 So I kind of struggled with the title today, because what I wanted to do is talk about some, I guess, some real world examples of what I've seen for battery optimization trends in the industry, as well as DR optimizations and DR being distributed energy resources, effectively virtual power plants, at least the way that that I've seen them. And I tried to tailor this talk towards stuff that you probably wouldn't learn here at Stanford unless someone like me came and talk to you so I don't know if that's true maybe you've already heard this stuff. And so maybe it's somewhat of an ambitious title but that's the intention of this talk. So really quickly, I've already got an intro, but I figured I'd just give you my own intro as well just to give you a little bit more flavor of what I've done in my career. So I started off my career in consulting, so I work for a company called Charles River Associates, or based in Boston, the working DC, and I was doing a lot of modeling of the electricity grid and trying to figure out different problems. The big problem at that time is cap and trade policy and federal legislation for, for difference, noble incentives. So, there was a lot of questions about how are renewables going to affect the economy, are they too expensive, etc. And so, did that for a few years and then was really passionate about this question, are renewables too expensive for the grid can we actually have a bunch of novels and, you know, to stay in our society. And so that's what I focused on the PhD at Carminello. And really, what was like very obvious when you start thinking about, okay, how do you have renewables to the grid, is you start something upon spending storage and battery optimization became really the focus for me because I was trying to figure out how to take these new projects, pair them with batteries and make money, not just some academic exercise, not some pilot project, but like what's the profitability from buying renewables and good products. And so that was a major focus of my PhD. And then I took a couple of professional distractions during my PhD, but now I'm a professor at Stanford, this is not too pleased about. But I spent some time at RPE, which is the Advanced Project Research Agency, as part of the Department of Energy, and I helped them think about, okay, how do you use new battery technology and different applications. And then I spent a year in London, which is really mad about working for EDF renewables, where I helped develop offshore link farms, and what perhaps is not super well known is that the UK has a lot of width, on the order of like 10 gigawatts today, and a huge pipeline, so they have really world leaders there and so I learned a lot about offshore width. Shortly after my PhD are actually towards the end of it, I founded a company called Waller. And so after all I've learned about batteries and thinking about how to optimize them and make money with them. There still really wasn't up until about halfway through my PhD there really wasn't a product that you could go and buy in volume. They're just the industry hadn't really taken off. But in 2015 or so Tesla now it's a couple of battery products that became very compelling in terms of price point and terms of volume that you could actually get your hands on these battery products. So the power wall which is a residential battery, as well as the power pack which is like a grid scale that. And those came in about one third of the optimistic cost projections that academia was projecting. So it was about $300 per megawatt hour compared to about $1,000 per megawatt hour, which was what most people were kind of projecting so that just transformed the industry. Everybody was rushing towards developing battery projects so your renewable developer you develop wind or you develop solar. There was a massive interest to start thinking about how do I develop a battery project. But unlike those other projects and really conventional generation batteries are really complex because you have to figure out what when you're charging and when you're discharging, because you're kind of both sides of load. Right so you can either consume or you can you can generate compared to a generator which is only on the generation side. That makes the dynamics and the optimization really complicated. And so I founded what learn to help the industry kind of think about this, hopefully alleviate some of this confusion and provide AI software to automate the operation at the grid grid scale. And so that company had a speedy growth we brought in early investors, early customers and then one of our customers actually made sense for us to get acquired by them. And so we had a good exit from that in 2018. After that I ended up joining Tesla. That's not the company that acquired us. I actually can't tell you what company did acquire us because it was part of our deal but just full disclosure was not Tesla. But I ended up joining Tesla and the auto bitter team and we'll talk a lot about auto bitter and what that is but effectively it's automated AI software for operating grid scale batteries as well as the aggregation of small batteries whether that's residential or commercial like on a campus for example, or a factor. And then after that, my co founder that founded what learn with me founded another company called power line. And after about six months of running that company and raising a pre seed round. I jumped on board as a co founder because I'm just really excited about what we're doing there. And so a power line we'll talk about that later. The summary is that we are building a platform to connect the drivers to be able to create opportunities where you can use your car as a power plant. Okay, so I have four sections of my talk, just, you know, high level background of the industry batteries that kind of thing just to kind of lay the foundation. And then the kind of the meat of the presentation are these these kind of three topics so one of them being great scale batteries so talking about a lot scale batteries like power plants. VPPs. So virtual power plants that's like the aggregation of a bunch of behind the meter assets like residential batteries. And then operations in both of those and then I'll talk about power line what we're doing now. And so yeah first off, I think this is probably obvious to this group, but I'll just start with this. And so the electricity grid is changing quite a bit so since the, you know, 1800 or so we've had basically a one way flow of like power generation you have transmission you have some transformers to step down the voltage distribution lines and then really just a one way flow of electrons, but the grid today is very different. The grid today has all sorts of complexities because now we don't just have thermal assets, we have wind solar, we have batteries at this stage you know as power plants. And then we also have batteries and other DRs was called distributed energy resources kind of along the supply chain of energy. And so you have batteries that could be at the trans, you know, transformers or at distribution level or even at individual homes or buildings. And so now, of course, solar and wind as well. And so you have electrons are going this way, electrons that are going this way. And it's very dynamic system, and very different from the grid that was really designed and hasn't really changed frankly, in terms of the infrastructure for for over 100 years. So that's kind of the current state that we're in. And with the advancements and AI and the advancements and various hardware components like batteries and solar wind. We're starting to see a revolution and how the grid is shaping up so it's a very exciting time to be working in energy. The other important background information that's necessary to kind of describe is the fact that there are deregulated electricity markets. What this means is that that entire supply chain that I talked about for a long time that was just owned by one company. So there was a company that just own generation, the transmission lines, the distribution lines, and then you send, they send you a bill, and they controlled all of it. That's no longer the case in a lot of markets. And so you have basically separate companies that are participating and providing different services along the way. And the way it works effectively is through a marketplace for electrons, which, at least over on this side with the generators that's going to be the main focus of my talk is the fact that you have a wholesale market with a huge amount of volatility and what the price is for electricity. That really drives decisions for generators. And it's not just like one generator is deciding to, you know, come coming up with some rule and how they're going to be providing energy on a given day if there's a marketplace for buying energy and as well as, yeah, buying energy and selling energy and there's, you know, market clearing price that determines what the wholesale level price is. Now, at the end of the day, you don't us like consumers so if you have utility, I don't know how like dorms or whatever work hit at Stanford but the utility that you pay is giving you like a fixed rate tariff. That's basically meant to cover their costs but they are exposed to a enormous volatility on the other side. And that's the volatility that I want to talk about and how AI can really help. Oh, and real quick on this is that I don't want to get into the nuances of different market structures but markets are organized very differently depending on where you are. For the most part in the US, you have, you know, California and Texas and New England and New York is effectively a nodal market. And so there's this like supply and demand marketplace wholesale market for each individual node across the region. So like in California, for example, there are thousands of nodes. So the price for energy like here versus like five miles away like maybe in like Menlo Park or whatever can be very different. Because there's a supply and demand difference between those different nodes. Not all markets are like this but in the United States that's how the structure. And we'll talk about Australia a lot in this, in this talk because it's a very cold market, which deviates from that slightly, but we'll get into that. Okay, so the benefits of batteries, why are batteries useful. One of the main points here that, you know, maybe it seems obvious but it's worth calling out. So first you can consume energy and you can supply energy. This is very useful on a grid that is extremely dynamic where electrons are again coming from one side of the supply chain to the other. And so if you can kind of control that. And then then you can somehow control this volatility and take advantage of that. And then there's fast responding, which means a few things. So one thing is that yes, they can respond quickly to energy to providing energy but they can also provide what's called ancillary services very well. Ancillary services are basically like services that you can provide that require typically like a fast response. And we'll get into some examples of that but not all generators can actually provide ancillary services kind of depends on what the requirements are for that service. And batteries are, frankly, like the best asset that exists for providing services. And so, if you have a battery you can provide pretty much all the ancillary services in the market, in addition to supplying energy. And so, just a couple stats on that so they're like 21 different grid services in the US that batteries can provide. In terms of capacity. Right now we're probably around like 20 to 30 gigawatts of storage in the world, like roughly, it's like 17 as of 2020 but there's been a bunch of progress in the past couple years. But we're tracking towards like an order of magnitude increase by the end of the decade. So this is a fast growing industry it's not like they're just a couple batteries around. It's growing at an exponential rates, especially in California or caught a Texas and Australia and in the UK, those are those are really hot markets for parents. So let's talk a little bit about ancillary services. So they're basically like three things you can do to provide energy to the grid or provide services so there's the basic thing. And what people I think mostly think of in a wholesale market is energy. So with a battery you can charge from the grid so you can, you know, bid to or bid to charge or to consume energy. You can also bid to discharge. And this is pretty straightforward, like for a certain block of time, like an hour, five minutes, 15 minutes, whatever the market is. For that entire period you're basically consuming or discharging at a constant rate. And so that's why this you know this blue filling here in the chart. You pay it in dollars per megawatt hour. Okay. There's another category which I'm going to call contingency contingent capacity, which is rrs for those that know our costs or it's like fcas in Australia. Basically, it's a reserve product, you're not expected to do anything. Most of the time, you just sit there, just in case something happens on the grid and they really really need capacity. And it's not that it's very rare, maybe like one time per year, for example, and you're paid for capacity you're paid dollars per megawatts so if you bid like 10 megawatts, you get paid. You know, dollars per megawatt so for those 10 megawatts that you're available. And most of the time you actually don't have to do anything with the battery so these are actually really good for cycling with a battery because you don't actually have to cycle the battery a lot. As long as you're available you can get paid. There's another category which is called regulation. And so this is typically called in a lot of places, people call it frequency regulation that's actually not true, because it doesn't necessarily have to be directly tied to frequency on the grid, it often is. But the regulation market, what's what's unique about this is that it's similar to the contingent capacity, you get paid for the block of capacity, you're providing, but the obligation is to follow a signal. And that is not from. That is not obviously known. So you have to interconnect with some kind of third party that's sending you a signal for what you should do at any given time and that signal is typically going to be greater than zero so you're going to do something over that period and it's typically a high frequency signal so like people refer to it as AGC automatic generator control. And it's typically like four seconds or something like that. So, so there's there's like definitely some uncertainty here and what your obligation is going to be. And so that plays into the algorithm, for example, if you like say, yes, I'm going to be available for 10 megawatts of capacity. If you don't have enough energy to deliver on that then you fail your obligation, but then again, if you over commit your capacity, then you're kind of missing out on other things you could do with that. So there's a there's a trade off there and that's why regulations actually really complicated, but it's the most common answers. And then there are kind of weird mixes. And I'm happy to go into that but in Australia there's something called F cast, which is a little bit of regulation and a little bit of contingent capacity, like provide regulation but only in certain situations. Okay, so that's it for the background. Let's dive into grid scale battery operations and some of the things that I've learned in the past several years. And perhaps this is obvious, but I'm just going to state it anyways, AI is absolutely taking over in terms of energy trading, and it really starts with that. Because for the longest time, most energy companies like the big companies like EDF, where I spent some time, next terrors of the world and all these other large and reputable companies have these massive trading desks in Houston. They managed to trade energy with their solar farms gas plants coal plants, etc. There aren't obviously some analytics involved in some automation, but for the most part, there are humans making trading decisions kind of like Wall Street. That has changed dramatically with the introduction of batteries, because it's very hard to figure out how do you plant a battery, which is like typically a two hour battery you're not going to build a battery that's like 10 hours 12 hours. It's going to be a two hour battery. So that's the short duration. That means that you have to figure out when you're charging when you're just charging for the key opportunities of the day. It's a complicated problem that humans do a really bad job. And so, and just a couple points on that. So first of all, like profitable, profitable grid batteries did not really exist until about 2015. And I'm giving that I'm giving the industry a little bit of slack there. It's probably closer to 2017. But pretty much every grid battery that was built prior to that was really a demonstration or it had a huge amount of subsidies. It wasn't something that was just fully quote unquote merchant where it was just bidding into the wholesale market and making money, because the cost of batteries were just too high and the marketplace wasn't really ready to accept batteries as a participant in the market. And so this is very new, which is great because there are a lot of advances in AI that really do help. And starting with a brand new technology to really excel. And so, there are kind of two points that I wanted to make here about why it's so much better to use AI. So in my experience, I built a operations like the operator operations team at Tesla so like I effectively like an operations team that managed the fleet of batteries that Tesla operate. And so based on that, and that experience and based on, you know, some of my knowledge about other other teams and some of my industry research, it's equivalent to about like so if you have a good algorithm. You need about half an engineer to manage a portfolio of batteries, roughly. Which is equivalent to about five traders. So it's about a 10x change in cost. And, you know, anybody who works in software knows that like most of the software cost is like human capital. Right. So this is like the majority of costs. So it's about a 10th of the cost and it's about a 40% increase. That's like the minimum that I've seen. You know, it could be like multiple, multiple is greater than that, but there's at least a 40x improvement in the revenue, because you just find opportunities to get to optimize over multiple products that you a human would just pick up because of the dynamics of markets and the dynamics of the battle. And so I provided here and this is just a screen show out of Tesla's product. This is, you know, something that's in their marketing materials and stuff. So nothing particularly like interesting there, aside from just introducing to those who haven't seen it before. And so this is like auto bidder for Australia. And they're basically, it's an example of a multiple letters just up there today. But what you're seeing here, and I'll just walk through it really quickly is on the top is price so this is, I don't know if like, I see a bit of a glare I don't know if you guys do, but I'm just going to point. So, so here is like now, like the time now and then everything to the right of that marker is a forecast. So you have some kind of price forecast that you get for different products. Here is what you have for your planned bid so it's like a planning schedule of what you plan on doing in the next, say 24 hours. And then down here is your state of charge. So how much energy do you have in the battery so when it's really high that means your batteries full when it's really low that means you have no energy left. And so what you can see here is that this green line below this line is negative prices. So I don't know if this is like a weird concept to you guys or you guys think about this all the time but actually in the energy markets. I mean, in particular that product, you can have negative prices so what does it mean to have negative prices. That means that there's congestion at a particular node. There's too much energy coming from either solar or wind, or some other generator that cannot shut down. And so there's not enough consumption at that node and therefore prices go negative to incentivize people to consume energy, which is great for a better. So this is an example of prices go negative. The battery is planning this green here to charge over that period of time as well as providing into a services. And so these like gray and blue bars are contingent products, and then these like purple and red bars are regulation. You can see energy is kind of going up over time to take advantage of those things. So that's just an example of like what an algorithm would do. And the cool thing about it is that this is running all the time. You have to make one decision and just hope that it works with AI you can benefit from that automation and so things change prices, prices shift. Then your algorithm can pick that up. And it never gets tired or grumpy or demands a raise. So it's actually really good. So these are just a couple of example projects. I worked a lot on these projects. So that's why I put them here, but there are many, many other projects. I'm not trying to be. This is not a Tesla talk. This is just, this is what I know best with good scale batteries. And so there are kind of two main ones that a kind of flagship projects from Tesla, one of them being Hornstail, which I think is the envy of the industry, because when it was built. Everybody kind of thought, okay, maybe it'll be profitable. Maybe not. But then prices really took off in the market and the performance was really good. And so it was kind of an example of what you can do with the matter. And so since then there have been many, many entrants into the Australian market, which is where this is because of that. So it kind of led the way in Australia, but then also kind of started, you know, folks thinking about other markets like Texas and like, like California, which we're seeing today. There's another one that, you know, both of these were the biggest batteries, their time, I help commission with them too, which was kind of cool. So, so this one, got up to like 150 megawatts in 2020. So it went from 100 to 150. And at that time it was the biggest battery in the world. And a year later, Victoria Big Battery BBB, which is also Australia 300 megawatts was also the biggest battery. I don't think that's the case anymore. There are a couple other battery projects that have been built, I think in Australia, in fact, that are quite a bit bigger. Okay, so let's talk about a general AI modeling framework for batteries. So how would you like try to think about, so this is mainly conceptual we're not going to go through like a bunch of equations. So I'll spare you that. But yeah, let's talk about the concept. So the first thing you got to do, which is pretty straightforward, I would imagine is like, well at least conceptually doing this is pretty hard, but forecast signals. And so the first thing you need to do is like, forecast, you know, wholesale prices, for example, that's a main signal that you're going to need. Obviously you need to know what the prices in order to make a good decision, potentially solar generation or load or demand or whatever other factors. These are just examples but whatever other factors that your algorithm is going to need to make a good decision. So the first thing to do is you need to have good forecasts. And so there's a whole like discipline and what techniques to use so you could use like linear regression obviously is a basic one to start with. And it can be very powerful. It can be very complex as well depending on what kind of variables you use. You could use neural nets of course, or you could use physical models. And so what's cool about the energy industry is it's not really just, it's not your typical data science world where you're like you kind of, kind of have these like, yeah, use SK learn and then you're like okay what kind of algorithm so I have let me try them all. There's actually kind of a unique set of tools that you can use in the energy industry, including physical ones. So like, if you wanted to, for example, try to solve for what the market would clear the price at you could do that. You build an economic dispatch model like a general equilibrium model, solve with the market clearing prices and then use that as a way to forecast prices as well. And so there are a lot of different techniques you can use. And this is just an example I just pulled this from Amo. This was a couple days ago. Oh, I have another one. That's a little bit better. But yeah, these are Amo prices. The cool thing about Amo, which is the Australian energy market operator is that they publish their own forecast as well from their, their market clearing model. So when they're trying to clear the market, supply and demand, and they set the price for that particular interval, they publish the forecast going forward for all future and they make that available to participants. And so you can see what the market is at least thinking in terms of what the price agree, which is very, very useful information. Okay, so the next thing is having an optimization or a planning tool. I say this because optimization techniques are useful. And that's sufficient, probably not, but it's definitely a convex optimization can be can be quite useful in this. Whether that's an LP a linear program mixed integer nonlinear. It kind of depends, or even a stochastic optimization model kind of depends on on your, on your problem but that's typically like, I guess some some tools that have been discussed in academia at length, including during as well as in industry as well. And so you have basically you try to solve for like a plan. So here's what I'm going to do for the next 24 hours or whatever it is 24 hours, two hours, whatever it is. And then the last thing is control. So, once you have your forecast you can then figure out your plan once you have your plan you can figure out what you're going to do right now. So maybe this changes over time or maybe not, but control probably will be changing a lot depending on what's happening. So, yeah, in summary, there are basically three main components forecast optimization and control. Okay, so let's talk a little bit about. So this is this is kind of getting into some of the nuances of trying to operate a battery that's running with these automated algorithms. You know, suppose prices are forecasted to spike. And so this is the chart I was thinking of. I just pulled this the other day. And so the cool thing about Australia is that it's actually a very volatile market and it has very high price cap. So the price cap in Australia is $15,000 per megawatt hour. Does anybody know what it is in California. Hi so the market cap. It's $15,000. The stuff like that. Yeah, in ERCOT it's $4,000. If you add price adders it might be a little bit more because ERCOT has this weird like price adder mechanism, but still like in Australia it's $15,000. So that's that's a huge amount. And then on the low side it's minus $1,000. So you can you can pay to charge at minus $1,000, which is really quite lucrative. And it's, and the other cool thing about Australia is that it's only a real time market. There's no day ahead. So there's no like pre scheduled capacity for any given time. It just runs every five minutes. And if you know demand is short, or sorry supply is short and prices will spike. And it happens regularly, which is why it's such a great market to have a battery because you can really take advantage of price bikes and it literally happens pretty much every month you're going to see a price bike of $15,000. And this is proof. I just, when I was preparing these slides, I just pull open a most dashboard and try to look at what the forecast and prices are. And sure enough, right when I pulled it up, there's a forecast for $15,000 prices. I don't think it actually materialized because I think that was yesterday in Australia time, but nonetheless, this happens a lot. Okay, so let's suppose this was the case. Here's your in the purple line here this is the price forecast. And you can see like a $15,000 spike around, you know, peak peak hours, but 1800, 6, no, 6 PM. So, here's what like, and this is like a really basic example I didn't like run any sophisticated algorithm to generate this but this is like, based on my experience this is what would happen. You would basically want to be doing a lot of reg up and down, prior to the price bike, why, because regulation typically comes, well, as we discussed, there's a capacity payment that you get for just being available and following that agency signal. And so you are going to have to follow a signal and so you cycle the battery but the good thing is that you get paid for the energy. And so if you're cycling high during like the morning peak you'll actually get paid for that. For that energy and frequency tends to deviate in the direction of higher and lower prices so prices are high for energy, probably the case that you're going to have an agency signal that takes you up to discharge. And so you could, you know, do a bunch of ancillary services before the price bike. And typically algorithms tend to get greedy. And so they'll charge like right before the price bike, just to try to, you know, take advantage of basically all the other opportunities ahead of time. And so it waits until the very last minute to charge charges up for a couple hours, and then discharges for the peak that that would be like a typical kind of naive algorithm, basic algorithms that that you could develop. Let's say like a basic optimization. There are a couple problems with this. So one of them is that the spike. What if the spike comes earlier. So if the spike comes earlier, and you know we have to shift, we want to shift our charging and shifting our charging might be really hard because we were planning on charging, perhaps right before the spike happens. And I've also plotted here the state of charge. You see that you're basically writing at like 20% state of charge. And if you're writing into a peak with 20% state of charge, that'd be a really bad day. And so on my team I'd be, I'd be pretty pissed. And so, so that could happen. The other thing that happens could happen is that the spike never materializes. And so you're preparing, preparing, preparing for the spike. Maybe there are some decent prices that materialize that work 15,000 but maybe 1000 or 500, but you didn't take advantage of that. And so you're leaving money on the table. So these are a couple risks that you have to think about when you see your algorithm deciding making decision like this and try to try to decide whether it's good or not. And so let's just walk through it like a couple of those cases really quickly and try to decide. What I encourage you to think about is that it's not just about like, Oh, was I right or was I not it's what is the alternative and trying to analyze what that alternative will be. And so let's talk about two potential alternatives to that forecast. So case one is that suppose that the alternative to the price spiking when we thought based on the forecast was that the price spikes two hours earlier. And so you see here in this charts, actual prices are in orange blue is the forecast. And so this is a bad situation so effectively if you were to look at like your base strategy just wouldn't be able to tell you to do. And then you look at some kind of perfect stretch run your algorithm and say perfect knowledge or something like that. If the spike does happen and you were right that in both cases, you make about 293k with the 10 megawatt two hour battery. If I know if no spike occurs, you know it's pretty huge delta and what the spike is versus whether the spike materializes or not. And more importantly, a perfect strategy that actually captured that spike earlier would have made, you know, about 170k. And so if you look at the risk here is about 175 k that you missed out on because of your algorithm not being correct. So compared to a reward of about 300 the delta between these two. And so I would say that's a pretty risky situation and you might want to intervene and maybe not go through 100% with what your algorithm told you to do. But here's another case like suppose that the alternative was just a normal day that the evening peak would be about $500, which is typical. And the price spike just never happened. Well, in that case, if the spike happens we have the same revenue. If there's no spike kind of similar revenue, it's like closer to zero. But in the perfect strategy and like just a normal day you might only make like 11k. And so the risk in this case is much lower, 7k versus you know making close to 300k. And so what you do at length in operations is scrutinize these types of situations and try to dig up as much information as you can to figure out what could be the alternative. And so one example of an of a normal day is that if you look at all of the signals that typically influence a forecast, and none of them really telling you that there's going to be a peak or a spike. And then maybe that spike is just an artifact of, I don't know some kind of perhaps a plug in your algorithm or perhaps it's overfitting on a particular day that's not relevant to today. And so, maybe you just let the algorithm ride in this scenario. Whereas, if it's this scenario and you're like there's definitely going to be a spike today, but we just don't know when it's going to happen. You have to be overly cautious and maybe start intervening and tuning your algorithm to take advantage of potentially a spike that could occur over a longer period of time. And so these are the types of decisions, and you can imagine all sorts of tools and tweaks for the algorithm to be able to achieve these types of scenarios and kind of tune the algorithm. And so that's why you still have half an engineer, at least today, and a lot of these situations so that you can actually make some of these calls and, you know, not get fired because you missed out on a huge option. So as a result of these uncertainties, what I would encourage you to do like if you're if you're ever putting the situation where you're responsible for like the P&L of projects that are driven by AI. In my opinion, the best way to evaluate these algorithms is by doing scenario analysis, rigorous scenario analysis with simulations and that. And so, but what I would encourage you to do, and, you know, as you get jobs and go different places and do awesome work. When you think about developing a forecast, for example, like one component of your system. Don't look at that component in isolation. Because if your MAPE is really like mean absolute percent error is really low, but you miss the price spikes, then I don't care about your MAPE. Instead of to look at the full system simulation, like, hey, I have a new forecast, for example, I'm going to test it with the optimization and the control. Run it and compare the revenue and compare that to the old forecast, as opposed to just looking at the forecast in isolation and saying, oh, this one has a lower MAPE so it's better. And so, yeah, so like a lot of the work that I've done in my career has been building tools and systems are like AI platforms basically to monitor the AI platforms. Like the police policing the AI by running a bunch of simulations and trying to figure out like what is the best set of parameters and different conditions and how can we learn this ahead of time. And then to our algorithms, or as you get more advanced automatically to your algorithm that can actually modify your algorithm. Okay. So that's what I have on grid batteries. I'd like to move on to VPP operations, because I think it's equally interesting. How are we doing on time? Let me just take a look real quick. Okay. I think we're good. 15 more minutes. So, so there's a similar concept here for BPPs. You know, you can take a virtual power plant, a lot of market structures and bid them into a markets, much like we were talking about participating in the wholesale market. This does require the right regulations, it does require the right market structures, but there are a lot of places where you can do this and in particular in Australia. Another reason why the Australian market is awesome. And so, in Australia, you can actually provide grid services and sorry services with aggregate aggregations of batteries. The idea is you take a bunch of small batteries and then turn them into a big battery, about one megawatt or more. Depends on, I guess, the market, but typically around a megawatt for minimum bid. And here's a basic virtual power plant setup, just to kind of give you an idea of what it looks like. I've designed a few of these in my career and I'm not sure that all of them are fit the structure, but this gives you like a general sense. So basically, you have a bunch of stuff happening in the cloud. You have a data collector that pulls data from a third party, maybe like, I don't know whether data, market data or other data that you might use. You're collecting data from sites. So you have maybe like thousands of sites or tens of sites, whatever. So you're kind of collecting data there. You have, again, these kind of three pillars of, you know, market participation like we were talking about before forecasting optimization and control. And then you have some kind of bidding engine, which then takes all this takes your plan takes your control and says, All right, what do we bid into the market? And then the control element is going to try to push the sites to be able to comply with whatever the award is. So if you bid for some capacity, you got to make sure that you actually deliver on it. So that's where the control comes in. And so this loop can run. Yeah, like every five minutes or every hour or whatever. But one thing to note is like obviously there, there's a lot more complexity because of this right here. So getting data from all of the sites and making sure that it's very reliable can be quite tricky. And same with this right here, just making sure that you can actually control and send the right targets and instructions down to sites in an efficient way. And so that's really where the rubber hits the road in terms of the technical design. And so there's a ton of software engineering and infrastructure work to get this working properly. Conceptually, it seems easy, but these two pieces right here are really, really hard. Okay. So why is it a little bit easier than good batteries, there are some things that are actually good with bbps that you don't have to deal with what the complexities that are a bit easier than good batteries, good skill batteries. One is that they're often finance the projects are often financed by multiple parties. So that means that like, I don't know, there might be a customer that purchased the power wall and you're just leasing it from them. And so in that case, the capital cost is partially covered by the customer or a homeowner, as well as the governments like so there could be some programs that encourage market participation by, you know, behind the meter assets and typically you can get more financing for these types of projects. And that's complex value stacking for market offer market products, because typically you're only doing a few products as opposed to multiple products at the same time, you know, instead of bidding 18 products maybe you're getting bidding to. So that makes the, the algorithms a little bit more complex, a little bit easier. But what's harder is again the telemetry issues getting reliable information being able to control the sites. And the fact, and this is maybe like one fundamental issue that is still something being walked out in the industry is often you have two agents. Right, you have the homeowner who wants you to do something with the battery. I don't know provide backup power. Maybe be 100% off grid maybe they have different objectives that they want to achieve. But then you have the aggregator, which wants to go and make money in the market. So how do you share the benefits, how do you control the battery in a way that is quote unquote fair across parties and make sure that everybody remembers what you agreed to. Which was actually a big, which can be a big challenge and so customer visibility and education is really important so what are you doing with my battery. I think this is a question that comes up a lot in VPPs because behaviors complicated and it's unclear exactly why you're discharging right now versus later. Even as we discussed before, looking through in a lecture. It's a complicated problem so explaining that to customers is really hard and tariff design is really important and incentive structures to get them on board and to make them excited about this. So, so those are the challenges. Here are a couple projects that I'd encourage you to kind of look up and read about so one of them is a virtual power plant South Australia that's really first of a kind. It's doing a bunch of ancillary services as well as doing energy arbitrage and wholesale markets while providing tariff optimization for customers so it's really cool project. So there's a VPP in Vermont. That's doing a bunch of cool stuff to including the first VPP ever to be providing regulation services. So following an agency across thousands of sites so every four seconds new signal, thousands of sites responding in sync. Very, very hard software engineering problem but that's actively happening right now, Vermont. Okay, the time check, I have 10 minutes left. So I will go through a quick introduction to the company I recently founded and why I think it's really exciting and follows directly from from the talk on VPP. So, first of all, a lot of people have cars. Not a lot of people have batteries. So 85% of families in America have cars and soon they will be electric. And that's just a trend is for the longest time is that transportation, you know, for cars about 5% of the time your cars being used as transportation, most of the time sitting in a driveway it's doing nothing. But recently there have been multiple companies that have popped up to try to better monetize that asset. And so whether you're renting your car through toro get around. You have a ride share, your right share drivers Uber lift. So people are starting to use their cars more than just for personal transportation, they're starting to use it as an asset. And EVs are by definition grid assets, because they have three fundamental characteristics that make them very useful. One of them is that they're bidirectional energy resource. It's literally the same battery. In most cases, that are being put on the grid. There's no difference in chemistry. There is obviously some hardware components that have to shift power electronics, but fundamentally there's no reason why that battery can't provide the same kind of services. They are software defined. And so anybody who drives a Tesla knows or any of the knows that that experience is a lot different than driving a conventional car because pretty much everything is software defined. You can connect to the cars easily you control them, you can get good telemetry from them, much like a residential battery. And they're mobile. So this is one thing that's that's actually pretty fundamental that people are talking about but not really implementing the fact that cars are moving they are moving electrons from one place to another all the time. There is a network of transportation that exists and then there's a network of electrons. Is there a way to connect those two together to make better decisions on where to charge how to charge. And fundamentally that's that's what we're doing at power line. And it's really to address this big problem the fact that EVs are growing especially in California but elsewhere too. But the grid is not really catching up. We need about an order of magnitude more energy to address the coming wave of electric vehicles. And it's equivalent to just a massive. I can't even say this number I think it's 50 quadrillion dollars by estimate on how much we have to spend in order to actually get enough energy to supply these cars. It's just massive. And so that's what we're doing a power line we are connecting the transportation. Sorry, the transportation network with the electricity grid via AI platforms in the cloud does connecting what's happening on the grid. And what's happening with cars who are mobile apps. So that's how we talk to different EVs and provide them incentives on where to charge how to charge, whether that's at your house at the office or on public infrastructure. And so conventional solutions today, kind of focused on this managed charging element, which is like your plugged in at home. Okay, let's treat your car like a like a smart, like a smart meter started off during certain times of the day. That's effectively what we're doing today. But there's an enormous amount of complexity happening on the grid and opportunities that pop up along the way. And so it could be the case that, you know, charging at different locations at different times could actually be better for the grid. And as a result, you could have be paid to provide different grid services. So that isn't really being done today is coordinating, you know, DRs whether that's at your house at your office or elsewhere with what your charging decisions are, they're typically treated in isolation. And so having some kind of integrated solution is required to really modernize the grid and address this coming wave of electrification. So that's what we're doing a power line. And so this is, I'll, I think I'll just leave with this point here, but this is, this is a great example of where the problems are. So like right now this is my house and like down in San Jose, this purple circle, and this red circle is where I dropped my son off at preschool. So these two nodes often actually have difference in prices and in this case, I just pulled a snapshot of prices here were about minus $150 per megawatt hour, whereas prices up in the red circle where I dropped my son off at school were about $150. So there's a huge spread but $300 per megawatt hour spread and prices. So just being able to know that and making a decision about do I charge right now at home or should I just wait, drop my kid off and then charge there. If I make that decision the delta is above $55 that I could earn, as opposed to paying $10. That's just one example at the wholesale level, but once you start thinking about distribution level inputs factors, and the fact that, yeah, like right now we are really not using cars. And the main characteristic of cars, which is the fact that they're mobile. It just unlocks a huge potential for, for how you can optimize the grid with your vehicle. And so we actually have a pilot project that we're launching now, or I guess in a couple days, starting on Monday, yeah. Where we are passing incentives to drivers on where to charge. And so we're pulling in a bunch of charging data, EV location, EV telematics so we have collecting a bunch of data from the car. Market prices emissions and grid conditions. And we are collecting that data running it through our algorithms generating dollar incentives where we will pay people to charge at location a or location be at different times. And that is in sync with what's happening on the grid, similar to the example I just provided. And so, yeah, that's that's our that's our company and we're we're really excited for this pilot in particular because there are two proof points one of them is technical and to show that we can do this very complex integration across multiple systems, but also to try to see what kind of incentives to people respond to how can we make how can we make people make better decisions. And this is a little bit information about how we can connect so participate in our pilot. So if you do have any be and you'd like to try it out. You pay to charge your EV. We're also doing paid interviews for those that actually tried out so up to $50 for 30 minutes of your time. Other thing is to join our team we have eight people so far we are busy backs Mason Sunnyvale. So for those that are looking for for a job soon definitely reach out collaboration opportunities, whether that's with Stanford or students or whatever. And particularly what we're looking for right now is B2G demonstrations. And so how we can actually get bidirectional chargers installed and start playing around with them so we have a few projects and flux to develop there. And then yeah, just follow us so companies called drive power lines or drive power line calm. This is my link or link to our page on LinkedIn and then a link to my profile. Oh, let's definitely connect on LinkedIn as well. Yeah, this is just, and I also have some more information about the pilots, if you want to come up take a picture of this. Great. So I think I want a little bit over, but thank you for your attention. The question is regarding the VPPS, they've managing DRs like, how do you actually manage, but the transmission constraints in like a non Nordic market like Australia like how do you aggregate the VPPS and coordinate. How do you aggregate the VPPS. Well, yeah, it gives us more of a, like a market design question, because so in Australia they have decided that. Well, depending on if there's sufficient capacity to install a battery and to install solar at your house, then they the regulators have decided that it's sufficient capacity to allow or exporting from those homes to the grid. In Australia, it's actually a regional market. So it's not normal. That means that pretty much everyone in South Australia, where you are, it's the same energy price and the wholesale energy price. And so, yeah, there's a regulator's decision to allow, you know, assets to be able to basically discharge into the grid without having to have additional capacity installed or, let's say, the price that they're that they're working. But yeah, like, it's a good point. You need to have the infrastructure at the distribution level to be able to actually export to the transmission. I'm just going to talk super interesting. I understand that for where you ask drivers to locations to charge their UV, do you also discharge the batteries. Yeah, yeah, so that's our vision. And so what we want to do is and they unlock this. But today, you do need to have a bidirectional charger in order to discharge your EV to the grid. So there has to be bidirectional charges and they're just kind of few far between right now in five years that'll change dramatically. So we are setting ourselves up to do exactly that. But even today just having better charging decisions has a huge impact. And so that's what we're focused on in this pilot. What's the driving distance when you're choosing a better location. Yeah, so we are so we're collecting data and collect preferences from drivers. And so if a driver wants to just, you know, is willing to. The short answer is we try to minimize the distance that drivers have to travel. And so we're looking for locations that are convenient and we have like a large penalty for driving too far. But there are different preferences that you can set as a driver and our intention at least, and it's a pilot so we'll see, but our intention at least is not to ask people to drive very far to provide the services. It was our follow up or. I'm just going to have a district part, because it's hard to predict people's background driving, right? What if you discharge a battery and they decided to drive all of a sudden? That's true. That's true. And that's, that's yeah, that's part of our algorithms to predict behavior and to propose offers that actually makes sense to drivers. So, you know, this is not like a one offer fits all kind of situation. It's like, drivers can elect into doing this if they want to. So if they decide that, hey, I'll take 30 bucks to just park for an hour, no problem. And they're willing to do that. And they can, they can kind of schedule around that. But we wouldn't like ask them or force them to kind of do anything that wouldn't fit their schedule. Any more questions. I have another question. Yeah, it's also about the grid scheme that we and like how it can provide services in multiple markets. So, like the algorithm, like how exactly does it decide, okay, this part can be used for like ancillary services and this part is used for like an energy market. Oh, I see. Yeah. Well, so value stacking is actually really easy with the battery for the most part, as long as you are have sufficient capacity to ramp up to provide an additional service, then you can provide both and you can stack. So for example, suppose you have 100 megawatt battery, 50 megawatts is energy. And you have another 50 megawatts, you could do 25 megawatts of reg up. So you could basically follow a signal to go in your baseline starts at 50. So you kind of move up from there. And then if you have another, let's say, so you do 25 megawatts right up, you do another 25 megawatts of contingency capacity, as long as you're able to increase your capacity for generation. During times that you're needed. Does that make sense. And so you can stack up to your full power capacity in a lot of markets. And also, if you are like bidding in the capacity market, does it also like certain megawatt. So can you use the same capacity and like the energy fit is there in the energy market. Yeah, typically not so that's where you have to have that mutual exclusivity so you can partition the battery into chunks for different services but you can't. You often cannot that's not always the case they're depending on the market structure but you often cannot like stack the same capacity to provide different services. One exception there is in Australia for contingency F gas products there are three different products that you can provide. But those products depend on how fast you can provide that service if there's an event. And because batteries are so fast and they can either do it slow or fast, you can stack all three for the same capacity. Any other questions from I I guess we'll go through the list then so Jeremy Pollack for community choice aggregators or municipal utilities with little technical skills and programming, AI, etc. What would be the easiest and most effective way for them to implement a VPP managing their customers DRs. Work with companies like power line. I mean, I think, in particular, in my experience and you know, this is one opinion, but I do think that defaulting to expertise and AI. It can be really valuable for utilities and retailers and to not try to because I've seen what happens when people try to do things internally at large institutions that are not necessarily, you know, technical in those in those ways so you know, you can go to aggregators and software providers and find ways to partner with them because you know, there are definitely ways to collaborate and I think, you know, there's no way to do everything together. And there's huge value that utilities and aggregators bring to the table that I think, you know, AI platform providers don't have and so think collaborating is really the way to go. Okay. So, wrong wrong isn't I'm going to butcher this name so I really apologize wrong zing. Ying says what kind of control framework currently deployed in power line centralized or distributed or sort of hybrid happening to nerd out about that topic after an NDA. I think, yeah, like a hybrid is probably closest to it. More of a hybrid. I'm sorry, Adrian a Rima. Sorry, there was a follow up. How do you consider the balance. How do you consider the balance out of the revenue. The revenue between the battery owner and aggregator. That's really hard. It's really about the market design or sorry the tariff design. And so that was one of the bullets that we didn't dwell on too much but it's probably the one of the most challenging things about a VPP is coming up with the right incentive structure. And that's why I think we're in a pretty good position of power line to do this. But that's really challenging. So it's more of a business decision but it's also technical because you want to make sure that like, you can provide good visibility and the kind of incentives that you're providing. The consumer knows what you're doing with their battery and can actually make, you know, informed decisions about yes this is a good, you know, you're doing, you're using my battery in the way that we agreed to or not. And so it's, it's a bit of a trick tricky situation. The tariff design and the incentive structure. Yeah, but it's quite complex. And I think there maybe was another one. But Adrian says doesn't discharging a battery from a Tesla car void the battery warranty. Yeah, so that there are some interesting. Considerations to have about ownership models for cars. And so that's actually one thing that that we're pursuing as well. So we're working with a leasing company called Zevi. They do leasing like cheap leasing car for a month or a month or whatever. And so I think ownership of the car and being able to manage that risk is going to be really important. And what I think definitively is that batteries are way over oversized and over designed compared to their warranties. And so a Tesla battery today probably closer to like a million miles than it is to 100,000 in terms of longevity. And so there's just huge potential to just use that. And yeah, whether there are warranties in place or not, I think that ownership structure and contracts are really important to get to manage that risk. But that's definitely something that we're actively working on terms of on the business side. Okay, there's another question I'm going to say, I'm going to skip to Eduardo. Very telegre the old part, if I got that right. What can the public do to encourage regulators and other entities in California to facilitate V2X and VPP in general. Yeah, I think I think first of all, we need to get the PUC on board to make it easier to install infrastructure at different locations. And so one thing right now is I have solar and batteries at my house to get that project installed to install to top grade my meter to service upgrade from 100 amps to 200 amps it took 18 months. And so it's a simple upgrade it should happen pretty quickly but it took 18 months. And I think I'm not trying to blame PG&E they have a lot on their plate and there's a lot going on but you know at the same time like it seems unacceptable to me if we're trying to radically change the grid. And so I think the PUC does have to get involved and try to put a little more pressure on on on the parties right now to just install infrastructure in a faster way. I think that's the biggest bottom back and where regulators can help. Anything else I think those are all the questions. You can. Sorry, there's one last question and I think it's an interesting one. Do you consider the battery state of health into the optimization? Or how do you consider or notify the health status of the battery along with those markets? The answer is yes. And that's a very complicated co-optimization that has to happen that was actually one of the novel aspects of the company that I founded called Wattler. Where we were actively, you know, leading telemetry and trying to figure out what is the state of the health of the battery and when is one megawatt hour or one cycle, let's say, you know, too costly. And so what kind of what is the threshold for going and using that cycle to do something useful in the grid. And so yeah, it's absolutely something that has to be considered. And typically it's for good AI platforms. Thank you very much.