on. Sean Kelly, CEO and Founder of Amperon, explains how artificial intelligence and machine learning are replacing outdated forecasting methods to predict energy supply and demand with precision. This conversation explores battery optimization, demand response programs, and why traditional grid models no longer work in a world of extreme weather and distributed generation.

Sean Kelly leads Amperon, a company providing AI-driven energy forecasting platforms for grid operators, battery owners, and energy traders. With nearly 20 years of experience in energy trading and risk management, Kelly has worked for major energy firms such as Tenaska, Lehman, EDF, and E.ON, where he led significant trading desk developments and transitions, including bringing nuclear power plants into new markets.

Key Discussion Points:

Traditional forecasting models look backward. AI models look forward and learn in real time, adapting to grid changes as they happen.
Battery storage is exploding. ERCOT expects 40% more storage in the next 5 years. Texas batteries run 1-2 hours, California runs 4-hour batteries, and companies like Form Energy are developing 100+ hour systems.
The critical hours are 6pm to 9pm. When solar goes down at sunset and wind drops simultaneously, batteries become essential. Amperon's platform tells operators exactly when to charge and discharge for maximum profit.
The grid needs 6 terawatt hours of energy storage but we're only one-tenth of the way there. Developers should build all projects battery-ready or include storage today.
Demand response works like insurance. Grid operators pay large customers to reduce load during peak times. Penalties are severe for non-compliance, ensuring the system works when needed.
Forecasting is no longer optional. Sean calls it "the operating system of the modern grid." Historical averages won't keep the lights on anymore.
Data centers and large industrial loads play a critical role in grid flexibility, backing down from 20 megawatts to 10 megawatts when grid operators need relief.

Connect with Sean Kelly

LinkedIn: https://www.linkedin.com/in/sean-kelly-0792626/

Support the show

Connect with Tim

Clean Power Hour
Clean Power Hour on YouTube
Tim on Twitter
Tim on LinkedIn

Email tim@cleanpowerhour.com

Review Clean Power Hour on Apple Podcasts

The Clean Power Hour is produced by the Clean Power Consulting Group and created by Tim Montague. Contact us by email: CleanPowerHour@gmail.com

Corporate sponsors who share our mission to speed the energy transition are invited to check out https://www.cleanpowerhour.com/support/

The Clean Power Hour is brought to you by CPS America, maker of North America’s number one 3-phase string inverter, with over 6GW shipped in the US. With a focus on commercial and utility-scale solar and energy storage, the company partners with customers to provide unparalleled performance and service. The CPS America product lineup includes 3-phase string inverters from 25kW to 275kW, exceptional data communication and controls, and energy storage solutions designed for seamless integration with CPS America systems. Learn more at www.chintpowersystems.com

The Clean Power Hour is produced by the Clean Power Consulting Group and created by Tim Montague. Please subscribe on your favorite audio platform and on Youtube: bit.ly/cph-sub | www.CleanPowerHour.com | contact us by email: CleanPowerHour@gmail.com | Speeding the energy transition!

WEBVTT

00:00:00.000 --> 00:00:30.379
Forecasting is no longer a nice to have. It's literally the operating system of the modern grid. The traditional models look backward. Ai lets us look forward and learn in real time, and so these historical averages just they're not going to keep the lights on, and we don't always keep the lights on. But as someone who's struggled with this over the last four years, keeping the lights on at their home. This is something that we really, really need to pay attention to in this new world of extreme weather and distributed generation.

00:00:30.379 --> 00:00:45.460
So are you speeding the energy transition here at the Clean Power Hour, our host Tim Montague, bring you the best in solar, batteries and clean technologies every week want to go deeper into decarbonization.

00:00:41.320 --> 00:00:55.600
We do too. We're here to help you understand and command the commercial, residential and utility, solar, wind and storage industries. So let's get to it together. We can speed the energy transition

00:00:57.460 --> 00:01:06.780
today on the Clean Power Hour AI forecasting for power markets. My guest today is Sean Kelly. He's the CEO of amperon. Welcome to the show, Sean.

00:01:07.319 --> 00:01:08.879
Hey. Thanks so much for having me. Tim.

00:01:09.780 --> 00:01:24.140
It's great to learn about this industry, because I don't know very much, so I'm I'm really all ears on this one. But why don't you start with a little background on yourself, and then we'll dive into what amperon is up to in the world.

00:01:24.680 --> 00:02:31.280
Yeah, absolutely. So I've been in power markets for my entire career, since the two weeks after I graduated from Texas, A and M walked onto a trade floor at tenaska. At tenaska started in ERCOT, the Texas market, happened to live, live here most of my life, and I'm currently residing in Houston, Texas. And went from tenasco back to Houston, and was at a great company called Eagle Energy Partners, which got bought by Lehman Brothers. I think we all know how that one worked out. And then EDF electricity, France trading came in and bought us. I guess what makes me uniquely qualified in this is I've run more than three dozen power plants, everything from wind and solar to five nuclear assets, including helping with the acquisition integration of two of them, Nine Mile and g'day in New York. Also moved to Chicago in 2013 to help set up eon, the big German utilities trade floor, as the first trader in North America.

00:02:27.020 --> 00:02:39.080
So yeah, been a really fun career on the trading side, which is really helps me understand our customer types.

00:02:34.219 --> 00:02:47.259
And so I was living in New York, moved there in 2017 for a girl who's now my wife, one of my best trades, or my best trade all time,

00:02:48.939 --> 00:02:50.439
that's awesome.

00:02:48.939 --> 00:02:50.439
Yeah.

00:02:50.439 --> 00:03:00.840
And in 2017 you could walk around New York and almost trip over a data scientist. They were they were everywhere, data engineers.

00:02:58.000 --> 00:03:24.860
Everyone was a co founder and disrupting and all of the buzzwords, and that made me realize that all my friends in Texas probably needed to be on top of this latest and greatest AI machine learning and and so the joke between my co founder and myself is he'll build it and I'll sell it. And so that's what we've been doing here for almost the last eight years at amperon.

00:03:25.639 --> 00:03:32.780
My gut tells me there's a lot of parallels between financial trading and energy trading. Is that accurate?

00:03:33.620 --> 00:04:00.159
There are definitely there. There are parallels, and then there are also diversions between the two. But I mean trading is trading. It is always buy low, sell high. And a lot of the financial trading and electricity is very has a lot of similarities to trading other commodities. I mean a little bit of equities, but you see the same different options available and things like that. So, yeah, it's a it's a really fun field.

00:04:00.159 --> 00:04:08.219
I was the weird kid who wanted to be a trader from the time they were about eight years old.

00:04:02.879 --> 00:04:10.800
And here I am, guess that's just a few years later, still playing with trader.

00:04:11.699 --> 00:04:17.939
And what is the status of amperon? How big is the company? And what markets are you serving?

00:04:18.600 --> 00:05:01.019
Yeah, so ampron is the leading energy forecasting company, operating at the intersection of AI and energy data. We have raised 38 million in venture capital. We have a little less than 100 employees, and we are active in 23 countries with offices in both Houston and London, but cover the entire continental 48 parts of Canada, Australia and almost 20 European countries. So just actually signed our first deal in the Middle East. And so we've built a truly global company that can stand up forecast anywhere. Because, as you very much know, wind and solar works in any country. And so we our forecast works in any. Country as well.

00:05:02.939 --> 00:05:05.339
All right, cool.

00:05:02.939 --> 00:05:09.300
Well, what is something about the power market that most people completely misunderstand?

00:05:09.300 --> 00:06:11.639
Sean, yeah, when I started in my career, people asked why I was working at night. And electricity is a physical commodity. These megawatts move across wires in real time. It can't it can kind of be stored in batteries, but it's not like an oil or a natural gas where it can be fully stored throughout an entire season. And so back to what we talked about a minute ago. They people do think of it like stocks or crypto. It's fundamentally different. It's actually more volatile. It's the most volatile commodity out there, as we've seen in some of these crazy events, whether it be polar vortex, winter storm, Yuri winter storm, Elliot and so yeah, it's it's definitely misunderstood, but it's also it's not going anywhere. Every single industry is extremely tied to electricity, as in, without it, no other industry exists. So in my unbiased opinion, it is the most important thing that we should be focusing on.

00:06:12.420 --> 00:06:54.459
Yeah, it's so true, we take electricity for granted, but literally, the good life completely depends on it, and the good life goes away pretty much instantaneously when electricity goes away. It's a double whammy, too, right, right? Because then when there's a big outage, like you're you mentioned in Texas back in 20 what? 2221 electricity prices spike, because it's a very rare commodity then, right? And so it's also super painful for those who have electricity. It's not just those who don't have it, but Well, what is, what is the problem, I guess that amperon was created to solve.

00:06:55.600 --> 00:07:39.860
Yeah, so in 2017 when I met my co founder and we started the company in January of 2018, we knew that there was a lot of changes coming down the pipe. Smart Meters had started getting adoption. You used to have back in the back in the glory days, the meter made breaking in your backyard, getting bit by the dog. Now you just drive around with the truck and you get on a 15 minute read from someone's home. We knew that was happening. We knew renewables were happening some behind the meters, such as solar. You've got places like Australia, which are all in on that, and then Texas and California and other western states that are also adopting it. So knew that was coming.

00:07:34.819 --> 00:08:24.860
EVs, I mean, the Tesla was popular. Other companies were talking about having EVs. Here we are in 2025 and they're absolutely everywhere that changes what your electricity profile looks like. So we knew all this was coming. I also knew that much of this was being done in Excel, and there were some of the largest utilities out there that were running these monster models overnight and just praying and hoping that they would get an answer the next morning, but oftentimes the macro would break because there's so many lines of just lines on lines on lines in these Excel spreadsheets, and that's not going to work. And so amperon was built to deliver a real time AI driven forecast so that these traders and utilities can make better decisions.

00:08:25.879 --> 00:08:50.679
Now, back in 2018 the definition of AI was maybe different than it is today. I mean, most knowledge workers and prosumers are aware of llms, and maybe, if you're knee deep in it HRMS, which is a new model that just emerged. But these large language models, do they have anything to do with the technology that you use at amperon?

00:08:51.759 --> 00:09:03.360
I mean, we look at them from a basically workforce efficiency standpoint, and so that is something we're using just again to speed things up.

00:08:59.919 --> 00:10:29.059
Got a very talented engineering team, data science team, they're always looking at the latest and greatest to continue to speed up their efficiency, how many, how much they can crank out in the in the hours they've committed to the team. And the way we've looked at it, though, is is the really the machine learning standpoint? And so our first model went live in November of 2018 and machine learning was involved in that very first model. And again, forecasting is hard. We're still always looking at and tweaking the models, but our models have retrained every hour since November of 2018 that is a lot of compute cost, and so we've continuously looked at that. We also know that one forecast is not always right, and so we're always running an ensemble of four to six different forecasts. And so looking at the simple regression to gradient boosted trees to whatever the latest and greatest is. And. That's why we're really fortunate to have a team of about a dozen data scientists, mostly PhDs, just always dealing with this problem because it's hard. I don't know what's happening in your house. I need to know when you bought an EV I need to know if you have solar panels. I need to know what a battery is doing. This problem is getting so much more complicated than back in the day, when everyone went to work at eight o'clock, they came home at five o'clock, every home profile looked very similar.

00:10:25.580 --> 00:10:35.179
Now, home profiles are all over the place, and so we we've got to stay on top of that for our 150 plus clients.

00:10:36.019 --> 00:10:53.500
So let's talk about this from the perspective of, say, an IPP, a company that owns fleets of wind, solar and battery projects, and let's talk about Texas, since you you know that market certainly very well.

00:10:47.860 --> 00:10:58.779
ERCOT is a unique market. It is now on a windy, sunny day, 50% wind, solar and battery powered.

00:10:58.779 --> 00:11:32.480
Who knew? Right? Ginormous. Red State is going green. Texas eclipsed California on an on an annual basis. Now, in terms of the install solar, right? We're, I think Texas did 11 gigawatts last year, something crazy like that. But if you're an IPP, okay, you're in the business of selling electricity to various and sundry off takers. Walk us through that business and what are some of the challenges and opportunities that they're faced with on a daily basis.

00:11:33.200 --> 00:12:01.980
Yeah, Texas has done a great job of just really being a pro business state and getting an entire full text or full generation stack. And so, I mean, we have 40 gigs of wind and a ton of solar. And so for these IPPs that are our customers, independent power producers, for those of you who aren't all in the acronym weeds, what they do with our forecast is they give us said asset.

00:11:57.279 --> 00:13:11.759
We'll pick a solar asset. And so we'll give them a 15 day, 360 hour forecast for their specific solar installation, and we'll call it 100 megawatts. And we actually give them five minute granularity, because, as we all know, Sun kind of does what it wants to. It's more reliable than when from knowing when it's going to be sunny versus not, but we've got to stay on top of this. So a five minute forecast is the bare minimum that you need to actually understand what you're doing with this, and it's what you can actually go and schedule. And so we tell you on this five minute granularity what it looks like, so that you can tell again, what was we're picking on, ERCOT here. ERCOT, what you're going to be showing up within a day ahead. And so every day, by about 10am you need to tell ERCOT, hey, tomorrow I will have this much solar. And so that's what amperage forecast provides to you. And then as we continue, the models rerun every single hour, and you're getting an update there. So very similar with wind we're telling you, we're updating like multiple times per hour, on five minute cadence, what your wind farm is doing so that you can best monetize it and run it accordingly.

00:13:12.120 --> 00:13:18.539
And does the IPP know, though, from that model, what kind of revenue they're going to be generating?

00:13:19.320 --> 00:13:46.659
Yeah, they can take, they can take our output that we're telling them, and then basically put it with the the clearing price, whether it's the day ahead price, if they bid it into the Ford market, or the real time price, if they if they took it, if they floated it to the intraday, and then they can figure out what their revenue generation is going to be in figuring out how much money amp runs helping increase on a return on investment.

00:13:47.259 --> 00:14:25.940
So what is let's just step back a little bit, because I think obviously some fraction of my listeners will be familiar with this energy trading business, but only some fraction so in the greater scheme of things, IPPs are signing what kinds of contracts to they need off takers, right, to develop the project out of the gate, frankly, right? And, and who is, who is a good example of a buyer? And then what are the what is the variability in that contract?

00:14:28.940 --> 00:14:42.460
So what they're going to do is, you're absolutely right. If I want to build 100 megawatt wind farm to keep nice round numbers, I want to finance it and not have to foot the entire bill up front.

00:14:39.799 --> 00:15:14.879
And so the best way to do this is sell a certain percentage of my of my wind going forward. And so they'll do a PPA power purchase agreement. And we'll say a 10 year PPA is what they're looking at. The most natural off taker is municipalities. Cooperatives, so towns. So they'll go in and say, hey, my town would love to buy the 100 megawatt wind farm. 100 megawatt wind farm does not produce 100 hours or 100 megawatts every hour, right?

00:15:14.940 --> 00:16:15.240
It's not a base load. It's probably never going to produce 100 megawatts. That's just the the nameplate capacity of it. So I'll go in and say to Tim city, I'm going to give you 30 megawatts, like around the clock going 10 years. And so I would sell you that, and you would say, Great, I have 30 megawatts coming. Well, you then get the money from that, and you're able to go build the wind farm agreement reached. Well, then you've got to deal with the imbalance. And so if only 12 megawatts are going to show up, I own Tim City, 18 more megawatts. And so that's where I need to know, with amperons forecast, that I've only got 12 megawatts tomorrow. It's just not a windy day. I need to go buy that like that interval of 18 megawatts from the market. So that's how, that's how these are being built, and these are how these are being financed. And again, the the PPA market, it's not, it's not an exact science.

00:16:15.240 --> 00:16:28.100
It takes away a lot of the commercial upside. However, unless you have just the ability to fork over the full amount whenever you build it. Then, yeah, it makes a lot more sense to go finance that PPA. And

00:16:28.099 --> 00:16:42.278
then, as the asset owner, I'm getting a guaranteed price on the wind power, but I'm buying grid power at some variable price, and that's also where this kind of a platform comes in handy.

00:16:42.578 --> 00:17:44.499
Yeah, you agree to said price, and that's where you look at what the Ford curve is trading. So you'd say, hey, 2026, through 2025, is at this number, and said like municipality or Co Op will go buy it. You. I guess examples of other people buying this, or obviously, utilities will go and buy this. Oftentimes they go build their own fleets, just because they're normally quite well capitalized. And so they'll say, hey, I need to have this for my rate payers. And then you'll also have now, what's really exciting is you've got data centers and hyper scalers who are coming in because they're normally in very good financial shape, and so they're able to come in and really incentivize this. And many of them, as we know, have made green commitments, and so they're more they're more likely to be excited about the wind and solar if they have a green commitment. But they're also trying to figure out just, you know how to, how to make sure that said data center can can run as needed.

00:17:45.400 --> 00:18:19.319
And how risky is this business in the greater scheme of things, you you know you mentioned, okay, if I have 100 megawatts of a wind farm, I might be able to sell a contract for 30 megawatts, but the wind goes up and down. Some days are it's blowing, and some days it's still for a week, it could be multiple weeks where there's very little wind, and then I'm buying power at some not necessarily known rate. So yeah, just how risky is this business in general?

00:18:19.319 --> 00:18:33.380
I mean, power is a risky business. I mean, again, I've been, I've been going through this for 20 years. I mean, you looked at summer 2011 in ERCOT, there was not a lot of wind. It was extremely hot.

00:18:30.259 --> 00:18:33.380
Power prices were very high.

00:18:33.740 --> 00:19:45.099
2014 you had the polar vortex. I lived in Chicago then, and that was a rough one. Being a Texan in Chicago, dealing with like minus 40 feels like in Fahrenheit. And so again, not the renewable generation didn't show up as much then. And so this is why having a 15 week forecast is so important. We're able to go out and for instance, amperon, you mentioned winter storm area earlier, so that that was a little too near and dear to me, from the standpoint of my wife, was six months pregnant at the time, and the power went out at 2am on February 15 of 2021 we had no marketing department at the time, so I wrote a really fun blog post on Valentine's Day, ideas to do in the dark, because February 14 is what day I thought the power was Going to go out. Went out at 2am on the 15th. And so we were able to tell our clients on February 3, our meteorologist nailed it and say it's going to get really cold. So multiple IPPs went and bought power at $60 per megawatt. It wound up clearing$9,000 per megawatt those days.

00:19:39.980 --> 00:20:11.819
So because these different wind farms went and bought power, because amperon, they did their own research too. We don't give full credit here, sure, but they, but they went and bought$60 power, and then it traded all the way up to 9000 and so that's the way that amperon Really. Really helps them understand that, because they knew they weren't going to be able to supply all the PPAs that they had previously agreed to that week of February 15, and lo and behold, yeah,

00:20:12.000 --> 00:20:15.480
but you literally were able to predict that the grid was going to fail.

00:20:17.880 --> 00:20:45.039
That's not exactly what we do. I knew from my experience that the grid was going to fail, and literally put out a blog post about it, or a LinkedIn post about it, that we were going to lose power, and we lost power, and then on. But we we put that it was going to be a very, very like epically cold event, and we started saying that at the very beginning of February. So we gave our clients, I mean, 12 days notice.

00:20:45.039 --> 00:20:56.680
We'll say 10 days notice, because it really got dicey on about the 13th on that Saturday is when prices went well into the 1000s, and then it just continued on. It was a one heck of a Valentine's Day weekend.

00:20:58.180 --> 00:21:05.519
And so who are the winners and who are the losers in a stormy area scenario?

00:21:01.200 --> 00:21:15.180
Obviously consumers are, are the big losers, consumers and business owners like not having power is painful and uncomfortable, but yeah, just walk us through that.

00:21:16.200 --> 00:22:07.619
Yeah, there were so many losers in that. And I would say the the biggest losers, you nailed it, definitely consumers, those rate payers at the end, there are people who are on variable contracts at their home that were literally getting their bank account hit, like, every like, every 15 minutes that were on auto pay from a company that's no longer around called gritty and so, like, you literally, like those were the biggest losers. And the fact was, the consumers, the business owners, obviously you lose those days of doing business. And then I would also say that anyone, anyone, I mean, natural gas, didn't have natural gas plants, also didn't all show up. So they had contracts to sell that they were they had to go buy those back. And that wasn't fun.

00:22:03.480 --> 00:22:25.160
Getting gas wasn't fun. There weren't a whole lot of winners out of that. I mean, it definitely put the state in a pretty bad situation. I do think it made people more resilient, because you look in the temperatures we had, I mean, as you're in Illinois, they're like, teens. Yeah, no big deal.

00:22:20.640 --> 00:22:53.559
Nothing to see here. I mean, when I lived in Chicago, we had teens for like months at a time, that 2014 and that 2015 but in Texas, our everything wasn't winterized properly, because we never seen teens before. And so now, I mean, I think ERCOT and the PUC and the rest have done a really good job of pushing for that winterization is now whenever we have a slight scare, it winds up being or so far, like knock on wood, has been pretty much a non event.

00:22:54.279 --> 00:23:56.440
The Clean Power Hour is brought to you by CPS America, maker of North America's number one three phase string inverter with over 10 gigawatts shipped in the US. The CPS product lineup includes string inverters ranging from 25 kW to 350 kW, their flagship inverter, the CPS 350 KW is designed to work with solar plants ranging from two megawatts to two gigawatts. CPS is the world's most bankable inverter brand, and is America's number one choice for solar plants, now offering solutions for commercial utility ESS and balance of system requirements go to Chint power systems, com or call 855-584-7168, to find out more. Yeah. All right, I could talk about just that one event, probably all day. But let's move on. So what sets what sets April and apart, though from other forecasting companies,

00:23:57.160 --> 00:24:06.359
yeah. I mean, what sets us apart is accuracy. I mean, we're two to three times better than grid operators.

00:24:06.839 --> 00:24:32.180
ISOs. TSOs is what they call them in Europe. We also, we cover the whole value chain. We forecast demand. But demand is not the whole story. It's net demand, as you and your listeners know, which net demand is pulling out renewables. So it's demand minus wind minus solar, because those show up and and make it much easier on the grid. And we also forecast prices, and so you're able to see this across the value chain.

00:24:32.480 --> 00:25:19.380
And then this is also built by traders and data scientists. Our first hire was a PhD data scientist who's still hanging out with the company over seven years later, and then my second hire, which the investors thought was a little weird, was a meteorologist. And why a meteorologist? Because forecasting, weather is so, so important to this. And so those are some of the things that setting up. Run apart, but also we get a little bit of a late mover advantage. Like I said, we started the company in January of 2018, so many of our competitors are 20 or even 30 years old. We've been cloud native from day one. We've been using AI since day one. This wasn't, Oh, we've got a good company. We've got clients.

00:25:19.380 --> 00:25:44.500
We've got to move out of, like, we've got to move out of just running this off our own servers. Oh, we've got to move this out of Excel. Oh, we've got to move this out of we started with a modern tech stack from day one, and so that's given us a huge advantage and an advantage that we don't take for don't take for granted, and wake up every single day trying to figure out how we can improve our models and pay attention in the next, latest

00:25:45.400 --> 00:25:53.380
and greatest. And are companies building their own systems for this? Or is it all third party platforms?

00:25:54.039 --> 00:26:04.319
It's definitely both. I mean, people will look at it as a core competency, especially utilities, but normally they also want to have a third party come look at us.

00:26:04.980 --> 00:26:21.859
Learn look at it. Their forecast. The reason why is, as I mentioned, we have 150 clients and over 40 million meters in the continental United States that are in some way on our platform, either as individual meters or in like aggregations.

00:26:22.160 --> 00:27:07.500
And so there's no utility that it has 40 million meters. And so we are able to have already seen some things. So when we walk into utility number four, we've already dealt with utility one, two and three, and so we already know what their problems are and what they're seeing. And so it really helps that kind of been there, done that before, as opposed to having to start from scratch. So a lot of people will still run their own internal models. We're going up against a lot of 20 year old models that they just don't trust because of how times have changed, but they'll run those with ours. And as I mentioned, we're a big, big fan of ensembles, and so ensembling is basically just waiting different forecasts together to make one super forecast. In layman's terms,

00:27:09.240 --> 00:27:26.059
I'm curious if you know the weather's changing, and you can no longer say, well, let's just look back at the last 10 or 20 years to predict how the weather is going to be in February of 2026 is that a factor?

00:27:26.059 --> 00:27:43.359
Now, absolutely. I mean, it is. Weather's gotten crazy. I mean, we're sitting here talking about a one and 100 year event. I mean, I remember early in my career, all we were worried about was hurricanes.

00:27:40.839 --> 00:28:31.160
And the reason we were worried about hurricanes is most natural gas came from the Gulf, and when a hurricane came in, it took out gas production, and natural gas went through the roof. That happened in 2005 right when I started. And now, when a hurricane comes, natural gas actually dips, because it takes out load, and we don't really drill offshore that much, and so that was really kind of the big change. Now we're sitting here, and I can rattle off. I mean, all of the events starting, I mean, ERCOT, 2011 was 62 days of over 100 degrees in Dallas. I mean, we've talked about 2014 with the polar vortex. We've talked about 2021, with winter storm Yuri. I mean, blacked out a whole bunch of people on Christmas Eve across the east coast with winter storm Elliot.

00:28:31.640 --> 00:28:48.220
I didn't know what a derecho was until last May, when it came and took out a bunch of power lines in my neighborhood and I didn't have power for seven days. I mean, you just continue to see these hurricanes are getting bigger every year. It feels like we're saying it's going to be a wilder hurricane season.

00:28:48.940 --> 00:28:56.799
Wildfires are more prevalent than I can ever remember, just in the last handful of years. So I mean, things are getting wild.

00:28:56.859 --> 00:29:06.839
Extreme events are here. And I mean, you just off the cuff, reference a one in 100 year event. But is it really a one in 100 year event? If it happens every two months?

00:29:10.019 --> 00:29:16.799
Yeah, it is crazy.

00:29:10.019 --> 00:29:16.799
So talk to us about forecasting.

00:29:16.799 --> 00:29:26.180
How is that the role of forecasting evolving as the grid gets smarter and more dynamic.

00:29:21.619 --> 00:29:37.940
You know, you mentioned electrification of transportation, data centers coming online, and, you know, the economy is growing, and it's electrification of everything.

00:29:34.339 --> 00:29:45.940
It's not just EVs, it's heat pumps, HVAC, industrial processes. So how is your How is your business evolving?

00:29:47.019 --> 00:29:57.940
Yeah, forecasting started as a nice to have and that, and now we're sitting here, and it's literally the foundation of daily operations.

00:29:53.920 --> 00:30:20.099
The reason it was a nice to have is back to what I referenced earlier. Smart meters is you just got a data point every 3060, 90 days, whenever the Meter Reader came around, so you didn't really know how imbalanced you were. Now we're sitting here and we're getting, I mean, some places are getting millisecond reads, a nest thermostat or something. Of that is picking up data all the time.

00:30:20.279 --> 00:31:40.720
So the amount of data that we have to take in has gone just absolutely through the roof. And so every single trading strategy, dispatch decision and even long term defense investment depends on it. And so if you think of forecasting, it's really the operating system of the modern grid. And so what we're doing, we knew it was going to be important. But you also, I mean, you referenced not being able to have look at the last 10 years of weather forecast. We also can't look at the last 10 years of load data, because I didn't know when I started this that covid was going to be a thing. It changed what a landscape of a city looked like for as little as six months to as much as three years, just depending on what the like, what the lockdown, what rules were when people went back to work, and now we're sitting here in a in a remote economy, and we have people in, I think, 17 different countries who work at amperon because We get to hire best in class, but they work at their house, so their house doesn't look like it would have in the 1980s and so these are all things that I mean, all things that are changed that when we started the company, we couldn't have realized we're going to hit but continue to evolve and make things more difficult, Which means we have to continue to get smarter.

00:31:43.359 --> 00:31:56.319
And what are the what are the limitations that you encounter with traditional forecasting methods, and especially with regard to AI?

00:31:49.779 --> 00:31:59.680
How have you leveraged AI to transcend that legacy technology?

00:32:00.460 --> 00:32:29.779
Yeah, and, and, and, just to be clear, we still use the legacy, I mean, linear regression models have been around for quite a while, and we still have that as a very core input into our models. But again, being able to automatically weight each model and having it retrain every single hour is not something that you can go do in Excel for 40 million meters. It's just not computationally possible. And so that's where the the AI in ML machine learning come into play.

00:32:30.019 --> 00:33:10.799
And so we're we're looking at that. We're also looking at weather. Weather has made huge strides here, just in the last two years. Whether it be from like the models that NOAA puts out to you've got the like, European AI models with much better granularity. So weather is something we look at. We currently use for weather vendors, and we're always looking for like, the next, latest and greatest. So I mean, weather is something that I'm very, very I mean, it gets more complicated and continues to get more complicated, so you need better models just to stay at where you were if things continue to progress like this.

00:33:06.539 --> 00:33:30.079
So that's where the AI really drives in. So weather has been a huge help, and that's why we literally, we've got a team that goes in and tests out all the latest and greatest AI models to see which one's going to hit is, I mean, the the people with the most money in the world, the the NVIDIA is, the Googles, etc, the world are putting out their own models, which is definitely, uh, definitely worth diving into.

00:33:31.099 --> 00:33:51.819
So tell us a little more about the Machine Learning, though, to the extent you can that you're leveraging it's, it's built on top of some other technologies, like open AI or Claude or Google. Yeah, I'm just curious how that, how that works?

00:33:52.240 --> 00:34:18.840
Yeah, so it's, it's not built on top of LLM technology, like some of the ones you mentioned, what it's really doing is going in and waiting our forecast and running different forecasts and saying that, hey, this like grading boosted tree needs to be with this deep learning model. We'll put in 15% of this one and 17% of this one. And then also different clients are different.

00:34:14.219 --> 00:35:03.900
Some, especially in the retail energy provider world, will be all residential. Their book looks very different than another client who's all commercial and industrial. So we again, don't have someone sitting there saying, oh, make sure to use the commercial model. It just automatically goes in and uses the commercial or the industrial model or the residential model. So again, you don't want someone sitting there across 150 customers trying to figure out which model to use at which time. And so that's where the machine learning really kicks in. And again, the data science team has done a phenomenal job of doing this since inception. So I'm glad this isn't something that we kind of freaked out about a year ago and said, Oh my gosh, we have to use AI in some. Thing.

00:34:58.719 --> 00:35:58.840
One of the coolest, I guess, accolades that we've received is Andreessen Horowitz, who we have no money from, put out their top 50 AI American dynamism companies. And right when AI became a thing about two years ago, and there was one company from Texas, and I believe there was about, there's like, eight to 12 energy companies, and we were, we were on that list. And so again, we've been using this since very inception, and the machine learning makes all the difference. That's where you get to go against a competitor running in Microsoft Excel. And that's also something that putting on an updated model, as quick as weather changes, and as quick as sun and wind and all of that changes, as opposed to so many of our competitors, we're literally running their model once a day, which just does cut it once or twice a day. And that that's that's not going to work with how fast weather is changing these days.

00:35:59.980 --> 00:36:08.519
So when, when the next storm Yuri comes? There's, I guess there's a couple of things on my mind about this.

00:36:09.059 --> 00:37:05.519
The grid is changing. We are hardening the grid now in Texas, we're also increasing the attachment rate of batteries, and the percentage of wind and solar on the grid is going up, but the attachment rate of batteries is going way up. So there's that. So the grid is more resilient. And okay, maybe our maybe our forecasting of weather patterns is getting better. So walk us through a scenario for another Yuri style event. How much warning does the grid operator have, and how much, you know, flexibility do they have? Like, one of the things that's noteworthy about ERCOT is it's it's famous for being isolated from other ISOs, and so they can't necessarily count on getting a bunch of power from some other pool when the shit hits the fan.

00:37:07.320 --> 00:38:01.199
Yeah, I mean, and that's where, I mean, that's where having reserves really kicks in, and incentivizing the reserves to act at the correct times. And then that's also where demand response really kicks in. And so a lot of the load that we're talking about, right? You can't, you can't get through a conference. Can't get barely get through a podcast without talking to demand growth. So might as well just go for it. But a lot of the demand growth that we're looking at, I mean, yes, we're seeing heat pumps, we're seeing EVs, we're seeing demand growth at your own home, but the real load growth that we're seeing is data centers and large industrial I mean, again, more more things on shore and just that stuff's pretty flexible at the end of the day. And so that's something that I think that the grid operator is doing a good job of paying attention to how much the grid operator gets involved.

00:38:01.440 --> 00:39:39.920
There's something that just passed in Texas called SB six, and that is basically really pushing on data centers to have to behave whenever times get tight. And so I think that's really the next step. So ERCOT, you're absolutely right. Is more of an island than most places, but I think they have taken the proper responses from a demand response, and incentivizing those large loads to be on good behavior and be good stewards of the grid. So that helps. And also, if it's during the summer, it's hot out, and that's when the solar is going wild, so you've got a few hours at the end of the day, kind of hours in the 1921 which is 7pm to 9pm at home, and those hours, that's when the batteries come in. So having those, although it'd be only one to two hour batteries, ERCOT spreading those out and kind of saying, These are the hours I need you that's really critical to keep the grid running. But ERCOT had a smooth summer PJM less so I think they had a dozen EA one alerts, and they also had something that in June, they blew through what they said was their peak load for the entire summer, and that's just because they have such a wide footprint that they did not expect that it would get hot from Chicago all the way to Virginia, and it did. So it had a heat wave that just kind of sat over that whole Mid Atlantic region. And so PJM saw that. But for us, that made people realize that you can't sleep on PJM forecasting, either or New York forecasting, because it was a hot, hot summer. And those, they've had some pretty mild summers for the last handful

00:39:40.340 --> 00:39:49.599
so the grid operator sees these events coming, right? They say, Okay, it's going to be super cold.

00:39:49.599 --> 00:39:58.000
People are going to be running their electric heaters, their gas furnaces, whatever, right?

00:39:53.199 --> 00:40:19.079
They have some ability to predict what the load is going to. Be over time. And then if they notice that the load is going to be greater than what they capacity of the grid is, they will send a signal to large off takers and say, hey, at, you know, four o'clock tomorrow, we want you to throttle your data center. Is that kind of how that works?

00:40:19.440 --> 00:40:21.679
Yeah, you, I mean, you pretty much got it there.

00:40:21.739 --> 00:40:36.920
They go through and they say, All right, the supply stack is this, and we think demand is going to be this. And this is why having a good wind and solar forecast is so important as you can, for the most part, tell a combined cycle when to show up.

00:40:36.920 --> 00:41:03.599
I mean, not a lot of coal left, and not a ton of nuclear unfortunately, but those are pretty reliable base loads, and so you've got to understand what the difference is. And so yeah, you tell people who are enrolled in these different demand response programs to, hey, 4pm that's your time. Like, go ahead and plan ahead. Go ahead and send your workers home. Like, I need you to do what you can do.

00:40:59.800 --> 00:41:03.599
So yeah, and

00:41:03.599 --> 00:41:25.159
then if they don't do what they can do, like if that demand response, for some reason, fails, does the grid? Is the grid smart enough or savvy enough that the grid operator can go, Okay, I'm going to shut down this, this circuit to this major off taker so that I don't have an uncontrolled outage.

00:41:25.880 --> 00:41:59.079
So the way demand response works is it's like insurance. And so I'm going to pay you $10,000 to be on good behavior when I tell you to be on good behavior. And so the penalties are nasty if you are not on good behavior. So if you're sitting there and you're running a data center, and you promised that you'd go from 20 megawatts an hour to 10 megawatts an hour, and I already paid you your $10,000 it's way more than that, but I already paid you your money to back down. Then you will back down.

00:41:55.659 --> 00:42:10.559
And so that's where. That's where. Again, there's not really, I see behaving on the grid and very similar in, like, especially in the East, where you have capacity auctions.

00:42:08.039 --> 00:42:22.400
That's, I mean, the capacity auction was obviously a record in PJM, and so they are saying, like, you are promising that you will back down then. So, yeah, the you don't have to worry about the good behavior. The good behavior has already been paid for.

00:42:22.760 --> 00:42:33.559
All right, so let's talk a little more about storage. You know, storage is kind of the new kid on the block, but there's lots of it, you know, and lots more coming.

00:42:34.280 --> 00:43:04.380
I think I saw that there's 40% more storage coming in the next five years in ERCOT. And of course, ERCOT is just one market. There's incentives in Illinois, New York, Massachusetts, New Jersey, Connecticut. I mean, many states now California have programs incentivizing the installation of batteries. What should our listeners know about batteries as this relates to what amperon does.

00:43:05.340 --> 00:43:59.019
Yeah, I love batteries, and the reason why is early in my career, I got to manage part of the generation fleet for Cobb County, so Atlanta, Georgia and we had a bunch of hydro as part of Southern portfolio. And hydro functions very similar to batteries, in that you charge it, ie pump storage during low price hours, then discharge and use that generation during high hours. So I've the battery concept, to me is made sense my entire career, and both ERCOT, California and other places are doing a great job, because batteries and storage in general is unique, because it's supply and demand. And so the storage lets us shift energy over time, not just balance it instantaneously. And so as I mentioned earlier, oil, gas, you can store for a season.

00:43:59.019 --> 00:44:06.659
Electricity. That's not the case, but batteries are slowly but surely helping us get there.

00:44:02.219 --> 00:44:53.980
I'm very intrigued to see what some of the companies like a form energy trying to do, like 100 plus hour battery are going to do, but, but what we have right now in Texas is one to two hours and in California, four hour batteries. From amperon standpoint, we're going to tell you what hours to use those batteries. We have well over a dozen clients who are battery operators on our platform. And so what they're looking for is they're looking for a time when so sun goes down at 7pm ish, it's summer right now. We'll call it right around there. And if wind decides to go down with it, that's when things are going to get dicey. Because it's 7pm families are still home. Kids are up. You're using that. Your office also might still be on.

00:44:51.099 --> 00:45:04.920
Your workplace might still also be on. And so those are the hours that used to not get as much love, but now we're. Delay from that like 6pm to 9pm is where you're paying attention.

00:45:02.400 --> 00:45:55.780
So you're going to look on amperage platform and say, Man, wind goes away the same time solar goes away and loads pretty high. This is the hour. And so that's what I mean. Most battery upgraded optimizers in ERCOT are using our platform for so it's a it's a new arbitrage opportunity. And again, I love when technology can stand on its own legs. And this is where you want to make batteries as much as possible, because back to what you're referencing, we need the more the merrier. We need batteries to be paired with renewable generation to make the grids go and for operators, this gives them more flexibility in balancing these renewables and also just smoothing out peaks so it but it makes Sam pron more important as well. So don't hate that, because timing and forecasting become everything.

00:45:57.099 --> 00:46:21.925
Hey, guys, are you a residential solar installer doing light commercial but wanting to scale into large C&I solar. I'm Tim Montague. I've developed over 150 megawatts of commercial solar, and I've solved the problem that you're having you don't know what tools and technologies you need in order to successfully close 100 KW to megawatt scale projects.

00:46:22.465 --> 00:46:29.844
I've developed a commercial solar accelerator to help installers exactly like you.

00:46:25.344 --> 00:46:51.429
Just go to cleanpowerhour.com click on strategy and book a call today. It's totally free with no obligation. Thanks for being a listener. I really appreciate you listening to the pod, and I'm Tim Montague, let's grow solar and storage. Go to clean power hour and click strategy today. Thanks so much.

00:46:44.889 --> 00:48:31.570
Yeah, I think I wrote in 2024 that the grid needs about six terawatt hours of energy storage, and so we are, we're only like 1/10 of the way to that goal, so to speak, and that earn, save, protect, tripartite that I drill into my listeners heads. I think everybody who's a regular listener knows now earn from grid services, save by attacking demand and capacity charges and protect from grid outages. But so storage is super powerful technology, and as prices have come down, it's making it much more realistic to adopt the technology. You know, there was just an article by John Weaver, one of my co hosts here at the Clean Power Hour, saying that there's good evidence that developers should build all of their projects now battery ready, if not including storage today, make them battery ready so that you can affordably And quickly attach storage because the price is coming down, and the service, the value of storage is so great, and of course, we have an extended ITC on batteries. Well, what else in our last minute together? Sean, should our listeners know about amperon? I really enjoyed this and appreciate you coming on the show. I definitely am thinking a little deeper about energy trading. And I still don't understand what is machine learning when it comes to energy trading, but we'll have to save that for another time, perhaps.

00:48:32.889 --> 00:48:58.195
Yeah, I mean, I have really appreciated the conversation. I'm definitely going to use the the earn, save, protect, that's that's great, and from, from a takeaway, from my standpoint, I mean it, this is no longer forecasting is no longer a nice to have. It's literally the operating system of the modern grid. The traditional models look backward. Ai lets us look forward and learn in real time.

00:48:58.434 --> 00:49:27.264
And so these historical averages just they're not going to keep the lights on, and we don't always keep the lights on. But as someone who struggled with this over the last four years, keeping the lights on at their home, this is something that we really, really need to pay attention to in this new world of extreme weather and distributed generation. So appreciate you giving me the giving me the microphone to share this with your audience and learn learn some from the conversation as well, too.

00:49:28.344 --> 00:50:07.269
And just to be clear, earnsay Protect was a framework created by intelligent generation, a software as a service. My listeners will have caught a recent interview with Jay marhoefer, the founder of intelligent generation out of Chicago. So I just have glommed on to that and really appreciate that framework. I'll have to listen to that one. Yeah, do check it out. It's not out yet, but I will be by the time this interview goes live. So all right, I want to thank Sean Kelly, CEO and founder of amperon, for. Coming on the show. Check out all of our content at cleanpowerhour.com Tell a friend about the show.

00:50:04.690 --> 00:50:17.094
That is probably the best thing you can do is just tell a friend. There are many, many people in the energy industry who do not know about the Clean Power Hour yet. So tell a friend and reach out to me on LinkedIn.

00:50:17.094 --> 00:50:23.755
I love hearing from my listeners. Check out all of the content at Clean Power Hour and Sean, how can our listeners find you?

00:50:24.500 --> 00:50:33.260
They can. I'm very active on LinkedIn, both under my personal and as well as amperon site. And we have a new, updated website out amperon.co

00:50:36.260 --> 00:50:38.599
thank you so much.

00:50:36.260 --> 00:50:39.500
I'm Tim Montague, let's grow solar and storage.

AI Energy Forecasting: How Machine Learning is Transforming the Grid

Listen On

Recent Episodes