Modeling the Energy Future: Ants, Birds and Dr. Strange

energy economy, just released in November. It's an Outlook all the way to twenty fifty, and his Beenest Flagship Report. The third book is called Kings of the Yukon, an Alaskan River Journey by Adam Weymouth. Here the author records his trip down the Yukon by canoe from Teslin Lake in Canada all the way through Alaska to where it lets out in the Bearing Sea. He's taking the trip to see firsthand the run of the Chinook or king salmon, and talk to the people that have been marking the

time of their annual return for generations. But for me, the most fascinating parts of the books so far have been when he goes into detail about the behaviors of the fish, for example, how they know exactly how to get to their spawning grounds down to the tiniest tributary or creek of the river without ever having been there as adults. Well, today on the show, we're gonna keep

going with the topic of animal behavior. Turns out the engine behind the analysis and NEO, the code that does the calculations and creates the scenarios in the outlook, was built after animal behaviors, specifically those of ants and birds. With us, we've got Ian berriman modeling analyst for PNF. He will tell us how the model, known as the New Energy Forecast Model or NEPHIM works and how he took cues from ants and birds and building it. We'll also talk quite a bit about Dr Strange from the

Marvel movies. So Comberbatch fans get set. Okay. Our discussion is based on BNS New Energy Outlook. Being a users can get this report on BNF dot com, the BENF mobile app, and the Bloomberg terminal. As a reminder, BENIF to provide investment or strategy advice, and you can hear the full dist claimer at the end of the show. I'm Mark Taylor, and you're listening to switch on the benf podcast Ian Welcome, thanks for having me. So today we're going to talk about modeling. I used to be

an energy analyst, and I'm not a modeler. You know, I'll be I'll be very blunt about that one. But for everybody that is listening that doesn't even know what a model is, can you just take us all the way back to the top and tell us what a model is and why they're important. It's a good question. So I think why models are important is because we don't always have time or space to try everything out. In reality, we don't have a physical way of testing

this hypothesis, don't have the materials. When we're talking about models that look towards the future, we don't know what that is until it happens. So models help us answer questions which we wouldn't otherwise be able to answer. Okay, So it helps us get through more data and ask more stuff of data that we wouldn't otherwise be able

to do. Exactly Okay, cool. I think it heard like could even be considered like like the model that of a building that an architect builds before they actually go out and build it, they can see what it looks like. You're good to see how it might work exactly. And related to that, there will be somewhere on that architect's computer, there will be computer model which has all the beams and all the weight in the building that is expected

to carry. And there will be another model which simulates winds or earthquakes to make sure the building cap with stand them. Okay, cool, Yeah, it sounds complex right, So speaking of that, how did you get into modeling? So? Um, if you can tell by my accent, I'm originally from Australia,

many many years ago. I was looking for a PhD opportunity, a kind of a vague idea of where I wanted to go, something energy related, but no, no great drive, sort of drifting for a bit, and then applied to a bunch of places and bam, I got this offer to do a PhD at Oxford. So I couldn't say no. It sounds like a pretty sweet place for to end up if you're just drifting. I mean, maybe maybe I'm

under selling myself, but it wasn't. You don't you don't just drift into there like that's there's a whole application process, like you can't just rock up and anyway. Okay, So you start this PC program at Oxford. Yeah, and the top pick which I got into was was solar power. And before I before I left Australia, I'm joking with my supervisors. I'm like, guys, like, I'm not bringing any of the sun with me. You realize you have to

provide that. It actually was. It was ended up being quite a significant part of my PhD because we did we didn't have sunlight to test a lot of our equipment, so I created computer models which could do that for us. These are ray tracing models, so you sort of you take rays of sunlight, shoot them out, bounce them off objects they will like reflect or concentrate, etcetera. So that that was probably about half my PhD was sort of working on that. That's cool, so you looked at how

the sun bounces off solar equipment. Basically, yeah, I mean we we're working on a solar powered oven. So basically a couple of mirrors which combined would result in like a very highly concentrated point of sunlight. So we could get past three degrees celsius easily and you could not just cooking bake, but you could fry as well, which was one of the big selling points. So now you're at BNF and you switch your focus from from looking at how the sun raise bounce to modeling the future

of power systems? Is that right? That's right? I mean one one naturally leads to the other. I assume can you talk a bit about that. So you're modeling the future, You're you're looking at different scenarios, you're looking at forecasts, you're looking at all kinds of things. You could model anything, right, why is the power system a good or bad candidate for looking at the future. A way to think about

this is the sort of how in the modeling. So there's there's different types of models out there, and some are more complicated than others. The simple ones will often sort of take an existing data series and just extrapolate forward a bit, and they might do some fancy stuff with the data to make sure that that that curve is accurate, but that that's sort of like a top

down approach. They're not understanding the fundamental question. The other type of model, and the one I work on is part of this category, is sort of what we call bottom up. So what we interested in is simulating the question we're looking at, which for us is the entire power system in fact, so we're not we're not blindly extrapolating anything. We were creating an entirely new power system, a virtual power system in our computer, and and we're using that in the model to look at the future.

All right, So before we started recording here, you told me of this super nerdy analogy that that I want you to make here about Doctor Strange, the Marvel character. So just for the record, everybody, UM, right now, we're London, we're on our third lockdown, and my wife has decided that we are actually going to go through and watch all the Marvel movies right now. She's listening to a podcast that goes into great detail on each and we're picking him off one by one. She's never seen any

of them. She's never seen Doctor Strange. So Ian tell us about Doctor Strange and his modeling so that there will be a spoiler alert here for your wife. Man, she didn't listen to the same rate. So in in Avengers Endgame, there's this big bad guy. The Avengers are trying to beat him. The odds are stacked against them, but they've got Doctor Strange on their side, and Doctor Strange has this ability to go backwards and forwards in time,

but also to look at different alternate realities. And he goes away, he visits all these alternate realities, and he comes back with this sort of really foreboding message that I visited fourteen million sive alternate realities. We only beat the bad guy. We only beat Sanos in one of them. So super long ards, super long ards. But this is a really good example of what our model is doing, what Nephem is doing sort of behind the scene. So you could imagine that you have a model which is

just mysterious black box. You feed a data gives you an output. That's not what we're doing. We're doing we're taking the Doctor Strange approach here. So when when we're modeling, we're actually creating these sort of alternate realities. We're creating different versions of our power system. Um, they will be slightly different, they could be very different, but they're all different from each other. And we're doing this because we're trying to find the version of our power system which

is the cheapest. The only way we can do that is by actually creating them, looking through them, and then trying to find the best one. So Doctor Strange had about fourteen million. I did some back of the envelope calculations before, and I think we did about semi million um of these. Yeah, so those are rookie numbers. Doctor Strange. Bumpers up. We're crushing Doctor Strange. That's that's awesome. That's all we need to know. So we're done, not even

in the same league. Yeah, that's really cool. So so if I get this straight, you are using you're putting a bunch of data in and creating alternate realities, you know, for the future power system, and then picking the cheapest ones to say those could be most likely. Is that yeah? Well, I mean we don't make a judgment call about which is most likely. That's yeah, that's that's the policy lessons that you draw from from the model. But we do

identify which which is the cheapest Okay, the cheapest to build? Right, Okay? When you say identified, like, that's that's the crux of this, because even even with all the interns in the world, we can't easily manually look through seventy million different power systems to find the best one. Yes, so how do you do that? So we we've got a solver, which which we use in our model what's a solver. So a solver is I guess the best way to describe

it is it's an optimization engine. And generally think about optimization. We're normally trying to either minimize or maximize something. For us, we're trying to minimize system cost in the power system. And so what a solver does is looks at previous data. So we look at we'll have a starting set of solutions and we'll look at those and then from the system costs its seasoned those from the inputs it knows it gave them, it can infer what a better guess

will be the next time around. We do this, okay, so it just gets better and better each time. So it tweaks something and and says this is higher or lower than the best or or how is that right? Yeah? And I mean to put it in terms that makes sense for for what we're doing, Like the solver might try a solution which has a bit more wind and a bit more solo. And if you're a faithful subscribe a b NF, you you might realize that, hey, those

technologies are pretty cheap. And if we added that to the system and the system costs came down, then the solver is going to recognize that, and it's going on, Hey, that that that wind and solar solution was pretty good. Maybe I'll try some more that is similar to that and see if I can't get an even better solution. And I imagine there's cases where will go too far and say that that too far and made the more

expensive again, and they'll go back exactly. So we might start adding too much, and then we'll start getting curtailment. Like some curtailment's fine, but you get to a limit where it's just too much, it doesn't make sense, and the model just backs off. What's curtailment in case anybody isn't now, So, curtailment is when the energy system producing more energy than we actually need. So this tends to be a problem when we've got things we can't easily

turn off, like wind or solar. Before we started recording, you also talked about an analogy about ants in how you build these scenarios. Really, can you tell us a bit more about that? Well, this is this is better than analogy, even better than an awkward analogy, which is my favorite type of analogy. This is this is more or less actually what's happening. So so under the hood of this solver, the solver is actually based off ant behavior in the real world, so it's an ant colony optimizer.

Is the sort of type of solver the way using is that the name we've given it, or is that the name that is out in the in the wild, in the industry, in the wild. So I mean this is this is not our software. This is like commercial software that we've bought. But it's very good. It was I think it was developed in partnership with the European Space Agency and it's holdsome world records for optimal orbital flight paths for different Shuttle launches, and I don't I

don't know exactly. I mean, it's it's it's good stuff though, Okay, amazing. I'm convinced. Yeah. The way the way to think of what what is happening? If we've got these these these alternate realities. The thing is we don't have all seventy million of them at once, like we have to fit these inside our computers. So normally we're doing, depending on the computer, somewhere between four to eighty of these simultaneously and what the answer doing, and they're carrying information about

these alternate realities to and from the solver. So we'll create a generation of ants. They'll go out, each ant, we'll go to a different reality, report back the system cost, and then our solver will create a new generation ants based on the information that the previous one provided. And at a very high level that that's essentially what's happening. And it's the mathematics the underpinn all this are based on like path optimization and and the way the ants

will do that in the real world. And in the end you find the least cost futures is that right? There will be there will be one lucky ant that finds it, one lucky ant, and so let's get into that. So you you've found that in this year's edition of well Napham the New Energy Forecast Model, but which was the basis for the New Energy Outlook, the or NEO, which is bens Flagskap report. Is that right? That's right? Okay? How does an ant, an individual aunt know when it's

found the cheapest system cost? So that's a that's a good question, because an ant doesn't know how to calculate a system cost. So for all the good work that the soul is doing, that's that's really just the top layer of what's going on in this model. So most of the code what's happening is the actual calculation of the system cost, and that fundamentally is what NEPHEM is the New Energy Forecasting model. It's creating from a given set of input data, simulation of the entire power system

so that we can actually perform that calculation. In the methodology document, it says the model solved for a capacity mix that minimizes system cost while ensuring I really demand is met for an entire year. You Mike, sure all the demand is met, putting all that data in, and then it says, okay, this is the this is the mix that minimizes system costs. Is that right? Yeah? So I think maybe the best way to think about it

is is each and is a different capacity mix. And what I mean by capacity mixes we look at all the different technologies we have available to us. When solar, coal, gas, nuclear, that exactly exactly, and we'll have a different mixture of them per ant, So some will have more wind, some will have more call, but most of the ants are different from each other. So what's happening with this AUNT carries the mix, and then Nephem is saying, from this given mix of technologies, this is how much the power

system costs. That's cool, and so then one lucky ant is the winner. Well, yeah, I said one lucky ant. There's there's a few ants because although those are global analysis, we break things down regionally. We don't model every country in the world simultaneously, so there's there's quite a few ants that win. What regions do you do? How do

you put it out? So Europe is nine regions, So the larger countries we all model individually, so UK, France, Italy, Germany, and then the smaller countries we lump together and they'll there's a north, south, east, West, and then we do iber Area, which is Spain and Portugal together. The US, there's thirteen different regions we do that by ISO. China is six different regions. India is actually just one region,

which is our largest region ends up being our largest region. Okay, we're going to take a short break and when we come back, the new Energy Forecast model with the environment, can you tell us a bit more about the data that you need in order to run the model and how you got it. Yeah, So there's a huge amount of data that goes in into this process. And I guess the question you're asking is how many data points do I need to like synthesize a power system for

a given region. That's one question, But I'm also just really curious about how you went and guard the data, Like, I mean, some some of it will come from the terminal, some of it's our own data. A lot of the cost data, particularly all the renewable cost data, is all internal. In fact, I think every cost cost data point is internal. There will be commodity prices which come from other teams inside ben f and Bloomberg. There will be demand data

which will come from Bloomberg Economics. There's also demand data which comes from the real world, So we'll feed in real world demand series for electricity demand from the countries where we have that data. How do you go about assimilating all this data and planning out the operation of the model. So, like, to me a non modeler, that just sounds this sounds like a nightmare logistical task or planning task. How do you go about planning this model for someone who has to do it? It is also

a nightmare logistical task? Okay, I mean we we so we have a database. There's a separate like well defined data BRASE which when everything's filled out, that contains every single piece of data you need to run the model and the processes to get the data in there. There's actually a couple of different ones. We use a lot of Python scripts to shunt data around, scrape data from website. It's there will be some manual points like not the big series, but some of the smaller ones will will addit.

Those the cost data. Again, this is this is all a huge data sets that are all coming in because when when I say something like the cost of solar, it's not just one data point. It will differ by country and it differs by year. So we need to know that what the cost of solar is in a bunch of different countries from now until. Okay, that's no

small task in it itself, right, No, Yeah, seems pretty intense. Okay, I was curious, would you say that is the hardest part or what is the hardest part of putting this particular model together. This has been so many hard parts. I think the storage algorithm was quite quite difficult. Obviously, when you're modeling an entire power system, there's different types of plants, and they behave differently. A wind or solar plant is relatively easy to model from the real world.

We have this data what window solar generation looks like across a given representative weather year, and we can put different weather years in the model if we want. But for something like storage, like a battery, you don't have a well defined data series which is what the output

is meant to look like. For a battery, the outputs the function of what the rest of the system is doing, because if there's a lot of cheap energy, or even excess energy, then the battery wants to charge during those hours, and if there's a shortfall and supply or want to discharge in those hours. So that's a completely different type of power plant and takes a lot of extra modeling effort inside the model to actually calculate. What was the

most surprising result from the exercise. I think there's a lot of cool results that come out of this. I think some of the more surprising ones are to do with some of the weird scenarios we've done. I'll talk about a scenario we we did where we had very cheap batteries, and I think most of us would assume if you have a power system and you add a bunch of cheap our ties, that the winners are probably

more likely to be renewables. And that's I think it's an intuitive thing that that we have, but it's it's not necessarily correct, and not at least not in the strictest sense, because if you think about a battery, what batteries are trying to do is to smooth out to

supply and demand imbalance. If I had a battery to this picture, what happens is that battery will see the supplied demand imbalance over the course of the day, and the battery says, hey, I want to charge at midday when there's like a bunch of extra solar, and then what I want to do in the evening is discharge because that's where all the demand does. And if you do that with a battery, you'll end up raising demand from the rest of the system at midday and lowering

demand in the evening. And hey, like that helps the solar guy. It helps the battery, but that's not the end of the story, because it also helps the gas plant. Without the battery, there's going to be a gas plant somewhere in California, which is watching demand fall at midday, they're like, do we turn off? We aren't turn off. There's like thermodynamic properties inherent to that gas plant that means it's very difficult to switch off and on instantly.

They can't really do it at all. So they're they're in a lot of pain at mid day. So they're turning their plant right down to the like minimal technical feasible limits they can, hoping that like demand doesn't go any lower and they have to turn off, and then they breathe this huge sigh of relief when demand comes back in the evening and they can turn their plant

back up and and make their money. And when I had batteries to the system, they don't have to turn down during the midday, or if I had enough batteries and they don't, they get to sit there and operate the day through because the batteries are taking the power

from produced both solar and thereby decreasing overall demand. Is that right during the midday, Well, it's it's it's all like if you if you're a gas plant, what you get paid over the course of that day is very very low at midday because demand or net demand so low and high in the evening, and if I increased demand at midday, then the price goes up and I get paid more in midday. I might also get paid

less in the evening. But like the gas plants will have extra additional costs if they're constantly turning their plan up and down all the time to sort of make room for the rest of the grid which is going crazy. So what is optimal? Is it optimal to have a bit more expensive batteries or what it's It's a difficult question to ask, and that the approach we've taken is

perhaps a little bit different. So you can write out all the equations for this behavior in a specific set of instructions and feed it to a computer and get a result. It takes an incredible amount of time, but you can fully optimize that problem. But we've got slightly different priorities when we're producing NEO, because we're not interested in a fully optimized, guaranteed perfect dispatch for a system. We're interested in in the larger picture. We're interested in

comparing systems. To make sense to this, I think it's maybe maybe useful to borrow one of the the strangers alternate realities. For a moment, there's an alternate reality where where we've already solved the energy transition, things have gone great. Being if still a company that's good. And in this sort of zero carbon utopia, people have all these new and interesting hobbies, have got all the spare time, and bird watching is huge. So imagine that in this alternate reality,

the B and b n F stands for birds. Birds. Birds are big business. Everyone loves birds. And the CEO of BENF comes down and he's like, man, people just love this bird. This bird content we're producing. We need more. And he goes to the best analyst and he's like, you have you seen those swarms of starlings. That's sort of clouds that undulate and pulse pulsate across the sort of evening sky, like I think most of us have seen them, and that it says like, yes, sir, it's

it's actually a murmuration of starlings. That's the collective now. And the CEO is like, you're the man for the job, or you're the woman for the job. So this analyst goes up north. It's very smart, smart person, and they're trying to quantify with a set of equations what this flock of birds is doing, and they're looking at the whole flock, they're looking at the macro and they're trying to figure it out. They've got a PhD in maths, but they just can't do it. It's just too complex.

They come up with sort of rough equations, but it doesn't hold in all circumstances and things are just falling apart and they're pulling their hair out. And the reason is because they've they've taken the wrong approach. So if you if you want to model this flock of birds, a better way to do it is to zoom in on the individual bird. So birds aren't particularly intelligent, like the expression like bird bird brain is there is there for a reason, like they bump into windows all the time.

And birds are only actually following a few simple set of rules when they're flying in flocks. So the first rule is basically like don't bump into my neighbors. That's pretty simple. The second rule is sort of go roughly in the direction that my neighbors are going in, and the third is kind of like gravitate towards the center

of mass of the flock. And it sounds crazy, but if you model those individual birds and you program those three rules into them, that amazing flocking behavior which we've all seen, sort of springs up out of those rules. So it's called emergent behavior. And that's a much better way of taking this problem and making it something digestible

that we can get answers from. And so in our reality, we we don't have birds, but all we do is we have power plants, and we can sort of program these simple set of rules into the power plants and then sort of set them free, if you will, to

to follow their own behavior. And that's how we stitch the whole system together, which is different like if we top down, if we tried to ask the question like what's the optimal mix at twelve o'clock on this given day for this given demand condition in California, Like that's a very difficult question, Like you can't solve it. It just takes a very long time. Our approaches say, well, look, if you will, like let the power plants decide for themselves.

We've given them all the cost and operational data they need. They can make that decision for themselves. So if I'm just random gas plant, you know in Nevada, or whatever I can based on certain conditions that you give me, I will make a choice on what to do at that given hour. Yeah, like if if the price goes too low, demand is too low, like I'll switch off if you switch the whole thing together. Essentially, what you have a bunch of ants scarring alternate realities, trying to

choose the cheapest flock of power plants? Simple? What can NEHEM do that you haven't yet investigated? So what Nehem can do that we haven't fully explored yet is a lot more work on on zero carbon. So we've done some emissions scenarios and they've been very instructive. But what we haven't done yet is eached up everything together into a cross sectoral optimization. What I mean by that is at the moment we've got a view on decarbonization pathways

for steel for example, and for other sectors. There will be electrification as a decarbonization pathway. So transport is a good example of that, where we get more electric vehicles as long as our electricity is zero carbon, that's fine, but it makes the electricity system bigger, so it makes that problem more difficult to solve. So we've done quite a way into sort of stitching those together at the moment,

but not everything and not fully optimized. So you have all these countries that are making net zero targets and you want to model out what that could actually look like in practice exactly. And I mean there's a there's a difference here as well, because we can take the sort of self nominated targets, sort of particular countries done and model lat That's a very different question from what's the global optimized least cost pathway to zero carbon system?

Because it might mean some other countries we haven't thought of do more heavy lifting than others. It's a it's a different story as a whole, rather than the individual country story stitched together. One final question I guess is what would you like to change on the model for next year? Definitely the name. I nominated a bunch of them, but they were all shot down. Most of them weren't serious conversations Global Renewable Energy Transition Analyzer. So the acronyms GRETTA,

which I think is very appropriate. Oh man, that's gold. I love it. Why would you not do that? That's that's great, that's so good. Ian, Thanks for joining us, Thanks for having me Today's episode of Switched On was edited by Rex Warner of Great Stoke Media. Bloomberg Guny App is a service provided by Bloomberg Finance LP and its affiliates. This recording does not constitute, nor should it be construed as investment advice, investment recommendations, or a recommendation

as to an investment or other strategy. Bloomberginn EPP should not be considered as information sufficient upon which to base an investment decision. Neither Bloomberg Finance LP nor any of its affiliates makes any representation or warranty as to the accuracy or completeness of the information contained in this recording, and any liability as a result of this recording. Did expressly disclose