 Welcome to the MACB7. Today we look at the Chauvin power. We're going to be talking about reading and charging. I just want to remind everyone that the last presentation for this call is in two weeks. I'm going to chat some from Matt, who talks about the grid that he saw. Our speaker, Chauvin, is a graduating PhD student in mechanical engineering. She's actually graduating in a month and she'll be doing postdocs at the PH, so congratulations. Thank you. And our vice-chair is Professor Ramana Chauvin and she will talk about electric vehicle charging today. Thank you for the introduction and also for inviting me. I'm really excited to be here. I've been to this seminar many times over the last six years, although usually in a different room, apparently, and it's really cool to have the chance to present and talk to you today a little bit about my research. So the topic of my talk will be long-term planning for EV charging, a large-scale modeling of control and driver behavior, but I'll start with a little bit background. So you're all very familiar with the transformation going on in the grid and in energy, but transportation is also undergoing a similar change and the two sectors together account for a large portion of direct emissions. This snapshot from 2019, you can see transportation and electricity together are more than 50% of direct emissions in the U.S. Many plans that look at getting to net zero, for example, this global pathway of our from the IEA, you can see those two wedges start out very big and very quickly disappear. And that's through policies like ban on the sale of internal combustion engine vehicles, targets for sales of electric vehicles on the grid side, renewable portfolio standards, bans on building new coal plants, different policies making this happen, and really what's unprecedented is that these two transformations are couples together. Electric vehicle charging couples, transportation and the grid, and this leads to many interesting opportunities and also challenges that we have to overcome. So the two impact each other in this coupling. It goes in both directions. On the left, I have this little sketch of the grid where I put some loads and generations, some storage, the network itself. On the right, I actually took this snapshot of the Marguerite system on campus because a lot of those buses are electric and we've done a couple of projects looking at their charging patterns. And so the grid impacts transportation in several ways. You have to have enough charging stations. Many researchers have found that if you don't have convenient, accessible, inexpensive charging options, people just won't switch to EVs. So that's important. It provides the electricity of the stations. Reliability becomes more and more important if you depend on a grid through your transportation needs. Lower electricity prices can actually help encourage adoption by making EVs more economically better than internal combustion engine vehicles sooner. Actually, this is happening a bit faster than expected with the increasing gas prices lately. And then the impact goes the other way too. So for example, charging when it's uncontrolled can cause early aging of equipment in the grid. We'll talk about that a little bit. It can force capacity upgrades if you imagine like a residential neighborhood, maybe 10 homes are supported by one transformer. If each home has a typical peak of two or three kilowatts and suddenly you're adding a bunch of EV chargers at seven kilowatts each, you can imagine how it quickly overloads what the system was built for and it can cause really expensive upgrades. They can also change the generation resource mix so depending on when these are charged, it could call for more gas picker plants or could be covered by a more solar. And then the benefit can also be benefit so they can provide grid services. We won't get to talk about all of these today, but improving this coupling was the focus of my dissertation. And we'll look at a few of these different aspects. Before we get into the details, we can talk about sort of what's at stake and what questions are we asking here. So I like to break the problem down into planning and operation. And in planning, we're talking to utilities like PG&E, policy makers like the California Energy Commission, grid operators in Palo Alto, and they're asking, you know, where do they need to put new stations? When do they need to install them? Do they need to upgrade their equipment? Do they have enough generation capacity? And the timelines for making these decisions are quite short. If we look at California in particular, state agencies are targeting 8 million personal passenger electric vehicles in the state by 2030. So that's a really big increase. But all of these decisions have a pretty long time right of them. So if you want to install like a bank of new fast chargers that takes several months to a year, at least if you want to build a new sub station that takes longer, building a new power plant takes several years at least. So we have to look at what's happening in 2030 and plan ahead so that we can make these decisions actually very soon. Operation is sort of the other side of the problem. It says once you've made these decisions and you have your set of equipment, people have adopted EVs, you have a certain grid configuration, then how should you operate and manage that day to day in the best way? So let's talk about both of these a little, but planning first. Quick question. Yes. Have the grid operators look at the battery technology? Yeah, they're looking at battery and everything. Instead of charging in two hours, you can finish charging in 10 minutes. Then the capability required for that 10 minutes would be huge. Yeah, I think fast charging is really interesting because from the sort of personal perspective, people think, oh, I'd rather just charge quickly, but I think it will actually be really expensive to put in that storage to really buffer such high charging loads. So it's definitely something that people are looking at. Well, it depends on the battery. If you can build batteries, yeah, well, it's that they might be really expensive. Yeah, so there's Yeah, exactly. And so I only really look at, I guess the charging side, but there's a whole area of research looking at the behavior and sort of what are people willing to pay to charge faster and sort of how do they value these different options from an economic point of view because you could make charging faster, but it's not good for the battery health. It's not good for the grid. It's expensive if you have to have batteries they to buffer it. So I haven't looked at that myself, but it's I know it's a big question sort of what is where does people's what's what's worth it, you know, to have extra convenience. Yeah, lots of really interesting questions in here. Yeah, thank you for the question. So before we get into, I guess, these planning scenarios, I want to break down electric vehicle charging behavior because I'll talk about that a lot. So one thing you need to remember is that mobility needs themselves are very heterogeneous. Everyone has a different pattern. If you think about yourself or your housemates or your parents at home or anyone, you can imagine they have a different how did they get here today? What did they do this morning already? This imaginary person I grew over here, you know, they left home, they dropped the kid off at school, they went to the office, picked them up, they ran an errand, they went home. Then what makes this problem really challenging is that on top of that, you have preferences around charging options. So this person might have waited until they got home and then plugged in once at the end of the day. Maybe they also topped up at work because it was free and they could maybe they didn't have those options and they waited and they used the cross charger on their way home. Maybe they didn't charge it all because most drivers don't charge every day. So you sort of layer these different mobility behaviors and then charging behaviors. And when you multiply that by eight million drivers that we're looking at for 2030, you'll end up with this aggregate low profile. You'll see a lot of figures that look like this in my talk. But I want you to remember that it's made up of all of these individual decisions and individual sort of charging sessions. The way you read this plot, you can say, for example, at 9 p.m., there's about two gigawatts in this scenario of home charging. And at 9 a.m., there's about two gigawatts of workplace charging. But really, you have to remember this comes from individual small charging sessions. And so that demand has an impact on the grid. We talked about it just a little bit about, you know, fast charging and buffering that impact. But even before you get to adding batteries, you can do a lot with charging itself because fortunately, it's a very flexible load. I mentioned most drivers don't charge every day. And a lot of sessions have idle time. So if you imagine someone arrives at work, they plug in at 9 a.m., they need to charge for two hours, but then they just sit there connected for the rest of the day. So that's an opportunity form of flexibility. So I guess the three main types of flexibility to break it down for you, we have within session shifting. This is the most traditional form of control. It's very inexpensive to implement because it's done through a timer, maybe in the driver's app or in the charging station. And that's exactly this case. You arrive at work at 8, maybe you don't start charging until 10. Or you arrive at home in the evening at dinner time, but you have the timer set until midnight because it's cheaper to charge later. Another type of control and flexibility is load modulation. So the same idea, you know, if this person arrived at 8, they needed to charge for a couple hours, but they don't have to charge sort of all or nothing. You can change how quickly they receive their energy throughout the day and reshape their demand in any shape we want so long as they receive the same amount of energy in the end. And the final type, these two types of flexibility are both done within a session. So once someone's already arrived and decided I'm going to charge at work today, but you can also think about changing that decision itself. Maybe it's less expensive to charge at work, maybe it's more convenient to charge at home. All of these different factors influence that decision. And several studies have looked at things like changing those prices or giving behavioral nudges where you know, you know, wait until tomorrow it will be better for the grid. These kinds of nudges to change that decision itself. And you can imagine that really changes when someone's charging and then there's total impact on the grid. The other, so we'll talk a lot about these two types of control today. And in our paper we're working on now, we look in much more detail at this third type, changing charging infrastructure. So if you have, you know, more workplace and public charging built out, more people have the option to charge during the day and some more load happens during the day. Yeah, very interesting type of control. And so how do we sort of use these types of control? If you were a policy maker, you'd have several different tools. Most commonly, it's setting rate schedules and electricity prices. So this rate is actually one from PG&E for residential EV charging. And you can see it's most expensive to charge at home between 2 and 9 p.m. with this great schedule and it's least expensive after 11 p.m. So in the charging data we have for PG&E area, we actually see about 30% of the drivers to income charging set timers. And so at 11 p.m. there's a big spike from everyone who just delayed their charging until that lower price period. Shaking charging infrastructure, using behavioral nudges, these are all tools for policy makers to use these types of flexibility. Yes. Do these people have automatic timers like built into the charging and they just, is it like just charge here or is it just a button pressed that says optimal charge? I think it varies by car, like manufacturer and by charging station, but it's very simple. You don't have to make the decision every day. It's sort of like if you have a home charger, you could put a setting on it that says, you know, never start before 11. Yeah. So automated. I think that that's one of the challenges that utilities design these rates as a way of shifting behavior and say, oh, everyone's going to wait into their laundry at night. But when you have these automated responses, it's like very precise. And so you get a very sharp spike at 11 p.m. instead of just sort of gradually shifting below. So it's not like a smart integrated thing where it's actually looking at the price? No, I think it's the same. It's normally that you just sit 11. Yeah, I think the driver gets to take a time. So in our data we see people have picked like 9, 10, 11, midnight, 1 a.m., sort of most people pick out of the hour, even though the only rates in this area are at 11 in the night. So the people who pick 9 or 10 p.m. are just thinking, no, waiters better. So it's not tied to the utility or officially tied to the pricing. Sure. So because we have like low shifting, and it is the automated response, do you think it can affect the market? So for example, so at 11 p.m., we see like because it's like many different types of charges getting power. So it means like the importance of pressure on demand, and hence like the price will change. There was a little bit of that. I think that that, well, you do see it started to see a big peak. Right now it's still pretty small. But the drivers aren't paying the market rate for electricity, right? These rate schedules are set years in advance. So for example, this rate, it goes through this long process with UCs, we got this rate put in place, and then it stays for a couple of years until the next rate comes through that process. So I think actually how people respond to this, it would change the market for, you know, the people who are selling electricity and the price, but they don't actually pay any difference because of this behavior. Does that make sense? Yeah. Okay, so this is the introduction. So we'll get into some of the details now, but I wanted to set this up. The sort of two foundational research questions that we try to address are what shapes charging demand, understanding driver behavior, and how people are making those decisions, and then how should we reshape it to improve grid impacts? We'll talk about that a little today in terms of, you know, how does control work at reshaping charging? And then the grid impact piece is I won't have time to go into that space, but we'll have a piece from the other person. And the last piece of motivation I wanted to include is that a lot of modeling in this space is proprietary. There are a lot of consulting companies that do this type of work in building these forecasts very heavily based on modeler assumptions and often very expensive to rent. So the other piece of motivation for this work, other than the problem, is also sort of methodological of building these open source data driven tools that we can then share and publish if we can use them. So this is the outline of the talk. We'll look at shaping charging demands for driver behavior, and then we'll also look at control. And starting with driver behavior, this is the paper we wrote about it if you want to see, you know, all the details of it, and I'm happy to talk more about it later too. So understanding driver behavior, we started by looking at charging data. We had access to a very large data set of charging data, and we noticed a lot of heterogeneity, a lot of people with different charging patterns, frequently they like to charge, what type of charging they like to do. We use this data set through collaboration with ChargePoints, that's all from the Bay Area. Most of it in Santa Clara County, which is where we are now, a lot of workplace charging, but still, you know, several hundred thousand sessions of public charging and residential charging. Focusing on 2019, it was about four million sessions from 38,000 drivers that were regularly using the charging in this network. And so studying this data, we observed a lot of different behaviors. We did observe some patterns, some people who always like to charge at home, or some people who have, you know, certain behavior showing up again and again, and so we decided to cluster the drivers. We used the broader clustering as a type of hierarchical clustering and described each driver with a feature vector that included sort of aggregate characteristics to describe their behavior. So their battery capacity, because we found that impacted their behavior. And then for workplace or home charging or public charging, you know, how frequently do they do that? When they do that, how much energy do they use? When do they start charging in each of those segments? All these different behaviors. So we create a feature vector for each driver and normalize and apply the clustering. And this centergram is an interesting way of visualizing the results. So the longer the clustering actually goes from the bottom up, but you can read this from the top down. So if you start at this top over here, you imagine all 38,000 drivers together. And then they split off into these two branches. We have this branch with several clusters that really only use workplace charging. And then we have a branch that breaks off over here with lots of drivers that use residential charging. And then drivers that mix public and workplace charging, some with large batteries, and that changes their behavior. Some sort of lower typical energy, different times of day that they like to charge, different mixing of these segments. We cluster the drivers into these 16 different groups. And then we wanted to use this insight about these different charging patterns to build a model of future charging events. So remember, this is current charging. These are 2019 drivers. When we looked at the literature of how people are doing this, a lot of models are travel survey based. So they'll take a travel survey, which has very detailed information about this person left at 8am, and they went to work, and it was this far, and then they got there, and then they left, you know, noon, and then this far. But that involves a lot of decision making for the model, or you have to model. Remember, we had mobility patterns, but then on top of that, preferences around charging. And so you have to model how these drivers would have made charging decisions if they had to use this challenging. It's good because it can be very detailed. And you have this data for a lot of cases, but it's also expensive. And so we wanted to use this analysis of charging data and bring it to long term planning. So we developed a model called speech, which stands for scalable probabilistic estimate to EV charging. You can look at it, the code is published, and you can read the paper about it and play with it yourself. I'll show a demo of how you can do that later. The key design features for this model were scalability. We wanted it to run quickly so that you can study different scenarios because, you know, you might change your mind and disagree and want to test assumptions and run sensitivities, and you don't want to wait days to find out the result of each change. It's open source. And then we built in these behavioral norms that I'll talk about. This is the graphical model. And so we have these driver groups that we identify through clustering. And then how it works is for each driver group. So conditioned on being in group one, you have a certain probability of charging and deciding to charge that home today. And then conditioned on being group one, and having decided to charge a home, we have a probability, a distribution of your likely start time and energies for that session and duration of that session. And so with the distribution over the different driver groups, you can sort of use this to build up an aggregate load profile. We can take another sort of closer look at the different driver groups. So these are the same 16 clusters we saw before. And this is just a normalized load profile for drivers in each group. So you can see we have those three clusters that were on the left that are people who use workplace charging, almost exclusively. But they have different sort of arrival time patterns, you know, cluster three uses much more energy than cluster one overall, if you just sort of imagine integral of this profile. Then we have some groups that use a lot of home charging. You can actually see here, each of them has these this timer behavior and the spike at 11 p.m. showing up. Then we have a bunch of groups with public charging and multi unit dwelling charging and workplace charging. I think some of the really interesting things when we did this analysis first was the mixed use of different charging segments. So a lot of models will assume, you know, this is a home charging person or this person prefers to charge at work. And we found a lot of drivers mix and use many different options. We saw different behaviors like topping up, you know, people who never go below 80 and they're always charging every other day and charging small amounts of energy will retain a high state of charge, whereas other drivers would go several days and then have a big charging session. So sort of tied to risk aversion. The timers, it was very interesting. And I want to draw your attention to these values here. So this is the weights in the original data set. In the original data, 16% of the drivers were in cluster one and 2% were in cluster four, for example. But changing those weights is how we can build new scenarios. And so we'll look at how to do that. This is the original distribution over those driver groups. So I mentioned you have 16% in cluster one and 2% in cluster four, but changing that distribution is a really valuable tool to create new scenarios. Yes, the other piece of the model first is this charging session model. I mentioned given, you know, given being in group one and given that you're charging at home, what's your likely session parameters. And so we modeled that using another type of question called Gaussian mixture models, the joint distribution of energy, start time and duration. And what you could do with that is sort of understand breaking down this aggregate, all of the drivers in this particular group when they charge it work, this is their charging profile. But you break that down into the different mixture components and you can interpret them as different behaviors. So yeah, so 20% of the time, you know, 20% of the sessions here came from this first mixture. And so you can say 20% of the time when a driver in this group was charging at work, their session looks like this and they're charging, you know, in the afternoon, or 15% of the time, their session looks like this and they're charging in the morning, sort of interpret these mixture components as behaviors as well, breaking down the data further. This is the base load profile. So you can see there is a lot of workplace charging in the data set. And you can see a little bit of home charging, you see that 11 p.m. spike, but a lot of workplace charging. And so that's not really what we expect even for today what's happening, but also for 2030. And so to start building these different scenarios, we started by changing that distribution so that there's more home charging to match sort of what we expect to happen today. So here's a base case. And here we said, okay, let's reweight those different driver groups so that two thirds of people are in groups that use home charging or have access to home charging, because that's more similar to what we expect in 2030. And this has actually the baseline we see 30% of the drivers who are charging at home using timers. We can change that. We have those timers are isolated as components of the mixture model. So working with the California Energy Commission in their modeling of 2030, they think that more people will use timers and sort of adopt this time of use behavior. And so they said we should increase that. And so here's what happens if two thirds of people who charge at home use timers, you can see the peak is much higher, just the simple change of more people following their time of use rate increases peak demand by a few gigawatts, which is really significant. But then we can play with this further. So let's look at, for example, those timers. If you look at the mixture model for this segment, you can see that the first mixture component was this, you know, typical uncontrolled home charging you arrive in your evening. And the second component was this timer behavior. It starts at 11 o'clock. And so if we change the distribution over those behaviors, we say, actually, this behavior is zero of the sessions following this behavior. And then we'll shift all those to this other one. We can create an uncontrolled version of the data that we started with, sort of identifying that behavior and taking it out. This is what the uncontrolled case would look like if you didn't have timers and didn't have time of use rate. We can play with workplace charging. So this was one of those groups. What if we said no one charges in the morning and more people charge in the afternoon and we're going to shift that and, you know, maybe workplaces are encouraging people to wait and plug in after lunch, something like that. And you can see how that changes the shape here. So we've reduced the peak in our workplace profile and sort of increased the afternoon amount of workplace charging. We can also play with the driver group distribution. So we have this base case we started with, with a lot of home charging. You could increase that or increase, in this case, multi-unit dwelling charging. So say, okay, today it's very hard for people in apartment buildings to have home charging installed because those challenges, if they're renting and who's going to pay for it, it's also very difficult to cite. It's a much more expensive project. But maybe in the future, that will be resolved. And so in 2030, we can look at a scenario where we have much more charging at multi-unit dwellings and you can see then what happens. In that case, that segment grows. The peak is a little bit higher in the evening. You could say actually, you know, home charging is not, it's going to be a thing of the past and we'll have more workplace charging and switch more drivers to those groups. Or you could say, everyone's going to have large battery capacity and switch everyone to the groups with large batteries. Basically, you can do anything you want when you're designing these scenarios for 2030. And what I think is really important is that I don't know how to make those decisions for each place, for each planner. So we published the code, published the model, and if you go and download it, you can actually locally run this tool. This is me playing around with it yesterday and you can go change those assumptions yourself. You can say, I think, you know, I'm going to change the number of drivers and run a weekend model or I'm going to, what did I do next? Take out the timers. You can see what happens. You can look at different scenarios. You know, some of the ones from the paper, like that high, multi-minute dwelling case, see what that does to the load profile. And so you can go and play with this. You can do custom cases where you actually go and assign, you know, different weights to the different driver groups. And so we've been working with utilities like PG&E to help integrate this into some of their planning and provide more low shapes and more behavior patterns for them to use. But really, they know best about what's likely to happen in their area. And that's why we wanted to publish the school and encourage everyone to use it. Has fast charging been factored into this? So we have fast charging. We have a pretty small amount of fast charging data because it's from ChargePoint. So if you're interested in fast charging, you can use, like, sort of increase the weight of the groups that use fast charging. And there's one in particular, I think I get to actually at the end of this playing around. Yeah, so this group in particular has a lot of fast charging. And so if you focus on that group, then you do start to see it in the profile. That's all this sort of specie behavior here. I know. So it is included. And I think also, if you ran this model on data that had more fast charging, you would be able to break that down further and see more sort of behaviors even within that group of people that use fast charging. But interesting when we actually find the distribution of start times is very similar to other public charging, the public slow charging. And so when you aggregate and you say, okay, here's my scenario with 8 million people, what if all these people using slow public charging use fast public charging, the actual aggregate low chip is very similar. So yes, it depends what you're studying. If you're studying the distribution level, it's very important to know whether it's fast or slow. But if you're studying sort of from a very large scale, like generation planning, then they kind of was the same. That's interesting. That's interesting. And they look the same because I mean, I was charging you almost like tripling the amount of problem, actually, maybe you'll find a lot more than tripling the amount of power that you draw for that charge, but it ends up all distributing the same. Yeah, people like the same time or like different because they might be kind of like strength that like a little company calls it with five hours to three hours. Yeah, it is each session looks quite different. And they're much shorter, like you said. But people arrive at the same sort of in the same distribution at fast or slow public charges. And so when you're at the scale of millions of people, you can't see that difference very much. Just yeah, it's kind of interesting. But you're right at like an individual level, it is very different. It's just because I mean, maybe in the future also, you'd have more fast charges that work places or other locations and you then get different patterns. But right now, fast charges, at least in the status that are mostly located at places where slow charges are like at grocery stores or places. And so it seems that they have the same arrival time distribution. And then you end up with the same agribusier. I think So that's the last slide about this model. One of the interesting future works about this model is applying it to other places, applying it to other data sets with more fast charging, for example. And so if this is something you're interested in, you know, I'm graduating, but there's a lot of work to be done. And I know there's a lot of interesting continuing this area. We are continuing this work and a lot of interesting interactions we're pushing toward. I hope you can mention a little bit at the end. Yeah, I will. And if any of you, especially your graduate students are interested in this area, just Yeah. Okay, how am I doing? It's pretty slow. Okay, well, so we'll have a little bit of time to talk about control as the other part of this modeling. And one of the types of control we talked about was that shifting within a session. And these timers are an example of that. So this is just in the charging data that we have from 2019, about 30% of people are using timers and that's what this looks like is the spike. But the other place that we looked at, we looked into doing control as a workplace. This is a really sort of growing area of interest. We worked in a project collaborating with Google to do some control at their workplace site, also at Slack, we installed a bunch of new charges into doing control. It's a really nice application because there's a natural aggregator. If you're thinking about people charging at home, it's very distributed and some people decide to set a timer or they do whatever they want. But at a workplace, you have this natural aggregator to manage it and sort of controls the site. And so they do this low modulation control and it's typically going to minimize the overall bill. One thing that this impacts at a workplace site, usually the charges are all supported by one transformer. And so we looked into how the charging impacts the lifetime of that transformer with different rate schedules and different charging patterns. We'll go through that really quickly. So, transformers have been studied a lot in the context of residential charging. You can see people arrive at home and there's a big evening peak that causes overloading very quickly. Several demonstration projects have confirmed this, but we really wanted to look at workplace charging and how this impacts the transformer at a workplace or at a commercial site. The aging of a transformer is driven by the highest temperature that's reached in its insulation. And so this is a model. This is kind of interesting on the right. If this is the total load of the site in yellow, you can see that the temperature on a different scale, but the temperature kind of lags that. And so it's really driven by that loading, but there is thermal inertia. And so we basically modeled a particular transformer at a Google site through that collaboration and said, okay, we have charging demand. What if we change that? How does that change the temperature? And then the lifetime is a function of the temperature and it's an exponential function. So after some point when you're going above, I guess, the rated temperature is designed for you very quickly to see the lifetime drop off. To formalize this a little bit, what we're talking about is the total demand at the site, so all the cars charging there out of together. We can call that D, demand. And then what you're doing is minimizing some function of that demand, subject to simple constraints like delivering the right amount of energy, only charging one of the vehicles are there. We don't look at vehicles with grids, so only charging between zero and, in this case, 6.6 kilowatts because it was all level two charging. And then this function can be anything you want, but in this case, it's most commonly done to minimize the electricity bill. And we also looked at what if you explicitly minimize the transformer's aging. This is an example profile. So in blue, this is the uncontrolled profile for the site. If they were 355 cars charging together. And you can see it's sort of this uncontrolled peak is highest around 10 a.m. And then it drops off as fewer people are charging mostly because most people arrive at this time and they don't have very long sessions, but also because fewer and fewer people arrive later. If you do the control, if you say, okay, this function is PG&E's E-19 rate schedule, which is a rate schedule for commercial sites that includes the demand charge, which is what penalizes the maximum rate reached at the site. Then the end of doing this load, like peak minimization, you see the green result is flat. So as flat as it can be, you subject to those constraints around people leaving and getting right amount of energy. If you do the transformer aging, as the objective function, it's quite similar, but it allows these kind of little peaks at the beginning and end. And that's because of that thermal inertia. So you can overload it more for a little bit, but just not long enough that it really heats up. And then it's whatever level you're sustained after a long time that drives the hottest temperature that it reaches. So very similar. And so we found a really interesting result. And so how to read this flat on the bottom, we have the number of cars. So we slowly increase the number of cars visiting the site every day. And on the y-axis, you have the lifetime of the transformer. So when you have, you know, 50 cars charging there every day, this is the 225 AVA transformer. Later at the maximum lifetime, there's no impact. And you add and add and add cars. And then here, this blue was uncontrolled charging. And so there you see quickly, after about 130 cars, suddenly you're really starting to impact the transformer and cause aging and the lifetime really drops off. These different controls are better and better and better. Here in orange is the case where we minimize directly for the transformer's lifetime. It does the best, which is good because, you know, this was the objective. And I'd like to note also that you can actually get much further. So you can have more than 67% more cars charging there before you have any impact on the transformer just from adding this control. But the really exciting result was these other series that are on top of it. So you have peak minimization and PG&E 19. And basically what this says is if you have a demand charge in your bill, you're encouraging this peak minimization, you're almost doing as well for the transformer as if you were explicitly making everyone apply for the transformer's health. So our conclusion from this work was that sites where the transformer is overloaded or at risk of being overloaded, it's good to have this charge and encourage, you know, when people are minimizing for their bill, which is what they're most likely to do, they'll also be protecting the equipment. This work sort of showed how powerful control can be. And the next step in this modeling was to try to include that in these large scale modeling tools. It turns out to be really challenging because it's very expensive. So this optimization on the scale of a few hundred cars is fine. But when you scale up to millions of cars, it doesn't scale very well at all. And you could take hours or days to run the full optimization problem. So a lot of other large scale models will do with different types of control and simplify the problem to represent it. Either doing like a fully centralized model, just modeling sort of choices instead of doing this type of load modulation and doing sort of theoretical approximations. And so in this study proposed a new method to learn this directly from data, how we represent this at scale. This was based on a couple of key insights. The first was that although there's lots of variation day by day for individual drivers, the aggregate load at a given site is pretty consistent. And once you do control it's still pretty consistent. And so we thought we could just learn this mapping directly, build a big data set of uncontrolled and controlled carers of profiles, and then forget the optimization and just replace it with machine learning, a very sort of stand for it. And so it worked pretty well. So we looked at several different rate structures, some with demand charges, some with just energy charges, some with both of them. I don't think I have time to talk too much about the different rates, but briefly the methodology looked at building training data sets, the building a big data set of controls and uncontrolled profiles, dividing it up, tested a bunch of different regression models. And we found that for each different rate schedule, it was either rich regression or random forest regression that performs the best. And the normalized RMSE was actually pretty good. And so when we look, for example, here are some profiles. This is uncontrolled charging, and then peak minimization control. And you can see there's like a small error between the model and the actual data, but it performs pretty well. And the test set normalized error is really 2.5% to 4.6%. So then we can use this and plug it into our tool. So this was an uncontrolled scenario I showed earlier. We ran this method and trained the model for peak minimization control. And then once you've trained it, it just becomes like another knob to play with when you're doing this scenario design. So you say, okay, this is my mixture of driver groups and this is the behaviors, oh, and I'm doing control. And you can just include that. And it takes less than a second to apply once you find the mapping. We also looked at when we created a time of use rate based on average grid emissions and learn the model of it, put it into the simulation, and this is the result. So you can very quickly then just include control as one tool here. And yeah, so before... Did that just shift the workplace? Yeah, this is just on workplace. So usually, like looking at rate schedule based control, different rates by the different segments. So you have like commercial rates like PG&E is E-19, and then you have residential rates like the one that causes the timer. And they're quite separate in practice. Just look at the far right one, kind of kill you in terms of the transformer. Yeah, it's really interesting. So this is... I guess I can tell you a secret because the paper's not published yet, but we've been studying this question. And it does increase the peak. So on days when you have a lot of solar, this is the type of control you want to do to align better with that solar. But at the local level, it does damage the transformer. And also if you have this sort of rate scheduled design control, it encourages you to do the same thing every day. And some days there's very little solar, and then you're just really increasing your peak at that time. So it's a really interesting research question. If you want to go into it, there's a lot involved in this rate design. I think it's an interesting conflict because, you know, it costs a lot of money to upgrade the transformer. But you're also at a large scale trying to align better with solar and transition to grid. And so there's some analysis to be done at the trade-off and sort of which cost is worth it. Maybe the answer is not the same across every site. And some places can afford to upgrade the transformer into better align with solar and other places that's better to protect the equipment. I don't know. But it's a very interesting research direction. Yeah. Shuman, is there a possibility of using like a cost associated with the transformer, and then like trying to minimize the cost and also the time of use? Yeah, you could definitely do that. So in this study, the transformer study, we did look at that. And we were going to have sort of this mixed objective where your objective has the cost of the transformer and the bill. But then we found that they aligned, and there was no trade-off because if you're minimizing for the bill in this case, you're also protecting the transformer. But if you look at a different objective, like sort of maybe grid missions and you're trying to align with solar, then you would have a conflict. And so you could do that type of optimization. I think it's really interesting as a direction because most rate design doesn't happen that way. But it would be great to know what is the best rate and how do these other rates compare against that benchmark? Or we don't have a huge investment in rates? Yeah. I think I have this noted in future work. Really quickly, some conclusions are that behavior is really interesting and heterogeneous and important to include in this type of modeling. Control is also really powerful and reshapes demand completely when it's applied. It has a really big impact on the profile. And then there's value to having these open source data-driven modeling tools, both to improve the models and not have to depend so heavily on assumptions, but also to make this accessible so other people can apply it to their own setting or change the assumptions. But there's a ton of future works. I mentioned this study of large scale grid impact is coming out soon. But then in terms of the modeling, there's a lot of interesting extensions. We looked at sort of a typical future day, but what about inter-day variability? This cuts both ways. I think it's sort of on Sundays, maybe everyone needs to charge us a long weekend. What does that mean? But on other days, maybe the grid is going to have a hard day tomorrow and you can ask people to wait an extra day because people don't need to charge all the time. There's lots of sort of flexibility that way. I think that's unexplored. Looking at other places I mentioned, future adopters modeling them differently. Looking at other segments. There's room to apply this type of model in medium and heavy duty. I know that's something that is going to be pursued here in the next couple of years. Looking at how this, we only look sort of at time, but there's also a spatial aspect I was talking about with Chinnu earlier. Like if you have these charging patterns and you're shifting people's decisions, what does that mean in terms of like a network of charging stations? Looking at V2G, I know, I probably mentioned there is a really big announcement about this just earlier this year of support for new research in V2G and V2X and V2Home. More of this sort of rate design question. What is the best rate? What does it mean to have this conflict between distribution grid and generation level impacts? Climate resilience of charging networks, I think also very unrelated to what I've done so far, but it should be tied in. I mean, I think that will be more and more important. Basically, there's a ton of future work and I'm graduating and really sad that I won't be able to do all of it while I'm here. And so if you have any interests or any questions or want to get started looking at any of these things, I'd be more than happy to talk about it. I think I should stop talking now because I only have three minutes left. There's even one more question. Yeah. Is there any reason why, and I'm guessing it's just because, answering one question, it's just because right now EVs have the most storage of anything, but is there any reason that it's EVs in particular versus all residential energy storage? Because what's the difference between a battery and a car and a Tesla's powerful? It's a great question. I think the value of just having EVs is that you already have them. Or you will already have them. And so if you could all sort of have them do double duty and also sort of for the modeling. Oh, for the modeling. Why did I only look at EVs? Because it's a lot of work. I think you could do a whole PhD on EVs plus home storage. But I think if you look more closely, so I focused a lot on workplace charging, but I think if you looked more at the home area as well, there's a lot of other research going on here about all the different loads in your home and how do we manage that? There's actually, I don't know if you talked about the lab. There's a lab here on the third floor that has some power walls and there's connections to EV charging stations and solar instead of looking at how to manage the home as a whole load. That would definitely be really cool to tie in here. And so that would just change. I guess it would change this. This is only showing the EV load, but also if you imagine this overlaying on like all other baseline demand, you're changing both of those. And that could be really interesting. And it's a lot of possibility to pay attention to the type of customer behavior. And it is easy to analyze. It's difficult to analyze the customer behavior. It's the same thing. The transition from the B2B to B2C, when you have the customer involved plus energy, there's a lot of interesting things to analyze. And the other thing that I guess I just observed from like that peak that you're showing, whether this be residential or with a transformer and just the way the grid is set up, if we were to be like on all solar, I feel like if we're maximizing these transformers and like residential energy storage or whatever energy storage, if we have all their energy, even if we store it all, are we like taxing our transformers that we're getting at all points or have them? Yeah, I think it's an interesting question. One thing I would add to sort of this thought experiment is what if you had that solar co-located with the chargers. So we didn't look at that. We looked at, you know, a site and you have the EVs and the transformer, but at the Stanford bus center, for example, the buses and a big solar array are behind the transformer. So then if you looked at it, it would change that problem a lot because if you're just doing EVs behind the transformer, well, I guess basically what we found is you want the load behind the transformer to be very flat. And if you just have EVs, you know, this is the best you can do. But if you also have solar, maybe this is the best you can do, because then the total load behind the transformer would be flatter if you want better with solar. Yeah, I think maybe that's the answer then. So like, save your distribution grid by putting solar wherever you have EVs, but that only works with daytime charging, right? If you put solar on a new home and you have home charging, they don't line up and they don't really help each other. I just say that because there's only a big debate on whether it should be like localized residential energy storage in a very small field or like huge governmental batteries sitting in places. Yeah, it's an interesting question. So in this few study, we found that these scenarios with high peak, I mentioned some days there's less solar, and so this is actually just increasing heat in that demand. It drives a higher need for grid storage. And if everyone, like if the grid storage and the charging infrastructure and everything were only operated by the same body, it would be really easy to evaluate sort of, is it worth it to do this type of control and how much does it cost to have that storage? But when the storage is distributed, and all of these decisions are distributed, I think the problems won't be complicated to you. That's not to say it's better or worse. From any other perspective, just it's really interesting to think about how sort of how many people are involved. And when we study this, you know, high level of birds-eye view 20, 38 million drivers impact, it's not very simple to think about, I guess, how you would change that and how you would implement different controls and is it worth it to have storage because it's all sorts of different people who are involved in each change. I don't have a good answer, but it's a very interesting thing to look at. Was this, you showed the control algorithm that you used, was that actually piloted at the facility? Yeah, so they do controls. Because I was just, I guess what I was curious about is, like these are like, like this? Yeah, was that, was that, was this like what he said later was also like piloted and data was collected as well from the facility? So some of it, so the project was affected by the pandemic. So it was part of a demonstration project. And so there is control happening at some of the sites that Google, and there is control happening at the sites that Slack, but we didn't actually get to test this like transformer control because of sort of the timeline of the testing ended up, we're going to start shortly after everything shut down and still no one's going to Google and they don't have any work they're starting to do anymore. So it changed the project a lot. But there is, so for example at Slack, we installed a temperature sensor inside one of the transformers there so that we could collect more data. I think, for example, this model of the temperature, it's very old. But it's really, really hard to get the parameters that you need to fill it in like on the nameplate of the transformer, you have some values and you make some assumptions and sort of do the best model you can. But one of the reasons for installing this sensor was to learn how bad the model is and learn how to calibrate it better. So I think that's also an interesting direction that they might continue like testing the transformer and using those temperature sensors to do some sort experimental validation of this. I think that Slack wouldn't let you overload the transformer to the point of breaking it. So this part might be always a simulation exercise, but at least using the real data to tune the model better and understand that, I think it's definitely a good direction. How would you also like keep out some noise from people who say charge for an hour and then unplug the car and do something else? Sorry, I will listen first. How do you like remove noise from like people who say presidents will plug in their car for an hour and then have to leave and confirm you take the system? But is that noise? Or is it just sort of part of how people are using less charging station? I think those are the types of sessions that aren't included when you have a very sort of model or based approach. But when you use the data, we do have a lot of sessions like that. And I think it's valuable to capture them and sort of see that this is sort of the messy truth of how people are charging at work. Yeah, how do you optimize for that? Because you're not going to ask like someone's like I'm going to find out how long I'm going to hear before putting your approximate hours. So how do you optimize for how long you feed down charge versus another place? Can you imagine, I expect my car to be charged to 20% and I show it to 10% because they are really expensive. Yeah, so that is another aspect of it, this sort of online control versus the control where you know everything in advance. We didn't look at that in the study. Actually, at Slack, one of the sites is with power flex chargers. And they do ask that. So when you arrive, you say I want to leave, you know, I need 30 miles of charging today and I'm leaving at 5pm. There's a lot of behavioral problems with that. Like people just always say the same amount and ask for more energy than they can receive and like don't have any idea about their state of charge. And it ends up being very flawed, but there are efforts to do that and collect that type of information. Yeah, I think we still have some way to go. I don't think people think about this very much, which is probably beneficial because then you can do control and do all this, you know, very beneficial, you know, take advantage of the flexibility for the grid and they don't care. They don't like even know or think about it. But if you then need to get information from the drivers about, you know, what's your state of charge or how much energy you need today, they don't know. I can't answer that either. So that's one of the challenges. Just have a button. It says most cheapest and most you know, or one that's more expensive. If you know anyone. If you have to work harder for the more expensive, less efficient one, people want it, which is always cheap. Yeah. I think the app also has like a history and so you can just like fail what I put last time. Probably shouldn't have. There's one question in the chat. Oh, I think we can probably wrap up for okay. Unless it's super quick. Oh yeah. So there's a question in the chat and then there's two. Oh, the one in the community. Does the M.L. approach generate individual optimized charging control sequence or the state of driven approach is used for the estimation? I'm not sure individual optimized charging control sequence. Um, it's only looking at the aggregate. So, for example, there are many other ML based approaches that look at individual acceptance and say, okay, this driver came at this time every day, and we learned from their habits and then we learned, you know, we can predict that tomorrow will arrive at nine. There's a different approach. This is just looking really want to understand the shape change from uncontrolled control in aggregate. So, when we train this, how we run it is in, you know, it was not really generated and control profile and we just apply the massing. And so what it generates is the corresponding control profile is an estimate. Yeah, nothing individual on the side. Great. Well, thank you.