Season 1 Episode 12: Sallyann Blackett, E.ON, on Using Data and Instinct for Energy Forecasting

This week, Sallyann Blackett, Head of Volume Forecasting at E.ON, joins us on the Hypercube podcast for a deep dive into the challenges and developments in energy forecasting.

Sallyann lifts the lid on how the forecasting teams operate within one of the energy world’s biggest players. She discusses her team’s work predicting gas and electricity demands both long- and short-term. She talks through the complex modelling and inputs used, as well as challenges around data analysis and accounting for external factors from the weather to the cost of living. She explains why forecasting is about more than just data, the value of a good team, and where AI fits into the energy forecasting operations of the future.

LISTEN ON SPOTIFYLISTEN ON APPLE
S1E12 Sallyann Blackett

In this episode, we covered:

  • Managing long- and short-term forecasting challenges
  • Using gut instinct alongside data models
  • The value of a strong in-house team to avoid blocks and bottlenecks
  • Playing catch up with solar energy modelling
  • A cautious approach to AI

The weekly Hypercube podcast sits down with leaders in the energy and utilities sectors to explore how data analytics can help businesses make smarter decisions and accelerate business growth.

CHAPTER MARKERS

[0:49] Sallyann introduces herself and gives a quick rundown of her role as Head of Volume Forecasting at E.ON.

[2:08] Sallyann explains some of the complexities of forecasting across various energy sources and business units in both the short- and long-term.

[3:31] Sallyann shares more on how her team uses different processes and models for accurate long and short-term predictions.

[6:53] Sallyann discusses the need to account for factors like weather and customer behaviour, and the need to employ instincts alongside data to get the best results.

[9:53] Sallyann talks about how and why E.ON develops its forecasting models in-house.

[13:39] Adam asks Sallyann about the challenges of analysing large datasets and prioritising areas of focus.

[16:25] Sallyann discusses the current tools her teams are using for data analysis.

[18:41] Sallyann explores the issues that arise from being too dependent on external teams, individuals with particular skills, or specific vendors.

[23:19] Sallyann discusses how AI and ML are enhancing forecasting capabilities but also stresses the need for caution.

[25:26] Sallyann shares where listeners can find out more about working with E.ON

TRANSCRIPT

Hypercube Podcast Transcript

Title: Sallyann Blackett, E.ON, on Using Data and Instinct for Energy Forecasting

Host: Adam Sroka
Guest: Sallyann Blackett

Intro: Welcome to the Hypercube podcast, where we explore how companies in the energy and utility sector leverage data analytics to make smarter decisions and accelerate business growth. I’m Adam Sroka, founder of Hypercube, a strategic consultancy that supports asset owner-operators, traders, route-to-market providers, and energy services companies to unlock the power of data.

If you’re interested in hearing real-world examples of how data and AI are advancing the energy sector, this is the show for you. 

Adam Sroka: Hello and welcome back to another episode of the Hypercube podcast. Today I’m delighted to be joined by Sally Anne Blackett, Head of Volume Forecasting at a little company called E.ON. Sally Anne, if you’d be so kind as to introduce yourself and let’s love a little bit about like, what does that mean? Head of Volume Forecasting, sounds interesting. 

Sallyann Blackett: Yeah, well, Sally Anne Blackett, I have been at E. ON for about 17 years now. My team is responsible for long and short-term forecasts for every business unit that E.ON has in the UK. But we are a company that supplies a fairly wide range of customers. So we have a lot of residential customers. We have small businesses. We also have a corporate.  If any big corporate business, and we have, I’m going to say small-scale generation and upset are all like generation suppliers.

So we have wind, biomass, and solar, either done through power purchase agreements or we. Self-supply and we also have quite a reasonably sized heat network, where we supply heat as a commodity, generally from small-scale CHP sites, and we forecast the CHP requirements for those heat networks. And a range of long and short-term forecasts. 

So when I say long, I’m talking four to five years out and supporting both the energy purchase and the financial part of the business. And then short-term forecasts are 14 days out when we’ve got a better idea of the weather. 

Adam Sroka : Yeah. Okay. So how are you going about doing that? For those that are unaware, it’s not straightforward, forecasting gas and electricity?


Sallyann Blackett : We do gas purchase requirements for the company. So we’re working out what our customers are using with power. We’re doing both import and exports. We’re working out what we need to buy to support our customers. So what they’re using, but also what any sites that we have the capability of exporting, what they’re generating. So we do ins and outs for power.

And where we’ve got agreements with wind farms, we’re forecasting, I’m going to say forecasting wind, we’re not forecasting the actual physical wind, but we’re forecasting the energy that’s generated off the back, the wind, the industry balances daily. So we are forecasting gas to daily granularity, long term. 

Daily across three to four years at an average temperature, because again, we don’t know what the weather’s going to do four years out. Power, we balance half-hourly. So we’re doing a within-day shape as well as across the days, months, and years. Your long-term forecasts are probably, less exciting if you’re a statistician, if you like technical models, because there’s, there’s only so much detail that you want to do, because we rerun the forecast every time things like numbers of customers change, what we don’t want to do is churn the volume too much because there is balance between getting the level of accuracy if you forecast. And that means that you’re then buying and selling energy over your four years to the cost that you get from churning that energy. So because we’re trying to do it for our customers, the lowest cost, because obviously, that’s a better benefit for the customer and our main aim is to get it right for the customers.

We don’t want to almost over-trade, which means that we balance how much we put into the forecast.  We do update them fairly regularly because we obviously don’t know how many customers are going to stick with us.  We also do it at a different level. There are different hedging and trading strategies off the back of the volumes, so my team’s responsibility is to produce what we think is the best view of the volumes across the timeframes. And it’s not our responsibility to decide how much of it you buy and when, that’s what our hedging teams and our traders do.  Short-term forecasts are almost more interesting if you’re a very technical person because they’re where we run a lot of the clever machine learning, AI. Very statistical models, because you’ve got, it’s almost your final cut to get it right. It’s a little bit niche, but what we’re forecasting when we get to the short term is actually not what our customers are using, but it’s what the industry is going to charge us for it. Because both Elexon for the electricity industry and ExoServe on behalf of the gas transporters, they have a kind of set of algorithms and balancing mechanisms that determine what they’re charging suppliers. So it’s in our interest to match what they’re saying our share of it is, even if we think they’re slightly wrong. And then we work with the industry to try and correct it afterwards. Um, because if we don’t match what they think we should have purchased, they’ll charge us in balance for it and that’s a cost to the business. We don’t get that back. Essentially, if we think they’re wrong.  In the long term, we probably wouldn’t roll that in because we would aim for where we think we ought to be because a lot of the company finances roll off the back of that, but in the short term, we would almost correct for how wrong we think they are so that we don’t get charged in balance. It’s a bit of a quirk. 

Adam Sroka: It’s a really interesting, yeah, is there a material difference to the work you’re doing in that sentence around you are trying to save the customer as much money as possible. You’re not trying to like just make as much money as you can trade it. Does that play out as like, big differences in the actual work?


Sallyann Blackett : I mean, forecasting is essentially the whole load of guesses. Try and see the best cases we can. What you don’t want to do is you don’t want to put any bias into that. So if you’re trying to, if you know that prices are likely to be moving in a particular direction, if you’re thinking about that, there’s a tendency for you to set your forecast one side or the other of where you think it should be.

And that almost leads to an overcorrection mechanism. So we try very hard as a team to not look at. What market prices are doing,  and to just try and almost nail the forecast. 

Adam Sroka: Is that part of the drive to do it then, the part of what motivates you, the kind of that competitive element, I guess?

 

Sallyann Blackett: I think we’re all driven to try and get the answer right. So I mean, we are essentially data people who like numbers and the short-term forecast particularly you get very rapid feedback on whether you’re right or not. And the guys like to essentially nail the answer. So you get a lot of motivation from getting it right.  There is forecasting, I think is kind of 80 percent doing the right thing with the numbers and 20 percent having a gut feel for where it ought to be is why it takes us at least six months to train a new forecaster is because you need to learn almost to trust your instinct. And that is also why you can only get so far with the models.  We have a lot of models and we run a lot of machines. You’re learning a lot of, I guess, technical input, but it will only get you so far. We definitely know having the expertise in the team and having the people know what they’re doing, gives us an advantage, and gets us a better answer.  

Adam Sroka: Talked about the teams and things, an interesting question about like, there’s a lot of things you’re forecasting and a lot of levels of granularity, like in terms of how many teams is that? Is that one big team? Is it, do you have like a team per forecast?


Sallyann Blackett: I split the team into power and gas, but we do try and move people between the fuels, but to a certain extent, the two industries, because historically they’ve  Developed from different backgrounds and because they have different industry codes, they work differently.

You have to have people that understand the way each bit works. So I have a gas team. I have a power team. The. Wind forecasting that sits within power, although it is slightly different to the main forecast, the main power forecasting. We do, I do expect people to do short and long-term forecasts of people who kind of lead on long-term forecasts and lead on the short-term forecasts.

But I find that we get a benefit from knowing what’s happening in each one.  So whereas the short-term forecasts are very weather dependent. Cause whatever else happens, whether it’s one of the biggest drivers that you’ve got on what people are using,  knowing how people are reacting to the weather and how much difference it’s making to their behaviour, you can absolutely bring into the long-term forecast.

So while you’ve only got seasonal averages in the long-term forecast, you’re still using a relationship between cold and warm and, you know, Speed of change between the two. So I think learning benefits both types of forecasts, although they’re done very differently. 

Adam Sroka: Very interesting. Okay. So like what kind of, what, what inputs are you using? You’ve mentioned weather, we’ve mentioned market signals. Are there like interesting kinds of keys, yeah, other models that you use or lean on or other inputs? 


Sallyann Blackett: All of our models are built in-house. So we do self-development because we find that’s what works better for us. Customer numbers make a difference. We do use expertise from our business as to where they think the market is going. Because essentially we are in, at the end of the day, working out volumes for our customers. So if we think that’s changing, then that’s obviously important for us. And that’s one of the other main drivers of the long-term. We use a lot of information on sort of where the economy’s going. What are the industry requirements are gonna be.The government has a big driver on what we think it’s gonna impact to the customers. So whereas things like the cap for standard variable tariffs make a difference to what our hedging people will do.  It also drives how people behave, and that’s what we’re trying to pick up in the forecast. We can use that to work out whether we think people are going to conserve energy. Obviously, the cost of living drove a lot of that, and we saw the biggest drops in customer consumption. We’ve ever seen, and that obviously impacts our volumes, but it also impacts all the relationships we build between how people react to temperature. We have had, I would say econometric overlays. One of the biggest impacts on gas usage is how well-insulated your house is. So a lot of the trends for subsidizing cavity insulation, loft insulation. They impact how much customers are going to use, and we need to layer those trends into our forecasts. In the longer term, you’re looking at how many people you’ve got in households because that impacts how much energy people are going to be using.

We do take a view on efficiency of appliances. We’re doing, unsurprisingly, a lot of electric vehicle modelling because that has a huge impact on the kind of size of load because electric vehicles pull a lot of power through when they’re being charged, which is relatively new in terms of the. Energy industry.

So all these things feed into both forecasts, but they start in the long-term forecasts because you’re obviously looking over a number of years. We look at the national grid’s future energy scenarios, which I think is in common with most suppliers. I don’t think you’d talk to an energy supplier who didn’t look at those because they do a lot of work with the government as well.

And that then influences what you are likely to get government intervention-wise, which. It comes back round to us and industry department, government departments like Desnes who are doing climate modelling and things like that, you know, picking up that information’s useful. We do work with the Met Office on climate warming because anything that impacts long term.

Ambient temperatures make a difference to us. So it’s kind of pulling things in from as many places as we can. And that’s partly what also makes it interesting because you’re looking wider than our immediate set of customers.  

Adam Sroka: Begs a couple of questions. How do you like to prioritize all that? When do you know when to stop? Like, because I always think it’s like a proper fractal problem. Like you can just keep going. What level of degree of detail would you get into and how do you pursue one avenue of exploration over another?  

Sallyann Blackett: I mean, you could, you could keep going forever and we obviously only ever find that amount of resource. Sometimes we get the prioritization wrong. So, I think we didn’t put a lot of effort into solar generation. Until it impacted the answers, possibly we should have started sooner. It’s a tricky thing because we have a confidence interval, a range of movement in the forecast where it might just be the way the forecasting models are going. It’s not necessarily a mistake or a trend or a move. And if you’re not seeing it as anything other than noise, you’re not necessarily putting your work into, into what’s driving that. So, so low.  Bigger than we’d liked before we’d done any modelling on it. And that’s definitely hit colleagues in other regions, The Netherlands have a lot of embedded solar, which has caused the forecasting to seem a lot of, a lot of issues. So we picked that modelling up later than we should have done and we then had to force a catch-up if like we try and make sure that we’ve got enough resources. By bringing in students across the summer and giving them the pieces of work, we’ve got a year of industry students who at the moment are doing a lot of EV analysis for us market-wide half hourly at the moment, pulling a lot of energy because we know it’s going to change.

The models, so some stuff that we’ve had working for years probably won’t be good enough when that comes in, which means that you then re-prioritise what you’re looking at. Smart meters, for all their faults, have given us a lot of data and data is always good for a forecasting team. So we have used students from our year in industry, and some of our intern placements to essentially do the basic analysis, you know, grouping it up, clustering it, looking at differences, identifying what might be issues, then getting the guys in the team who’ve got a bit more experience to focus, because sometimes you find that if you’ve got a lot of data, you could spend. You spend all your time just looking at the data and never actually get any answers.

So you have to find some way of focusing into the key bits that you need to look at so that you can get some benefits from it.  

Adam Sroka: What about the technologies then that underpin all that? What kind of tooling are you, you loving, hating, like, what’s the kind of crunchy bit of like? At the moment, our little world is entirely in a bit of software called Sass, which is our main modelling tool.

We are starting to move into coding some of what we do into Python. Like I said, we self code bit that mostly causes this problem is how to get all the data into a structure that allows you to look at it nicely and easily. We are pulling a lot of our data at the moment into Snowflake, which guys love Snowflake because it has a lot of inbuilt functionality that lets us do historical trending problems.

The problem we quite often have is that our business looks at customers as they are now. This is great, but when we want to do a new forecasting model or come up with something different, we need to look at our business as it looked historically. So we need to build up a snapshot, which is why we have a lot of data and probably have the biggest server that the business has got, just for us, because holding that historical snapshot for decades. And when we’re doing weather analysis, we like to go back to 1960 as a minimum, if not further back, because 62, 63 was the coldest winter that we’ve had for a long time. So if you want to do cold weather analysis, you Go back a long way. So mostly the good bits are there’s a lot more ability to handle big data, you know, and we’ve got Azure data lakes and all sorts of new ways of holding big amounts of data, but still being able to get at it easily. Historically, that was the problem. You know, you put your data into something and it’s too big. You know, we had Oracle hypercubes at one point, which. We’re forever giving us issues because we could never manipulate them fast enough. 

Adam Sroka: Not associated with hypercube controlling, I must say.


Sallyann Blackett: It was probably wrong, our IT department got very excited because it was a new shiny thing that you could do with Oracle. Which is great if you wanted to just query it, but we then wanted to change a bit of base data which then meant that they had to recreate all these hypercubes, which took a long time. 

Adam Sroka: And you find you end up pinned to like a particular skill set or a particular individual sometimes, like you end up with like a handful of people in the company that can support the technology choice and there’s no one else in the market or it’s hard to find new people in the market so that you’re queued up behind in other situations anyway I’ve seen it where you end up queued up behind one individual or one team that are overloaded because you’ve gone a certain route and it becomes really hard to get the fluidity that you need to like just move quickly and things.


Sallyann Blackett: yeah and that is partly why we don’t like our IT department have too much to do with us, which is why we do it in team.

It does mean that we need enough people in the team who can code things. So when I first started in the team, we didn’t really do the coding. IT did the coding for us and we just, we were users, but we have found that to be able to do things fast enough for us, we have to be able to do it ourselves. What we can’t afford to do, and we had an instance of it where, um, we’d had an IT team  that had been trying to do a new model for us, and they’d said they’d got this new shiny bit of AI that could do us something that was useful.

And we’ve been trialling it, but in December we found an issue with Christmas. Now Christmas Day only comes around once a year.  It moves every day. So for us, we also only get Christmas on a Tuesday,  almost once every seven, eight, nine years. This year we had Easter on the same day as clock change, which it was nine years ago that we’d had that.

So essentially you don’t get these odd days very often, which means that you need to tweak them quite often because they’re a, very infrequent. So your data’s, you’ve got less data, but also they’re quite important. And first week of December, we’d said, right, so we need to change this model because we don’t think, well, how are you treating Christmas days any good? And the answer was, that we’re in a change freeze because we don’t do changes to things across Chrysalis because it might cause issues. And it’s like, well, okay, so I can change it for 2025, but it’s no good for this year.  Which is why we like to do things ourselves because we feel the pain if it’s not right.

So we’re incentivized to do it,  all right, in a controlled manner. Cause we don’t also don’t want it to break.  We need it to happen fast. Like I said, we are starting to think about trialling bits with Python and moving our code base. That will take us years to fully roll onto something new. So A, we don’t do it fast and B, we will do the bits that are most important first and make sure, so things have to be really controlled. Cannot afford for this to break, but we also need to not back ourselves into a corner. We can’t be beholden to one or two people who can do something because if they’re on, have issues. And like wind forecasting, we do every 15 minutes, so that’s a lot of forecasts a day. It’s seven days a week.  Yeah. A lot of it is automated, so that is also a benefit in that we don’t have to have people in the middle of the night, but we do have models running in the middle of the night. We have controls around them, so caps and collars, and we don’t let them move if we can’t if it looks as if it would be outside the range of tolerance. So we, A, can’t move quickly, but also we can’t.  Not move if we’re going to back ourselves into a technology that’s going out of support.  It’s a tricky balance.

Adam Sroka: and it’s trying to find like early wins to build up the organizational confidence that this programmer works the right thing to do and kind of defeat some of the naysayers and also like pick open technologies that are in the direction of travel for the industry as a whole so that you’re not in the future trying to hire like Informatica. Skills in 2024 and stuff, things that are like a lot less interesting to the young grads and stuff that are coming through. I’ve seen it a lot of like the bigger, it’s funny, the industry, people get really excited about it, but what’s at stake? Like what’s at play? You have to be so careful and that, that leads you down these paths of a bit more rigor and a bit more reliance on sort of tried and tested methods. 

Last couple from me, you, you mentioned AI, interested to hear your thoughts on how you think AI, machine learning and kind of modern approaches are influencing, yeah, volume forecasting. And then do you have any big bets, like a big prediction for technology that you think is going to make a huge difference in the future? 


Sallyann Blackett: I mean, it’s interesting because everyone thinks AI and machine learning are new because they’re new and buzzy. I mean, machine learning, neural nets we’ve had for decades.  The ability to use them, I suppose, at scale is probably new. The maths that underpinned them is not new. It’s been there for donkey’s years.

So, so we’ve had neural nets for decades. So I’m in an interesting position of saying, we’d like to try these new models, but yes, we’ve got machine learning because we’ve been doing it. So it’s not new. You know, we are, I guess, expertise at using it. I think the benefit is that there’s a lot more opportunity to try different technical mathematical models that you probably wouldn’t have had, you know, a PhD level understanding of the maths that underpins it that you would have needed before.

You can now do it by chucking some numbers into a program and pressing go, and it gives you an answer.  So, I mean, we’re a bunch of nerds. We quite like things like this. So the guys loved playing with them. I guess as long as we can explain the answer, then that’s good. I’m slightly cautious that people get a bit carried away with them and almost trust the program without checking the answer because whatever you’re doing, I mean, you saw it the minute that stats moved from us doing it with graph paper and mathematical models to a little computer program that anyone could use. Just because you can do it doesn’t make it right. So you do have to still be able to explain the answer and almost know where it’s got to, even if you can’t quite explain how it’s got there. 

And we are using things like E.ON’s version of ChatGPT. Because it can give you some advantages, you just have to, I suppose, not go a bit wild with it. And make sure that you always have an expectation about where the volumes are going to come out. And that’s back to the people and the expertise is important.

Adam Sroka: Love it. Well, look, I’m really conscious of time. I’ve loved speaking to you today, Sally Anne. I guess one thing I always ask just before we sign off, if people want to find out more about what your team are doing or what you’re up to, job roles you might have or things you’re looking to promote or plug, is there anywhere we can point them to? 


Sallyann Blackett: Well, the E.ON website. All the roles that we have available, and we do have quite a few coming up. They will be on, on the E.ON vacancies website. So if it’s branded as E.ON Next, if it’s our residential business or MBS, if it’s the corporates business, but E.on as a whole, and most of the peop le I work with are on LinkedIn,  all the fairly common places, I guess.

Adam Sroka: Excellent. Well, look, it’s been a pleasure. Thank you very much.


Sallyann Blackett: No problem. Always happy to talk about data and forecasting.  

Outro: And that’s it for this episode of the Hypercube podcast. Thanks for tuning in today.

If you have any questions about the topics we covered, you can reach out to us on LinkedIn or check out our website at  wearehypercube.com. You can also join Beyond Energy, our Slack community of data leaders from the sector. There’s a link to sign up in the episode description. We’re just getting this show off the ground, so if you like today’s episode, please leave us a rating, review, or subscribe wherever you get your podcasts.

It all really helps. See you next time.