AI in the Energy System: data, risk, and real-world deployment

Stephen (00:02)
energy sector hasn’t always been the quickest of taking a lot of technology. obviously there’s a there’s rightly so there’s a lot of risk aversion, you know, they’ve got to keep the lights on. You know, any disruptive technologies probably have to be safely checked. Having said that as well, I think maybe there’s too much

risk aversion to trying some of the things, which especially if you put sort of the human in the loop type things, I think you can make sure there’s safeguarding

Adam Sroka (00:45)
Hello and welcome back to the Hypercube podcast. I’m joined today by Dr. Stephen Harbour, Senior AI Consultant at Energy Systems Catapult. Stephen, great to you on board. Can you tell us a bit about who you are, what you do, what ESC does and your journey into AI and all that sort of jazz?

Stephen (00:57)
Hello.

Yeah, sure. So yeah, thank you for having me, first of all. Yeah, so I mean, I am a senior AI consultant at the Catapult. So if you’ve not heard of us, we’re a ⁓ essentially, know, part government funded not for profits, support for innovators in the energy sector. This, you know, we’re one of nine catapults out there trying to support, you know, various sectors. So they’re in different areas as is energy sector.

yeah, essentially we, you know, we’re, sort of a relatively small team. You met probably a couple of us already. Well, the other two, maybe, we, sort of try and help understand what the use cases are for AI out there. You know, what supports needed, what are the blockers for use utilizing AI in the right ways, across the sector, you know, whether that’s, Is it small innovators who maybe need support? You know, it might be just even looking at data, maybe looking for collaborators.

Or it’s maybe responding to consultations, understanding what regulatory needs there might be or what other requirements there are. So most generally, we’re there for trying to help wherever we can and also get kind an independent view of the sector and seeing where AI might be needed and where it’s not really needed as well, think, actually, which I think is equally important.

Adam Sroka (02:19)
Yeah, and I mean like some of the stuff you produce you’re seeing you yourself produce a really valuable will get into the weeds of but I’m interested actually just in the title because now I don’t even I don’t know the answer this you’ve probably not always been an AI consultant I’m interested to know what these systems catapult like when did they change it how our front of mind is like AI large language model stuff because a lot of the stuff you’ve done is like forecasting

Stephen (02:35)
yeah.

Hmm.

Adam Sroka (02:45)
It’s been around for like, we’ve been doing AI and machine learning for ages, we’re kind of reticent to call ourselves AI professionals a few years ago. So just talk me through that story a little.

Stephen (02:49)
Mmm.

Yeah.

Yeah, sure. mean, so my background is low voltage analytics really. So I was a postdoc for about seven or eight years, if I can remember, at Reading University and then Oxford University. So, you know, we were working primarily on one of the low carbon network fund projects, which were sort of precursors to this sort of NIC, NIA projects. And we’re working with DNOs to try and look at solutions, which could help them prepare for, you know, all the low carbon technologies.

So we were working with Scottish and Southern Electricity Networks. the intention was to sort of help them, you know, look at sort of other solutions on the networks, how to do some better modeling, how to better understand what they were doing. On my part, we were doing a lot of forecasting, which is quite an interesting area really, because at the time there wasn’t very much of the low voltage. But also it was a modeling the low voltage networks. We were working with battery operation as well. We were working with others.

more engineering focused individuals. I’m not so much a power systems. So we needed the inputs from the people who like Timo Unisov, who was working on sort of power electronics. Yeah. And it was very much key to sort of see, you know, what the capabilities were, and, understand some more nature about that system. I kind of got to the end of that. And I was interested in sort of doing something else a little bit more wide ranging, you know, and understanding where we could sort of look more.

More generally data. always, I always kind of do approach things and say it’s actually more about data driven things than necessarily saying, everything’s AI. but I think the name AI is just helpful in terms of understanding, you know, I think that’s where, where people are looking for in data intelligence. think that term has sort of entered the lexicon, but you know, we did start off as a data science team. And I think then originally the idea was we do a lot more sort of,

Adam Sroka (04:23)
Hmm.

Stephen (04:32)
data-driven modeling, bit more sort of look at sort of creating solutions for innovators in that respect. But I think we found that the needs of the sector perhaps were a little bit less, there’s less required for some of that. I think what we found a lot of was that was a lot of people getting used to some new data they’d never really seen before. So it didn’t really require anything too advanced.

on that side, or if it was more on the sort of landscape side and required a little bit more thinking about, know, this sort of, what was going on in AI, what needed to be done to prepare the system for AI, you know, what did people need to be thinking about? were the blockers? It kind of came down to that’s where we sort of split from the data science team emerged sort of two teams, really want to sort of focus on each of those, you know, so we felt like actually there’s a lot of people who’ve just got some new data. We need to start,

helping them and understand what to look at and just get them on the first rung of that ladder. So I think it made sense to split that focus in some ways for what the sector needed.

Adam Sroka (05:30)
That’s really cool. And it’s, an evolving place. And I think it’s quite forward thinking look actually, because when we go out into the market and we talk to people, everyone’s like on one side, like swamped in all this hype and news and the about AI and everything’s changing jobs and all that.

The reality on the other side is that the energy sector is pretty slow moving, pretty risk averse, like technology is pretty behind on the software side. Like hardware gets this enormous investment and it’s doing leaps and bounds and the tech’s kind of an afterthought. The software tech, So it’s quite a problem thinking approach, but I guess that’s the point of the catapult, right? It’s to be in front and sort of shape some of the paths for where the industry might go.

Stephen (05:54)
Yeah.

Yeah. And, know, I think a lot of the work has almost been on the sort of data and digital side, you know, our team here, Richard Dobson leads the digital team here. And, they’ve done a hell of a lot of work in trying to support data best practice, so you had the energy data task force, and then the, subsequent energy digitalization task force, which was really trying to set that foundation. And a lot of that work is still ongoing now, you know, things like the data sharing infrastructure and things like that, or,

on the go and AI is really dependent on those things, so almost some of the thinking is now but some of it’s sort of five, ten years down the line as well.

Adam Sroka (06:49)
Yeah, 100 % and like from what I see from doing some of the data conferences, that kind of circuit is the metadata cataloging, like data qualities, that’s all sexy again, because actually everyone’s realized that if we are going to really reap the benefits of all these tools, and it’s done properly.

Stephen (07:07)
Yep, I got absolutely right.

Adam Sroka (07:09)
you’ve you know shy to writing a piece of content yourself you’ve in fact co-authored a textbook the first textbook on dedicated load forecasting but distributionally right? How much fun was that?

Stephen (07:17)
Yeah, that’s right.

Yeah, it was really coming towards the end of my postdoc career really when this came to fruition. I maybe regretted it once I started it, but the idea was initially to just wrap up all the work that we’d been doing in this area with colleagues who’ve been also looking at this and write a book that we said we hoped we had really.

we found it was very difficult to find information. Everything was kind of piecemeal or it was focused in a different way. It was more like, oh, let’s apply the snazzy technique rather than think about the fundamentals as well. So yeah, we put this together. It was a lot of fun at times. And then other times it was like, there’s a lot more we could have in. And every time we looked around one corner, we’re like, actually, we probably need to build something else on this. yeah, I certainly enjoyed that. I think it’s…

it probably actually could do with an update. There’s so much changing in this area as well now. we’ll see if I have the gumption to do an update.

Adam Sroka (08:18)
Yeah.

Yeah, I don’t like a free time DNA. It’s a big undertaking. got a friend who’s writing a book and it’s swelled from 150 pages to 600 now and he’s like three calls a day through writing it and he’s thinking this might have to be three volumes and I’m on the journey and it’s taking longer than I’d hoped, even with all the AI-powered tools and stuff.

Stephen (08:26)
and energy.

Haha.

Yeah, yeah I haven’t, I didn’t use those tools, but now I’m thinking if we do update that might be really good place to start. I think writing with others helps a lot, getting that feedback, you know, as well, so.

Adam Sroka (08:55)
Like going to the gym, isn’t it? You don’t want to let them down, so you do the work.

Stephen (08:59)
That’s it, yeah. There was times they

were ready and then they were like, they’d just done a lot of work. You’re like, okay, I carry on with this, yeah.

Adam Sroka (09:05)
So right, okay then, so you’re probably world expert on how we forecast distributed systems and things like that or in the upper echelons of that group of people. What are EVs and batteries and heat pumps going to do? How much of a mess are going to make to the distribution system?

Stephen (09:18)
Yeah, sort of.

Yeah, you know, actually that’s an interesting question, I think. You know, when we wrote the book, actually, we probably could have written a bit more about sort of net load anyway, you know, things where PV systems are on there. We sort of focused on the demand really. But actually, heat pumps are changing the nature of, how these systems respond to weather. And, electric vehicles are quite complicated because, you know, they’re not always charging where they…

know where the house is, they’ll be charging somewhere else and we don’t know what that demand is. Everybody uses them in different ways. So I think it’s a very interesting area. how much we need the tools for what we’re trying to do there is still an open question. I think one thing that we found when we were doing this research in general is that there’s still a lot we don’t know actually. You know, when we, just as a kind of anecdote, we started trying to look at

You know, forecasting like 100 second substations, that was sort of a lot of the work we’re doing in Bracknell as part of the Thames Valley vision. And just trying to understand, how do we forecast this? What features do we need to include and those sorts of things? And one of the things we found was that, weather wasn’t actually necessarily a great driver of that demand. Probably because a lot of the urban demand used gas for heating, right? So.

Adam Sroka (10:31)
Yeah, okay.

Stephen (10:31)
We, you know, but

that’s not completely, you know, that’s, that’s not, ⁓ you know, that’s a very small case study, right? But we did find those other papers that started to say the same thing. you know, this was not, it was not affecting it. and we also found there was kind of a bi-modal sort of features happening when we found those feeders with large numbers of overnight storage feeders. So they, there’s the fact is here we have one system which probably is dependent quite a bit on temperature and then the other less so. And therefore,

We found there wasn’t a one size fits all solution when we were trying to forecast these things. So as sort of heat pumps pop into this system, we’re going to probably see these things change quite a bit. and AVs of course. but also if we’re starting to implement flexibility on them, the big kind of issue here might be that our nature of our time series changes, you know, how do we train on a system that has, these changes, right? How do we track them? You know, how do we know what the counterfactual was, and obviously these systems then.

have parts of them that are less relevant, right? So I think there’s gonna be a lot of issues. we are trying to think of ways to make this, because it’s very easy to sort of say, well, we’ll just do lots of local methods. We try and train on individuals, but actually this sort of transfer learning might actually be the way we deal with this. And we have to learn how to do that in a sort of more efficient way, I would say. So my rambling answer, I guess, to this is I think there’s…

Adam Sroka (11:42)
You’re

Stephen (11:43)
going to be quite a lot, I think, of changes. I don’t think it’s going to be easy to know exactly what because I think we still, even the systems we had here before they had all these low carbon technologies, had a lot of questions we didn’t understand fully on them because there’s just not been enough data out there. So I think probabilistic methods will be key to dealing with this. If we can sort of try and develop methods which are a bit more explainable.

that maybe understand sort of causal relationships in there could be really important. Especially if we’re trying to design solutions for consumers as well. Ones that suit them and which can fit within their lives and within the nature of their demands. So yeah, I think it’s gonna be quite complicated.

Adam Sroka (12:21)
bit that always itches at the back of my mind and like a lot of my forecasting background energy is always about price really you have to you get into demand and supply but really it’s price and options and stuff but the bit that always bothers me and it’s like the Ukrainian war problem it’s like everyone’s forecasting this is the price this is energy this is the way it is all gas supplies been banja axed and now the whole training set needs chucked out because

Stephen (12:32)
Mm.

Mm.

Mm.

Adam Sroka (12:49)
We’re in a new paradigm. Everything’s totally different. And for months, like you’re trying to make guesswork about how is this market going to settle? Like, is this back to normal? What we doing? And I feel like I have this like perpetual chronic concern, but that’s just the energy sector. there’s always some sort of event like that. And whether or not there’s fast moving acute events like that, or the electrification of vehicles, like massive, like system changing events or

like actually yeah the price of solar planets through the floor and radically changes what we’re going to look at and so I guess what I’m getting at is like

That causal bit that you mentioned is really important, I think. Can we start, instead of thinking about forecasting for forecasting sake, forecasting to uncover causal relationships isn’t something that I think people really deeply think about or I don’t think they go that far in their thinking because are we just always going to have two years of data that’s really valid and then everything before that where it was just so different, it’s not there. What do you think of my rambling?

Is that a question or not?

Stephen (13:56)
It’s a really interesting, no, I mean, you know, I didn’t mention it, but I think that’s a key feature of these networks. And I think it’s even worse at the distribution level. Right. So on top of that, you have the switching behavior, right? So when you decide that you need to reconnect one network to another network, because you’re doing some maintenance or whatever reason, you get this sort of, I guess you call it sort of, regime switching, right? Is sort of the technical term.

but you’d also get the concept drift as well, right? So when it’s technologies get installed, as we get more efficient bulbs or more efficient appliances, you know, the demand will change all the time. So how you deal with that? think one of them is sort of the adaptive forecasting approach. You know, we make sure that we have models which are always discounting older information. You don’t maybe want to eliminate it entirely because there’s probably some features in there you want to retain. The other side of that, might, yeah, I think causal

It’s an interesting area. I don’t know if you’ve ever spoken to Pierre Pinson over in Imperial, but he’s recently written a paper on looking at causal inference within price forecasting and the impacts of wind and solar. So he’s definitely looking, you know, I’m seeing more and more people looking at this to try and understand things. And I think like the counterfactual is a really important area for a lot of energy research, because it’s like,

Adam Sroka (15:01)
out really.

Stephen (15:13)
especially, networks that want to know why something went wrong, and also what if analysis, right? They want to say, well, if I changed this, what this happens. So it could be a really interesting area. And I think it’s an area I’d like to know more about. whenever I can, I stick my nose in the door of other people working in that area and see what they’re doing. It’s, requires maybe some more, more sophisticated thinking knowledge. I think these, these things have been used.

Adam Sroka (15:29)
Yeah.

Stephen (15:38)
You know, as far as I’m aware, I’m not expert, but these things have been used quite a bit more in sort of medical research and places where you can do a bit more sort of trials that you can control. You know, we can do these obviously with things like living labs, but, I think that’s an interesting area. Yeah. think the other area I guess is foundational models. People are looking now at these foundational time series models, which can do these zero shot forecasts. Right. So, you know, with very little information.

I haven’t seen them how they deal with this sort of regime switch, but I think they’re a really exciting area that could be, linked into sort of dealing with issues like this. I certainly know, I think that distribution network operators find this switching behavior is a tricky problem.

Adam Sroka (16:18)
And it’s funny to talk

about foundational models, because it pings the large language model kind of alarm bells in my mind. We’ve got a demo of… I it’s called… Have you seen that?

Stephen (16:24)
Mm-hmm.

I haven’t seen the demo yet. I’ve been seeing a few posts about it now.

Adam Sroka (16:31)
And you just think like, so, and it’s like incredible. So for those that are unaware in the audience, AWS have effectively gone through the cost of training a large language model, but they only fed it, so language is a kind of this number there. They only fed it time series data, bucket loads, like more than you can imagine. the kind of data only Amazon can get hold of. And then, so you feed it in a time series and it spits out, it does what a language model does, but for data series in a time series.

It’s like really impressive. The demo is demo as well. But the reason I bring that up is that, yeah, you’ve got what’s right for the industry and academia and what we think, right, we should be doing this staff. But the nature of the world we’re in is all the biggest companies in the world are going to go after whatever they think the best tech is for general purpose. And you’ll get leaps and bounds of performance improvement from some of those technologies.

And the reason I think it’s relevant to this conversation is that that is wholly unexplainable. Like, that is as far away from explainable as you could possibly get. this like 150 billion parameter supermodel, just that’s the number it’s about. And that’s that. And you could do some grumpy stuff to maybe get it to tell you why. But that’s also stochastic, so good luck.

Stephen (17:40)
Yeah.

Yeah, and that area is really interesting. There’s a of black box models out there. And the explainable methods around them aren’t quite what maybe would satisfy a network operator to say, this is explainable. It just sort of highlights what weights are in the function maybe. Image explainability tells you that the image is looking at the eyes, right, and the nose and the mouth. And it’s like, well, OK, that doesn’t tell me when it goes wrong.

Adam Sroka (17:57)
Yeah.

Thank you.

Stephen (18:11)
There is, I think there’s a consideration to sort of understand the stability of those models and that helps a little bit. I think it always has to be tied to, I think benchmarks are really key here. You know, if you’ve got to create something which is quite explainable as well at the same time, I don’t think you have one model. It shouldn’t be like this is the model. But if you can, I think getting models which are interpretable is really powerful.

And I also think that’s probably where a lot of value is. know, I think sometimes we were looking for really, powerful forecasts and actually, I mean, when we were doing forecasts, we’d have, you know, in the postdoc team, we’d all be trying to outkeep, compete each other. Right. And often we’d find that like the average of the last four weeks was the best forecast, so you’d be a bit frustrated actually, I wish I could use this really sophisticated, complex thing. I don’t think, I think getting the basics right.

Adam Sroka (18:49)
Thank you.

Stephen (19:00)
You know, it’s key first,

Adam Sroka (19:01)
And it’s totally, we’ve run a couple of hackathons and I think both times someone using logistic regression and just really good features as well. And you’re just like, oh, and it’s like, I to build a really complicated crystal ball out of like millions of layers, just being good old fashioned stats.

Stephen (19:11)
Yeah.

love the stats methods, that’s my background really, know, originally. So it’s very much my wheelhouse. You know, the book we had was Marcus Foss’s was from T.E. Berlin who wrote in the book as well and he was much more machine learning. So we had the nice mix of the two specialties really.

Adam Sroka (19:33)
Yeah.

The first person I ever hired was a stats postdoc, because I was just like, I need that rounded off. That’s like, yeah, it to be weird. ⁓

Stephen (19:38)
No.

Yeah, it’s an interesting

area as who makes really good data scientists as well. Actually, physicists I found have been brilliant, actually. Yeah, I mean, because I think it’s like that nature of like understanding science and hypothesis and, also like, understanding how to look for errors. And it’s like about, that natural way of thinking, I think is really useful.

Adam Sroka (19:48)
Thanks very much. ⁓

Yeah, and do know what, I agree. mean, you get great data scientists from all flavours of other areas and they’ve all got different strengths and mixing them up is what I think is best way, I hope, give us the way it is, I was a physicist and I think, I agree, the one thing I believe physics teaches you, a formal education in physics teaches you, is whenever you’re given any situation, any model or whatever, to immediately go, right, how can I break it?

Stephen (20:11)
Hmm.

Mm.

Adam Sroka (20:25)
What if I stick an infinite load in? if I stick zero? if I stick niggas in? And really push the very, very edges of it. And it’s that acceptance that all models will break. And it’s my favorite quote, all models are longsumming useful. I love the edge of the box. I love that quote. And I think physics, like every, and I do it in like, I do a lot less sort of technical stuff now, more like strategy stuff and business design and things.

Stephen (20:28)
Mm.

Mm. Yeah.

Adam Sroka (20:50)
But even that, I’m like, when does this make? When does, like, what if I’ve got a thousand meetings a week? What if I’ve got no meetings? And it’s like that way of thinking just comes naturally to physicists, whereas others, like, it’s totally alien and they’re trying to concentrate on the refinement of the bit in the middle. But it’s, yeah, it’s a nice place. And I think as well, can do like chains of events, like how to understand all this affects that, affects that, and there’s a feedback loop. Similar to electrical engineers.

Stephen (21:06)
Mm. Whole systems approach.

Yeah.

I mean, I remember talking to just to finish up, but also, you know, you can get, as he say, all flavors. And I think I remember hearing from a sort of lead data scientist in the big ARM. And they were talking about like a journalist, ex-journalist being one of the best data scientists had because of the way they, could tell a story with the data, you know, that they thought about. So there’s different strengths, isn’t there? Really? Yeah.

Adam Sroka (21:28)
Yeah, yeah, yeah, yeah.

Yeah, definitely. And being able to ask the right questions and understand why we’re doing this, what’s the point, what’s under the hood. Yeah, I don’t think it’s the kind of domain of any one class of education or career. I’d love to move on a little bit to, yeah, smart energy data service and some of the work you can do there. What is the project, what’s it trying to achieve?

Stephen (21:35)
Mm.

Yes, sir.

⁓

Yeah, okay. So this is kind of a relatively recent project really. this is the Smart Energy Data Services. said it’s one of the smart data research centers, which are ESRC funded data platforms, essentially. So these have been developed to sort of try and help share data in really important areas, really. I did mention before the off-jam.

energy data press practice principles. And that was very good for certainly the presumed open principle, which was, let’s presume data is open and if you can’t open it and there’s valid reasons for that, then, you know, you should, you should look, you know, triage it in the right way so that you release it in different ways, but perhaps you have to have closed in some consensus. so that was very effective.

I think we’ve still a place where there’s key applications in certain areas where we need specific powerful data sets, smart data sets. It’s in this case that talking about, which can open up innovation, which we can’t really do because that data doesn’t really exist. so these sort of SDR platforms are trying to sort of support that by, concentrated specific data sets around these areas. So this project, it’s like.

by Richard Snape, who’s in the energy systems catable. I’ve got one job, it’s to sort of support the researchers, the work stream in that, which is probably the fun bit, to try and get different academics to use data to try and focus on different innovation. And in this case, we’re sort of focusing on two major areas, which are equitable EV charging infrastructure. So that means how do we make sure that we build,

EV infrastructure that can support wider communities and groups and make sure that also we’re building them in the right place, how are they going to be used? How does the community use them? How does people react? And the second one is essentially a non-domestic energy use behavior as well. So a lot of the research, as you probably know, there’s a lot of research at sort of domestic level. Some smart meter data is out there. Obviously, DNLs are releasing some aggregate smart meter data in some forms.

but you know, non-domestic, it’s a bit harder to get hold of some of that data and sort of generate, insights. yeah, so this, in short, this platform is, is a way to try and help support over the next couple of years, but hopefully, you know, who knows, we could go for that after this, in these specific areas. So it’s like targeted data sets rather than sort of let’s generally try and release data sets, you know.

Adam Sroka (24:19)
And the use case then primarily is for researchers to pick it up, do interesting work on real stuff.

Stephen (24:26)
Yeah, so when we release the platform, probably in the next few months, I think, we’ll try and encourage to link up, I mean, we’ll say academics primarily, but I think we do want to try and encourage innovation more widely as well. Obviously, the conditions around the data will be dependent on who’s sharing data with us. so it won’t be for commercial.

products, but I think it still could drive innovation, which could eventually lead to really good innovation and products out there and services.

Adam Sroka (24:52)
So okay let’s look at the two the problem sets that were mentioned. Equitable EV charging is really interesting. had Zapmaps on, I’ve spoken to Connected Curve, I’ve spoken to MyEnergy, Omi, like all these people that have massive data sets of like people plugging vehicles in and out. do you do with them? What’s your take?

Stephen (24:59)
Mm-hmm.

Mm.

of all the different datasets that are out there.

Adam Sroka (25:11)
Or

just like what’s it going to take to do it? What’s the most promising sort of route to unlocking equitable EV charging infrastructure?

Stephen (25:14)
Yep.

Yeah, I mean, there is research out there in terms of EV charge point locations optimisation. There’s a lot also about smart charging now. I think Centre for Net Zero released some research recently showing that you could sort of help flex things through different tariffs. So I think we have all the building blocks for it.

But I think we haven’t quite looked at what’s the impact for communities. How do we make At the moment, there’s kind of almost a split of who’s using EV. You’ve got sort of wealthier individuals who can afford the infrastructure and the charging. And then you’ve got maybe others like actually taxi fleets, which would require this infrastructure. And it has to be in the right places for them to use. But also,

it could have a wider impact on others. So I think we definitely need to understand a bit more about how those charges are used and who’s using them. That’s not easy to do because, that maybe requires also looking about journeys, looking at obviously that transaction data, looking at how they’re charged and all that information together, which not many people have access to, right.

Adam Sroka (26:26)
Yeah, PII GDPR all the lovely stuff like.

Stephen (26:28)
Yeah,

all the favourites. So, you know, I think, that’s where we’re thinking, there needs, we’ve got sort of like this, it’s funded by the ESRC, of course, the social science is a big part of this. So I would say it’s not all necessarily AI, obviously. So the sense platform is actually we’re looking at, we’ve got people who are social scientists, we’ve got people who involved in policy. So we need those people to be communicating to use the sort of maybe cliche, but it’s interdisciplinary is really important here, I think.

you know, how to make that data accessible in a nice way. So we are looking at putting sort of, lots of interesting data sets together. I can’t say too much necessarily because we’re not released yet and what we can, what we’ve, there’s some that’s just still speculative that we’re still negotiating, but we’re hoping that we can have, some really cool data sets that aren’t typically out there for research.

Adam Sroka (27:03)
Yeah.

And how do you see that interplay with things like, is it Faraday from Center for Net Zero, and other synthetic data sets? That’s like, it’s a topic I love, synthetic data sets, and you see arguments on both sides. How do you see that playing and mixing in with what you’re actually doing?

Stephen (27:19)
Mmm.

Yeah.

Yeah, it’s a really good point.

it’s great that sort of, Center for Net Zero have obviously access to all the Octopus data, which is really advantageous. Synthetic data is getting a lot of traction now in terms of people talking about it as a way to release data that doesn’t exist or maybe it’s not available. So

a big part of this will be looking at synthetic data actually. So there’s certainly some data we will not be able to release in anger. So we are trying to look at things like public charging profiles as part of this. And I think to do that, we’re going to have to try and do some synthetic. Now there’s lots of ways to do it. We mentioned foundational models before is one thing with like zero shot, but, you know, that’s difficult

without some data out there, it’s difficult to actually develop those in the first place. So there’s also ways you can look at it from more of a, principles based approach where you bottom up and you can build things based on, you know, understanding things like some journey data, which does exist out there. Now there is some charging profiles from other countries as well, which you can line things up with.

But it’s a complex question because there’s the sort of this mix between making sure there’s utility in the data, but also, the more realistic that data looks, the more there’s also a question about making sure the sensitive data doesn’t get released. it has different uses for different people. And some people still just want the real data, even if you create a very nice sort of,

synthetic data set, but it could still be used for other ways, other things to just make sure that, you know, maybe it’s about proof of concept for a product or maybe it’s about, developing something which, you’ve at least showed us some robustness in it as well. So.

Adam Sroka (29:02)
Yeah, but I just think that I could see it’s such a painful topic because you can very easily run afoul of the GDPR rules around using data sets for things they weren’t collected for by just joining them up like sort of innocently to try and get better insights. And all of a sudden you run afoul of some pretty hard to interpret rules. And that all just means lawyers know.

Stephen (29:14)
Mm.

Adam Sroka (29:32)
or Procurements as Snow or compliance as Snow and you get stuck. I mean, just as we expand and yeah, see more adoption of the synthetic approach because it’s great.

Stephen (29:37)
Yep.

We do have a work stream. So it’s not my work stream, but we have a work stream looking at the data ethics and responsible data use as well. obviously everyone we sort of taking, know, using their data, we’re making sure it’s all got the sort of right license conditions. It’s, got the right checks. and we can, with some of the data sets that will be behind sort of more privacy barriers rather than sort of just having them open on the platform as well. So.

they are definitely trying to make sure we try and comply to, to everything that needs to be complied to. But yeah, you’re right in terms of, yeah, these things are very complex. I think synthetic data is still, there’s a lot of questions. mean, there’s some interesting research I saw the other day where we are releasing sometimes real data, which actually does have privacy issues. You know, and I remember hearing there was a paper or something that someone had done in the medical field where they think they’d taken MRI scans.

don’t quote me on this, but they take an MRI scans and then you could link the MRI scan data to camera scans, right? Of faces. And then you could find out whose MRI scan was connected to that. I’d have to dig out the paper, but that’s an interesting problem. Nobody thought that was real data, you know, was being released. So in some ways there’s gotta be a question here about

Adam Sroka (30:48)
Yeah.

Stephen (30:51)
How do you deal with, there’s a lot might be unintended consequences. So with certain things, and I think people need to, first of all, try and do the due diligence as much as possible with synthetic data. but on the other side, be ready to sort of make sure that, any issues that are found or dealt with quickly, I don’t think, with the synthetic data, the main issues will probably come around when there’s very unusual edge cases, probably, you know,

Adam Sroka (31:13)
you

Stephen (31:13)
The general behavior will be, cause there are lots of it, you’ll probably, it’ll all be sort of general. But if you sort of, imagine you’re trying to model them, EVs or something, one person’s got an EV on that street, you see that part spike profile, you’d be like, ah, that’s their profile. Right. So, you know, that’s certainly one area you’ve got to make sure you do your due diligence on this area.

Adam Sroka (31:26)
Yeah.

minefield and that’s why I it. Yeah, I find it fascinating but scary at the same time.

Stephen (31:32)
Mmm.

we’ve also set up a synthetic data group to sort of make sure we’re doing it right. So we’ve got the teams working on that, but we’re working with a few people who researching this precise area of privacy and utility to make sure we’re doing the best practice really.

Adam Sroka (31:51)
So to change gears a little bit, like, let’s get rid of all the ethics, who cares? What about just pure AI? What’s the kind of fun? Like, what are we going to do? Is AI going to take over the energy sector? Where’s the most ripe for innovation and disruption with all the new tech?

Stephen (31:53)
Yep,

You

Yeah, as you said, energy sector hasn’t always been the quickest of taking a lot of technology. obviously there’s a there’s rightly so there’s a lot of risk aversion, you know, they’ve got to keep the lights on. You know, any disruptive technologies probably have to be safely checked. Having said that as well, I think maybe there’s too much

risk aversion to trying some of the things, which especially if you put sort of the human in the loop type things, I think you can make sure there’s safeguarding whether you you wouldn’t expect, but to be kind of more, let’s be more like open about, think, you know, what we, what could be applied out there. I mean, I like forecasting, so it’s obviously a bias here, but we’re talking to a lot of DNOs and they are really, a lot of them are trying to sort of develop their capabilities in this area. Some are very.

already quite far along. others are, are actively looking. I think this could be really keen, obviously understanding, anticipating those things on the networks. And then it opens up a whole different, broad, different types of use cases, right. And services you could offer, offer, right. You some of it’s about, delaying traditional reinforcement, but other cases it means you can, utilize that network, to its full. I think forecasting.

the distribution level is at its infancy, but I think it could really make some strides quite quickly. We’ve seen work by, I guess, Open Climate Fix, but they’re doing some brilliant work with solar forecasts and, I know they’ve been working with Nezone developing some really cool methods there, very short term sort of stuff. That’s a really cool area because it’s sort of incorporating time series, which was sort of my tradition, but sort of imaging technology and satellites. And I think that that’s really exciting area.

and extending that, guess is, I’m seeing a lot of advancement in AI for weather, which I think is very cool. I think it’s, it’s got the potential for, really improving a lot of energy applications. you know, It also means you can, you can generate, obviously you, generate probabilistic forecasts in, in weather by, often you have to run the entire system, right? It’s a, it’s a big, it’s Navier-Stokes fluid flow equation.

Adam Sroka (34:00)
Yeah.

Stephen (34:00)
You know,

you have to, to, You only get like 50 ensembles from like, ECMWF is one of the best out there. They generate, you know, only 50 I say, but you know, obviously that takes a lot of compute power and they’re one of the best out there. With AI being utilised in there, I feel like there’s an opportunity here to really generate more and more of these ensembles, you know, create a really big, good picture of like the sort of spread of the potential weather scenarios. And I wonder how much that could help certainly with.

with forecasting applications and operations which are a bit further along.

Adam Sroka (34:29)
So a big spike and buzz in the Google DeepMind weather thing, but I’ve not seen or heard of anyone using it or seen much since it was launched. There was a press release about six months ago.

Stephen (34:32)
Yeah.

I’m,

I think, yeah, I mean, I’m hearing more people trying to understand how weather can be applied in these systems. First of all, I think weather is a very complex error. That’s, know, I, my PhD was in, numerical weather prediction. Actually it was more on the optimisation side. And so we were there. It was the, it was the fluid flow equations for modeling it all. And I think for a long time, there wasn’t a consideration that you could actually, AI maybe wouldn’t do much in this area.

that it was very much, it’s physics, know, it’s, is how fluid flows. don’t think you can do better, you know, and I didn’t see much movement. I left in 2011 in that area, but I hadn’t seen much change until the last few years where I’m seeing more and more sort of, use of, saying, we’re generating these ensembles. So I know ECMWF have now got their AI system, as well, and that’s working better in some areas.

There’s probably not enough incorporation of weather anyway at the moment, or at least not the advanced weather products. So I think, we’ll see that thing might be further down the line. I think before, as you say, I don’t think anyone’s sort of, we’re already instantly using, AI ensembles.

Adam Sroka (35:38)
Do I find though and it’s like this is a generalization I guess so I’m all these people unfairly but I am When you in the film I when we’re talking to customers in the field that want to incorporate really detailed with a full car since their systems and stuff You find that they hear their visions or they have they’re aware of some of our advanced stuff Actually, they just get blocked by integration and the API is and like actually like okay. Yeah, this is the best one

Stephen (36:00)
Mm, right.

Adam Sroka (36:03)
team to set up a tool to use it. So I’m going to go back to a MeteorMatics or a Windy.com because it’s easy peasy and it just works and gets me like 95 % of the result I want. And I just wonder, like, it’s a shame because, and obviously I’ve not gone on through these integrations myself, but we’ve been through a lot of them. And some of them make it quite difficult, like, because they’re coming out of organizations that aren’t engineer focused or commercially driven. They’re coming out of places that are not our priorities. So they’re not thinking, right, how do I commercialize or make it?

Stephen (36:11)
Yep.

Adam Sroka (36:31)
make my service easy to digest for people that are maybe less technical which have bought me out.

Stephen (36:37)
Yeah, I would agree. I’d agree. mean, this is where I say, you know, think first, yes, simple methods can make get you a lot of the way anyway. You know, it’s only if you’re sort of you have a particular advanced application that requires that number of ensembles, for instance. think there’s a there’s been a general issue we’ve come across a few times with whether products are difficult sometimes to access, understand pre-process. I think, you know, if you talk to someone like Jeffer Brown at Glasgow Uni.

he knows all these issues and how to deal with them. You know, it’s, it’s quite complex data. I mean, if you’re not a weather expert, there’s, get a product like, oh, temperature two meters or, then one at 10 meters. which is going to be valuable for what I’m trying to do. There’s numerous radiation variables out there as well. and then, if you get your probabilistic forecasts, you need to recalibrate them, right. And to use them, probably process them.

And that’s not what there’s not much awareness in that area. there’s not necessarily easy knowledge about how do you do that? So yeah, I think there’s a translation. Opportunity here actually, you know, and I’m sure they do, obviously these groups like Met Office, they do have consultancies in the work in these things, but maybe it’d be useful to have some sort of tools out there to sort of say, well, automatically do it, especially for innovators.

Adam Sroka (37:49)
If anyone wants to fund it, I’ll do it. That’d be great. I’ve got a weather translation services. I’ve got quite a lot of sound about it. We ran a hackathon with Petro and it was a 24-hour thing. I think the average, like it was ages ago, can’t remember exactly, but the average time to just ingest and process the weather data file was like 18 of the 24 hours.

Stephen (37:51)
Yeah, yeah, that’s it.

Well, that’s it. Do I get a commission?

Hmm.

Adam Sroka (38:16)
because so many teams had come from other sort of tangential parts of energy and never worked with these 4D datasets before, didn’t have a clue how to turn them into something that their time series models could eat. And it was really quite interesting to see people really struggle with just that foundational step.

Stephen (38:22)
Right, yeah.

And they all come in different formats as well, right?

Adam Sroka (38:35)
what we like in energy,

high standards.

Alright, what kind of horizon scanning stuff are you looking at at Energy Systems Catapult and like, yeah, what do you think is going to keep you busy in the next couple of years?

Stephen (38:52)
Yeah, I guess it’s different horizons. To some extent, A lot of the work that we do is, where is the challenge with actually implementing some of the basics in some cases, what just needs to be set up just to make sure that we can facilitate that. I think there’s still a lot of challenges out there and you probably heard them all in some place before, but

skills is a big one. You know, there’s still the risk aversion to some extent. There’s issues with data, data cleanliness issues And, and also, you know, some of the products require, quite advanced digital infrastructure to be implemented. And those are sort of like all the foundations. So, in some ways, the issue isn’t the techniques and the models and those are, you know, we’ve got some great, great models out there. They’re ready to be implemented.

But I think maybe there’s also a bit of a requirement of sometimes maybe not the, there isn’t the, sound of sort of environment to be able to implement some stuff, we need, if we’re moving towards some more household flexibility, there’s, needs to be sort of easier ways to access that as well. You know, how do we make sure that we can support people who are early movers as well in different areas.

But I think to be more kind of sort of in the more exciting kind of horizon scanning area, I’m quite interested in these sort of foundational models in how they can help a lot of people get very quickly up the ladder. How do you move quite quickly to implementing something? think sometimes starting from scratch can be a bit daunting.

but if there’s models out there, I’d be very interested in seeing how, you know, I’ve seen some stuff around foundational grids modeling as well now. it’s not very easy to create, network products. If you haven’t got lots of networks to try and model things. So I was recently reading a paper on sort of, creating sort of superficial networks, by creating sort of these foundation. think those could be very interesting, very exciting, sort of maybe rate very much in the research space, perhaps, but.

Adam Sroka (40:47)
Yeah, okay.

Stephen (40:53)
I think one thing that was off-gem recently released the idea about a technical sandbox, you know, and that idea is quite interesting in terms of how do we really test things. And this could go some way towards helping with risk aversion as well. and that maybe could be coupled. One thought I had was, well, could we do something which has this sort of synthetic networks? We could generate lots of different networks and.

you know, test lots of different products on it.

Adam Sroka (41:17)
It makes a lot of sense, right?

Because it’s not like, it’s feasible. I don’t even think it’d be that hard. Because the network data is public, right? So we own it. yeah, yeah. And you’ve got loads of it. There’s loads of data. You just have to polish it and see if, you could totally see how European Commission, you consolidate the World Energy Council or something, consolidate network data from a few places. And you’re away to be able to come up with light synthetic.

but systems that behave like a sort of simulated network or whatever and if that was publicly stamped by someone like off-jam then yeah if I were going to take risks in the system I know well where’s the regulator going?

Stephen (41:57)
Mm.

Or even like, you know, someone like Niso or, know, so that could, mean, they are running some sort of, test beds. I think there’s different projects they have, which look at testing things that go onto the balancing market, for instance. yeah, I think. in terms of the risk thing, the only thing I, you know, there’s often, there’s a worry about all these thousands of IOT devices, that being operated on, that could affect the network. The one thing we sort of thought about when we were sort of responding to the,

off-champ consultation in this area as well. Actually, a lot of these devices are driven by the same things, right? In some cases, it’s weather. In some cases, it’s the markets. if they’re all correlating with each other and they’ve got the same confounders, is there a potential danger here of, cascading errors? But, I think, what contingency would need to go in place? So actually, I think testing these things is the way to go about some of this to just really understand a little bit more,

Adam Sroka (42:43)
Yeah.

Stephen (42:43)
If you’ve got all these EVs, which are responding to the same market signals, you know, we’re not there yet, but is there a smart ways to sort of understand what impact that might.

Adam Sroka (42:50)
I genuinely, my gut feel is that, see the EV thing, when the big kind of roaming virtual power plants will come online and start with Krakens and then a couple others will pop up, there’ll be some sort of big fight and a catastrophe that will fundamentally change the way we do everything and it will cause, like I think it’s just almost unknowable as to what they will need to make to make that okay. And it’s just coming down the pipe. And so we just need to be ready to react to it.

Stephen (42:59)
Mm.

Yeah. And I wonder whether we need things like this sort of, you know, these tests, know, lot of testing to sort of at least, at least iron out what, know, many things we might see.

Adam Sroka (43:20)
Yeah.

Cool. Last question, guess, someone was coming to you today asking what’s the best way to get more value out of our data or AI or technology, what would be your advice to it?

Stephen (43:34)
⁓ crikey. There’s quite a few things, some of them quite obvious, I think. We obviously talk to lots of different organizations and they’re all at different levels. are way up there, they’ve got their data in nice formats, they’re using it in the right ways. Sometimes it’s just the fundamentals that people need to get right. I think there’s a lot of data scientists we talk to get a little frustrated that…

they’re actually working on formatting the data and doing the data engineering tasks, you know, ⁓ which is, know, obviously a key part of it, but that becomes, that was the, you know, that wasn’t necessarily the role they went for. So the planning that team and getting that right, you know, I think we find other times there’s sort of an innovator there. ⁓ you know, just one, they have some, maybe one person in their data science team and, and, know, they sort of were brought on from one project and then they weren’t quite clear what to do with them. So I think, you know,

It’s sometimes it’s like, we wanted to try something And also, that person works in isolation. So it’s not always clear. often called magicians, you know, and you’re like, well, they’re doing science, right? so there’s sort of these structural things. I think this sort of T, you know, building teams in the right way. Obviously we’re a bit outside, we’re, not sort of building our own teams in that way. But I think just from, I’ll just talk from maybe my experience with this little low voltage area, which might be interesting.

I feel like there’s a lot of lessons you get by, say marinating in the data, right? You know, you need to be embedded if you’re going to really understand how to respond to things. You know, there’s a bit of domain knowledge that’s useful. it’s a bit of like that technical skill. Other coding skills, of course, as well. but there’s this understanding of the deeper knowledge of some of the data. And

this is kind of the point I always use in terms of trying to understand data. When When we were first looking at LV networks, we tried to do household level forecasts. And we did a whole range of them and then we did some rubbish benchmarks deliberately to sort of see how they did. And despite a lot of sophisticated methods we used, the best ones, often the best forecast was this flat forecast we did. That just was a flat line across the day,

which didn’t have any of the features of the real system. It didn’t have, the spikes in the right place. It would do very well, right? And we saw, well, we figured out quite quickly that this was caused by what we know as like the double penalty effect, right? So we the first to sort of see that this effect was very unusual. And it’s basically because, you know, we use point wise forecast measures, right? RMSE, make, things like that, where we sort of say, right, there’s a…

The difference between each time step, know, 3pm versus your 3pm forecast, right? What’s the difference? And obviously if you’ve got household level data, it’s very spiky. Yeah. It’s got a peak at a certain time. It might move though, right? If you get home from work late, you’re probably going to do all your activities. You’re going to cook your meal. You’re going to put the kettle on TV at different, just a slightly shifted time. So that peak moves, right? So if you’re then trying to say, well, what’s a good forecast for that? You might have one that has this

spike at 5 PM, but if you’ve moved to 6 PM, you’ve got now two big errors in your forecast, right? You got your first 5 PM error, and then you’ve got a peak because you, didn’t forecast right at that time when there was actually a peak. So, our traditional error measures were no good for that, but it took us a lot of kind of experimenting and failing and like looking at the data. We did a big sort of look at a big study and found like basically 50 % of these forecasts could have this double penalty effect in them.

so my long-winded point basically is that, we wouldn’t have known that if we hadn’t had like a lot of time with the data, you know, I think often there’s, and I understand there’s a pressure to sort of operationalize things to get a product out fast. but in some cases, that can, that can falter, right? We could easily release. I mean, if we released the flat forecast and said, that’s a great forecast, it would be no good for charging and discharging a battery, right? it would, it would give you no results because it has no

It has no utility to it at all. But objectively, according to the measure, it’s very good.

Adam Sroka (47:09)
Yeah.

Strange world eh? I do love it though. It’s been an absolute pleasure to have you on today. Thank you so much. been a great conversation. Before I let anyone leave, I always like to ask if people want to find out more about you, about Energy Systems Catapult, or if there’s anything you want to promote or plug, where should I send them?

Stephen (47:13)
Yeah, it was very interesting.

Yeah, you too.

Mm-hmm.

Yeah, so we obviously have it, you know, more generally the energy systems catapult, have our website, so you can find us just by Googling us, I guess. But also we do, I run a webinar series where we bring in, you know, brilliant SMEs and businesses and academics who can share like the latest learnings and what we think is interesting out there. Yeah.

And so yeah, you’ll find mainly we advertise that on LinkedIn. So find me on LinkedIn is probably the best place. I’m quite relatively active on there, often, sharing some of the stuff we found and, you know, and especially the webinars.

Adam Sroka (48:05)
Well look, thanks once again for coming on and yeah I’ll see you on Highland Art very soon.

Stephen (48:10)
Yeah.

Yes. Thank you very much.

Adam Sroka (48:14)
Thank you once again for listening to the Hypercube Energy podcast. As always, if you have any feedback or comments or you want to get involved, please get in touch. And if you want to find out more about what’s happening in data and AI across the energy sector, please join the Beyond Energy community and I look forward to speaking to you soon.