 Okay, so today I'll talk about different experiments with the campus that we're seeing, that including this room we're seeing right now. And so that's one of the fun things about the research that I do is I guess, as you see, well, you don't see, but actually you do see when you have with sensor data, you see the results of the work that you do around you and I think that's pretty cool. So the Stanford campus is really a small city in the sense that we have, so it's sizable, it's not huge, but it's sizable. The daily population there is partly people coming in, actually that number of 30,000 is more based on electricity consumption of the campus as a whole and that's roughly equivalent to a 30,000 people city in California and as you all know, there are different usages, there are different reasons why people are on this campus, some people live on this campus, work on this campus, do research on this campus, teach on this campus, so there are many different things that we do here. One of the things, so this is a picture, I think, taken from the engineering quad looking towards the main quad, the way I, so this is more of a schematic for the way that I see this campus and why this is such a great place to do both real life and thought experiments and I'll talk both about real life and thought experiments today over the past five or so years at this point. So this campus, you've basically got from the point of view of the outside world, you could think of the Stanford campus almost as an island and we're buying gas in an aggregate way, we're buying electricity, so there's one, physically there's one power line going to the outside world, there are two transformers on campus, one of them that's for the central energy plant that I'll talk about quite a lot today and the other one is for the 200 or so buildings on campus, and so there's one master meter, if you like, there's one electricity bill for the entire campus, which means that there's also one, to simplify, there's one entity footing the bill, that's not exactly true, but there's the Stanford University and we've got this depending on how you want to see it, either a very dictatorial regime or a communist regime at Stanford where there's sort of the one community decides what's happening for all the buildings, so that makes it a good place to try experiments, to test new ideas. And so if you look at this campus, so if I start in the middle bottom of this diagram, buildings, so depending on how you count, 150 to maybe 200 buildings on this campus, and the buildings like the one we're in right now, they use electricity, they want to have heating and that's coming in in the form of hot water, and they want to have cold water, chill water for cooling. And the heating and the cooling are something I'm very interested in here, the heating and the cooling are produced on site at a place called the central energy facility, which is roughly that direction. And so that's the place that John was talking about just now. In that energy plant, you have different types of machines that use electricity, that consume electricity, it's almost only electric, there's a bit of gas that we use for heating in the winter, but mostly we use electricity to produce the heating and cooling. And so these are different types of heat pumps, so the basic idea of a heat pump is you put in electricity and you have some hot source coming in and a cold source coming in, and basically what you want to do is you're making the hot hotter and the cold colder by moving heat from one to the other with electricity. And so we have different types of heat pumps, there are two types of heat pumps, they're conventional electric chillers, industrial chillers basically, that produce just, where we just use the cold water and we reject the heat to the atmosphere, they're cooling towers, and then there are what are called heat recovery chillers where we use both the hot water and the cold water, and so that's a different type of heat pump, also what you would call industrial scale chiller, and the heat recovery was a big part of the original design. But so different options to produce heating and or cooling, and then we have some storage, so big tanks and if you've driven or walked or biked next to that building you'll have seen, the tanks that's most of the footprint of the building, there are tanks for hot water and chilled water, I'll show a picture of them in just a second, so there are cooling towers and there are decisions you can make about when to consume electricity for that plant to send heating and cooling to the buildings, the heating and the cooling is sent to the buildings through pipes. So what we've got on campus is what's called a district energy system, so we bought both a district heating and the district cooling system, and so we've got different systems of pipes going around campus to bring the heating and cooling to the buildings, I think for each system there's roughly 10 miles worth of pipes going around campus, and actually in 2016 when we upgraded the system a big part of the upgrade was ripping out the steam pipes, we had what's called a district steam system before and we moved to hot water, which is what's typically put in place now, and so that was a big part of the upgrade. And then so the campus buildings, then if you move to the left, and so I'll talk a bit about, so actually the part on the left I'll talk more about, I won't talk so much about electric vehicle charging, but these are other assets, energy assets, if you will, that are sitting under this umbrella that are active, we've got solar on campus, there's something like, I think it's three megawatts worth of rooftop solar on the roofs, and then the building ventilation systems, so these are big fans in the buildings that blow around air, those actually consume quite a bit of electricity as well, and those actually we're looking at in this work. So I'll talk in sort of about two pieces of work that are separate, but they're really one continuation. The first part of work was actually mostly done during my PhD and kind of takes the supply view of the campus energy system, and is looking at planning and scheduling things at the central energy facility, assuming that everything that happens on campus is one consumer, one demand, one black box for demand that we has to be met. So I'll talk about that first, and I'll try to be done relatively quickly at that part is mostly that's introduction to the second part, which I currently find more exciting because that's what I'm working on right now, which is more at the demand side of the system, which is taking the point of view of the buildings and seeing what can we do inside the buildings to provide more degrees of freedom for scheduling the central energy plant. So the base decision making problem that you're dealing with at the central energy plant is about energy operations, is how much electricity do I send to the different machines every hour? Roughly the time scale that I care about here is 20 minutes to an hour. And why do I say 20 minutes? Because that's when you talk to the people who manage that building roughly when they make a change in the machines that they're working with, that's the time for the buildings that sort of at the other end of the loop to start feeling of the change. So that's roughly the time scale on which you're making decisions, 20 minutes to an hour. And what they're trying to minimize is their electric bill for the aggregate campus, which is the sum of what's going off both those transformers. So the heating and cooling portion of the electricity loads and then the sum of all the other buildings. And so you're minimizing your energy costs and I'll talk in just a sec about how we pay for electricity. And then, so your main decision variables are how much you heat water and cool water with each of your different machines. And you also have the opportunity to store water. And then you have constraints on how you operate the different machines. You have one constraint, which is you want to give people their heating and their cooling, otherwise people are not happy. So you have a demand constraint there. The way we, so this is a bit important to understand what's coming next. The way we pay for electricity is in what's called a two-part tariff where basically we pay a variable price for energy hour by hour. And that, so Stanford's what's called a direct access customer on the California market. So we get a price there and we have a scheduler that helps us place our orders for energy on the California market. And so we pay a price that, so one component of what we pay is variable hour by hour of the day and the price changes. And I think I'll show what that price looks like in a couple of slides. And then we also pay what's called a demand charge. And the demand charge is essentially looking at what's the maximum power draw over the course of the month and then you pay a price for that. And so one analogy for that, so actually not so long ago because I remember I used to pay my phone like that before. So and I'm not so very old, so it's not too old. There was, when you paid for your phone, you would pay both for speed, so what speed at which you were getting your internet and then you would pay for the number of minutes that you got through. So you pay both for volume. Yeah, I guess that's another analogy is water. That's one I like is volume and rate. You're paying both for the size of the pipe and how much went through the pipe over the course of the month. Okay, and so mostly you're looking at scheduling the building that's in the box on the top right there to meet the demand from the buildings that are through the chill water loop. So one big source of flexibility that I kind of hinted at is, and a lot of that work was looking at how much flexibility was there in this thermal storage. So this is an air review of the central energy facility and so we have two big tanks for cooling, one for heating. One of the things we were looking at in this work was trying to say how, is there an equivalence between this kind of storage and electrochemical storage, which was, I mean, what it still is, a lot of people are very excited about electrochemical storage and there are a lot of good things about batteries, but there are also things that are not so good, like they cost a lot, whereas this is pretty cheap. So what are things that are similar and not so similar between electrochemical batteries and this thermal storage? What are the services that you can provide with the thermal storage? That was a big part of this work. So the first thing that we did was to write an optimization program to say I'm going to make optimal scheduling decisions at the central energy plant and that's actually a replica of the system that's running the central energy plant right now. So there's an optimization program being solved. I think they update every 10 minutes, actually, the one that they solve in their real version. So we built a replica for that to say if you're paying this two-part tariff, how would you want to make your decisions? And the output from the optimization program is what you're looking at here, where the green here is the electricity that's being consumed by the central energy plant. So that's what we're paying in electricity to get the heating and the cooling. And then the orange line there, that's the total campus load. So that's how much the campus is doing and the blue is the total. So that's the master meter. That's what PG&E sees at the end. And the dashed line there, that's the price that we, the hourly price we pay for electricity. And there's also a fixed component, which is the demand charge. And I think the left, yeah, the left there is a summer period and the right is a winter period. And so what this, the main message here is that the blue is trying to be as flat as possible. So, and that's mostly given the prices that you're paying for the demand charge and for the variable energy price. There is, you kind of do want to try to shift consumption if you can to when the price is more expensive and you see a bit of that happening here. But really, the game is, look at when the buildings are consuming power and try to schedule the heating and cooling when the builders are not consuming power so that the aggregate is as flat as possible and that you don't pay so much of a demand charge. That's mostly what this is trying to do there. And actually, I think this is one of the big, or at least now not so much, but at the time one of the big revenue streams that people were trying to get in all the company, the startups that were doing electrochemical storage and trying to provide services, a big part was this and I think they called it demand charge management was this. So you can kind of do the same thing and that's what the campus is doing right now with this thermal storage. Then a second question that we asked, so this was not so much just about what's the cheapest thing to do but also if we start to say we want to make decisions about carbon and so here we had a proxy for if Samford were to pay, were to self-impose a carbon price. If we start to price in carbon in our decisions and the plots on the right there are showing what are called. So I won't go into that too much. I'll talk actually a bit more about that on Thursday. But average emissions factors and marginal emissions factors, but so basically I guess the main message from that plot is already today and increasingly the future, we can expect that electricity will be cleaner during the day and California than it is at night. So one of the things we'd like to do is to shift electricity consumption towards the middle of the day. And what the plot on the left is saying is by having a fully electric system, there's sort of two benefits to Samford if you're looking at Samford's operating emissions for heating and cooling. One is, well as the grid gets cleaner, well sort of automatically the carbon footprint of Samford is going to go down because the electricity we're getting is just cleaner and that's the cleaner grid here. And then the second is, and this will be especially true if we start moving towards a world where there's a big difference between the day and the night, well then it's going to have more value to do the sort of shifting of throughout the day and that's something that you can do actually really well with this thermal storage. And so just to re-emphasize that point quickly, so here I'm showing you, so what I'm calling grid imports are basically everything that Samford is consuming from the outside. So this is the total electricity that Samford is getting and so on these heat maps, every column here is a day of the year and every row is an hour of the day. And so this is sort of showing you a full schedule for a year of what Samford would be taking in this what if scenario we're paying for carbon and the left there is, so the left BAU that's business as usual, what you would do if you sort of had perfect foresight, everything that was going to happen for a full year and you're just paying for the regular price for electricity and most of that is sort of a is a color that's as close to as possible to the same thing because it's trying to be flat and you can kind of see the different months or different shades and then in the afternoon that's when typically the price was more, so this was back in 2018 actually, don't know what it looks like now, but there are sort of two peaks during the day for the electricity price, so you're trying to lower consumption there to pay a bit less and then the right side, that would be okay, I just care about carbon and I have this proxy signal for what's the carbon intensity of the grid and I'm just trying to minimize my footprint based on that. Well, and then basically, so if you're like this is also, one way I was looking at these was as a way of saying, if you wanted to say, Samford wants to absorb as much solar power as it could, what's the size of a solar plant on campus or we could just absorb all of that solar power in the middle of the day, well that's roughly what the plot on the right is telling you. So a second question, so that was mostly a thought experiment because we, Samford doesn't self impose a carbon price. So a second type of experiment that we did that was very much in the real world was saying, can you provide some flexibility to PG&E or to the California market? Can you provide some services to the grid? This is another thing that people want to do with batteries. And so we took a real world program, which is PG&E's, so PG&E is Pacific Gas and Electric, their capacity bidding program. And so I'll use the acronym CBP. And basically the game there is every month you have to tell, so it's actually 25 days before the start of the month, you have to tell PG&E, this is how much capacity I can give you and that capacity is measured in kilowatts or in megawatts, so it's power. And then as you go through, so you make that promise to PG&E, I can be available for five megawatts this month. And then as you go through the month, basically what PG&E can do is they can tell you up to, I think it was 2 p.m. every day, tomorrow I'd like you to give me that flexibility for a certain number of hours and their constraints to how many hours they can call or how many times they can call for the month, how many hours they can call each day, things like that. But mostly the idea is you say the power and then they tell you tomorrow, I want you to give you that for two hours or three hours or four hours and so forth. And so there are two problems that we were looking at to participate in this program. The first one is more of a planning problem and that's actually the work that we did before the summer to try to convince the energy operators on campus that it was worth participating in this program. And the planning problem is more about offline decisions is before the month even starts, how much capacity can I provide? And so we did a bit of work on that and we used an approach that's called two-stage stochastic programming where the basic idea there is in the first stage, I make decisions about what's happening in the second stage, which is uncertain. So I have different options for what can happen next, what's the decision I make now so that on average or in the worst case, I make a, that decision is good given the possible futures. And then there was an operational problem and this one we had to do well because during the experiments there was a screen in the central energy facility giving recommendations to the energy operator. So it was pretty important that we get that part right. Programming had to run. The second one is about as you go through day by day, how are you making decisions to the machines given that every day up to 2 p.m., they could call you for the next day. And so there are decisions to make about the storage. You don't want to be empty when they call you basically. And you can always assume there's a certain probability that they may call you or not or that you might run out, things like that. And so there we were using a technique called model predictive control, which is a, so I can go more in detail in the questions if there are some, but it's basically at every time step you sort of make a plan for the future. Then you solve an optimization problem saying what should I do if that plan is true? And then you implement the first step of that plan and then you move forward to the next time step and you rinse and repeat. So we developed software for both problems and this was deployed live, I think I said, and the bottom left there, that's a picture that's coming from, so this was made after the fact to show what had happened. So the picture is similar to the one that you were showing earlier, where the green is what the central energy plant is consuming, the orange is the campus and the blue is the total campus load and the parts have marked out in sort of that pinkish color there. There was a three live events in July of 2018. So I think the first one was a three hour event, the second one was a one hour event and the third one was a two hour event. And so in this case I think they could call you up to five times per summer and so that was a particularly hot period where they sort of three days in a row they said, can you give us some flexibility tomorrow? And so this worked pretty well. One of the nice things is that Stanford made a bit of money out of it, so that also on the research side bought us some goodwill to do some more work, which was always nice. To give you a sense of scale, we were participating with five megawatts of capacity, those were the bids that we were placing, which was conservative given the two-stage stochastic program that we'd been solving. We thought that's something we could do and we were able to do in practice. One of the good things about is that flexibility is actually pretty big. We were big participants in that program with a relatively small number of machines. The number of machines you control at the Central Energy Plant, that's actually pretty easy compared to some of the things that people want to do. It's like if you compare controlling the machines in that plant to I'm going to control a million fridges, it's much easier to do this for a relatively big size, so that was nice to see. One of the other takeaways for me was there's actually a lot of work to get something like this implemented and if we wanted to get this baked into the industrial software that's running the system while you need to get Johnson Controls, which is the company that's making the software that this is right now to integrate it, but they would want to do this if it makes sense for them to sell it to their other customers, which might be in different tariffs. So there's their barriers to adoption of these kinds of programs that we saw firsthand and yeah, I can talk more about that in the questions, but that was very interesting to me. So if there's one key learning for me from that first part, when we were looking at thermal storage and what can thermal storage do, it was that district scale thermal storage provides complementary services to electrochemical storage and the reason I say complementary is because you can't go as fast. You know, with electrochemical storage you can get sub-second responses. There are a lot, you know, on the power side that you can move things very fast, you can do things like frequency regulation, sort of a wide array of services, but then on the energy side, thermal storage much bigger, much cheaper, slower, yes, but also in many ways easier. So let me just, yeah, okay, I'm about right. So then the second part, actually maybe let me stop and drink a bit. Are there any questions? Okay, so either no one understood or was perfectly clear. Okay, so then, so to move sort of fast forward to more current day work, if I go back to my picture of the campus energy operations and how I schedule it, a big part of what I'm going to talk about next is about if we start having decisions that we can make inside the buildings, can those add value or maybe a better way to say that is would having decisions in the buildings give you more value than the cost to implement those decisions? And both in terms of actually doing it because there's a cost to just implementing things and also in terms of cost to say occupants. So DR there stands for demand response. So I'm looking at my original decision-making problem from the central energy facility and really here what I'm looking at is what if I added another screen, another lever, or set of levers that they could pull in that building, say I'm going to do things inside the buildings and that's what I'm generically calling demand response. So I'm going to pay a cost for that and in terms of what's happening there and the types of, so there are two types of decision variables I want to add into my problem there, there are thermostat set points. So I want to control, is there one? I don't see the one for this room, but so in many rooms in the buildings on campus, on the wall, you'll see a thermostat. So we can actually connect to those and send commands and those are more sort of, well, when they're continuous decisions and they're what you might call soft demand measures and then I always have the option to just walk to a building and shut off the valve for the water that's going to the building and just disconnect it completely which is going to have higher costs but is pretty reliable. You know that if you do it, it just turns off. And then, so if I look at my constraints so I still have the constraints I had before, now when I'm looking at satisfying loads, I'm looking at that problem on a building per building basis now and so there's somehow I'm going to have a state equation somewhere that's saying, well if I don't give a building cooling, it's going to get hotter and so I might pay for that down the road. So I need to start tracking more things and so I'm keeping what's the decisions I want to make at the central energy plant, I'm just looking at adding more of these decisions. So to start planting a seed for why could this be valuable and there are actually several different ways in which we think it could be valuable to have decisions at the central energy plant, one of them is to think about our capacity needs and I think I have a picture in the next slide where I'll show what happened during those events but part of the motivation for launching this program, John mentioned there were supply issues on campus so actually first in 2017 but not many people heard about that then in 2019 and there are a lot more people heard about that and in some ways that was not only but it was a capacity problem is sort of it's you have a certain amount of cooling you can produce every day and there we ran into so the first one was a period of two hot days where we just couldn't produce enough cooling and so we had to do the chill bar grill time and that second variable I talked about operators went and shut things off so one of the first questions is well can I reduce loads on the very hot days reduce my needs for cooling and if I can well I have to buy fewer machines and that's sort of a home run because it means less in terms of costs less in terms of carbon, in terms of space just in a lot of different things and so the plot on the right there that's plot that Ryan in the back there made and is data from I think 2020 if I'm not is that 2020? Yeah, 2020 so that but I think it's more conceptual what I'm trying to get out of it here this is what's called a load duration curve so what it's showing is every day of 365 days there every day of the year what was the total campus cooling load and then it's sorted left to right from maximum to minimum and one of the reasons people make these curves is to look at how peaky they are on the left and the peakier the curve roughly the more value there is for doing demand response because if you can reduce demand from the buildings on a certain number of, so here on five days well you have 10% less capacity than you need if you have 15 flexible days well you can get 20% reduction in capacity needs so that's actually a lot of money in the Stanford case if you can reduce capacity needs by this amount what this plot is not telling you anything about is how do you get the five flexible days and is that something that's easy or not to do and so a big part of this project is about well how do you do this so from this plot it looks like there's potential for this but if you wanted to do this you need to make some decisions or to be able to make some decisions locally you need to be able to pull triggers inside the buildings and you also need to make decisions centrally at the central energy plant you need to decide whether to call on these or not so what is this about so my catchphrase for this project is that it's about making large modern buildings what we call commercial buildings so typically like the one where right now more energy efficient, flexible and low carbon so if I think about, so I do want to highlight this because when you're thinking about managing buildings there are multiple different things you could do there are some constraints, some research constraints that we're imposing on ourselves here one of them is that we're not looking at new buildings we're looking at buildings that already exist and one of the reasons to do that is that buildings are around for a very long time I think the number from the commercial building energy consumption survey in the US is one to 2% per year is roughly the turnover so buildings are roughly there for at least 50 years on this campus many of them have been here for more than 50 years so buildings don't change that often that's one reason we want to look at existing buildings and what are we trying to do we're trying to build methods that can learn from real data so a lot of this is about running experiments and I'll show some experimental data today and then build some modeling methods to be able to learn from the methods and to relearn the behavior of the buildings day after day so why do you care about that so even though the buildings are around for a very long time the way that they're used changes in at least two ways so the bottom left there is things can happen inside the building that will modify its behavior and it can be a retrofit like a HVAC so HVAC is heating ventilation and air conditioning an HVAC system retrofit well that'll change and that's the example so this is the Wallenberg building that's an example here where the red is before the retrofit and the y-axis there is giving you a sense of what the cooling load inside the building is on a weekly time scale and so the left is before the retrofit and the red, sorry, the red is before the retrofit and the right, that's supposed to be green is after the retrofit and so you can see that the building consume less energy what's not so apparent here is actually even the patterns of the daily consumption patterns of the buildings changed before and after but another example is I can't remember which building but is there was the building was designed with the assumption that the windows were going to stay closed and then at some point that was changed the windows were now allowed to open but you're not going to rip out the HVAC system for that so somehow that's something we want to adapt to so all these are sort of things that are happening year over year or day after year, day and then there are events and we want to be able to adapt to these and so a big example for those are these cartelments that I talked about in 2019 and the plot on the bottom right there is showing these cartelments in June of 2019 so basically the first day was okay and then there were two days that were very hot with temperatures up to 100 plus Fahrenheit which for Stanford is a lot and then the third day there was also pain people were also unhappy on campus and but it wasn't so hot and the black curve here is the cooling that was actually measured coming out of the central energy plant and the blue curve was done by me and is a sort of a reconstruction if you like of what I think the buildings would have liked so if you'd given the buildings everything they wanted during those heatwaves how much they would have consumed and so to a first order the difference between the blue and the black here is pain you know this is why people weren't happy you know and so you can kind of see here on the second day it's much flatter you know basically what's happening behind the scenes there is on the first day all of our storage ran out you know and so during the night it was too warm so we weren't we were still trying to cool the buildings down we weren't able to replenish our stock of cold water so we started day two with very little water or very little energy in the tank if you like and then the reason the second day is kind of flat there flattened off is there were people from the CEF running around campus and actually physically turning these valves to disconnect buildings and so you know so I don't know how many how many people were on campus in 2019 okay actually a fair amount so yeah so I don't know how many people were where I guess in June of 2019 this is really the kind of time I care about but you know if you were there and well you probably remember the emails in any case so you know so the difference between the blue and the black is my explanation of why you got all those emails why people were angry and so one of the questions we're looking at here is can we do things differently in the short run so that the committee that John was talking about what they decide to do is to buy more capacity you know to buy more machines so that you know the idea was so that this never happens again but of course you know the campus keeps growing we're going to have to retire capacity there so this whole this entire project is looking at is there another option instead of going and you know and shutting down the valves inside the buildings can you do things that are a softer you know sort of in other words turned what are binary variables today per building into continuous variables you know so instead of zero or one I want to have a zero to a hundred kind of variable that I'm tuning down inside the buildings and I also want it to be more spatially precise you know I want to be able to shut down certain parts of the building and not the entire building at once to produce one of the issues that the campus has today is that if you have one critical room inside a building well the entire building becomes critical so if you have the option of doing distributed temperature adjustments well you can work around that so that's one thing we're very interested in so and I need to start okay I'll go quickly on this because I want to talk about some experiments we did but I can come back to it later so one thing that was fun to look at I don't know if fun is the right word interesting to look at sorry I tend to use the words like fun what I mean is interesting I mean I enjoy the research I do a lot so to me part of it's fun but here I wanted to talk a tiny bit about COVID because COVID gave us a natural experiment if you like where people were sent home so actually occupancy inside the buildings went down for very brutally you know you can kind of say exactly the date when this happened and for a prolonged period of time and one of the things one of the questions that you might immediately want to ask is well how much does occupancy affect the energy loads inside buildings and so let's just look at the top three plots there where the left one is cooling the middle one is heating and the right one is electricity and so that's the y-axis and the x-axis here are mean daily temperature so that's the trying to see what's the relationship between cooling, heating, electricity and the temperature it is outside and so I have 2019, 2020 and 2021 and if you remember roughly March that's actually the last time I saw John before COVID was I think March 15th or roughly of 2020 is when you know people started to go home so you know 2020 is a good place when this started and what you can kind of see here is cooling and heating very little impact things didn't change so much and their caveats but to first order reduced occupancy on campus didn't really reduce our heating and cooling loads electricity a lot you know so what one of the things that this sort of tells me is that first order if I wanted to say tomorrow is a demand response stay on campus let's send everyone home because we really need to consume less energy right now it would if I did it tomorrow you know if I implemented it right now it would have a big impact on the electrical loads it wouldn't have an impact on the heating cooling loads which means there's work to do and so the first thing we did is in 2020 we started running experiments back then it was in three buildings on campus where we were changing set points so when I say set point I mean basically the control so in the thermostat the control at which it's trying to keep the room and there's typically rooms are controlled to what's called a dead band so a heating and a cooling set point where if you're inside the dead band the energy system doesn't do anything it's happy and then if you go above the cooling set point you cool if you're below you heat or something like that so what we're controlling here is the temperature dead bands and we're and so the method is controlling the temperature dead band what we're trying to do is impact the cooling loads in this case inside the building so in 2020 there are two kinds of experiments we were doing we were doing experiments where we were changing those cooling set points day by day and trying to see how much that impact the building's cooling load and you can kind of see why we thought okay there's hope for doing this kind of thing on the bottom left where those are showing four days in a building over there this is a mechanical engineering building and these are four days with relatively similar outside air temperatures two days where we had a low set point two days where we had a high set point and kind of see the load goes down when you increase the temperature set point so we thought okay good this is something that has potential and then on the right we were also doing experiments where we were changing temperature set points throughout the day to try to get a sense for the building response because we want to build models for the buildings to see what they're doing and so oh gosh I really need to move faster sorry so the bottom right there those are experiments where we're increasing the temperature set point at noon and seeing the building respond and so here you can kind of see so the middle picture there is the temperatures and you can see that going up gradually and it also has a shape which I like because there's a physical model I want to put onto this and kind of matches that shape so I like that too and then the bottom right that was the campus cooling load and you can see that drop sharply in this case at noon which was also nice so a lot of potential and then in 2021 we did a lot more experiments on double actually eight buildings and there are six that I'm taking in here and so I'll talk more about this but basically what did we take from this well you know there's potential for doing this sort of thing so we were testing a two degree Fahrenheit set point change so actually pretty small and we got a pretty big response 12 to 29% in the four office and conference buildings and then laboratories we also saw response but smaller and let me jump through this just by saying that one of the things we saw is demand response can be a reliable resource so just to say a bit more about how we were doing this so the picture on the top left there is sort of a schematic for how building is operated so I'm bringing the heating and the cooling from the outside and there's one place where I extract the cooling from the chilled water and I pass it to the air and then I blow it to the different zones and then inside the zones I'm changing these temperature set points so we have a software system that's talking to roughly a thousand zones at this point spread across multiple buildings and so every day I'm changing the set point and seeing how the building responds let me jump through this and just say we want to build models on top of this this is to kind of give you a sense of the data the sort of data we're collecting so here I'm showing six different buildings and each dot on these plots corresponds to one day's worth of data so the y-axis is served cooling so this is how much the building is paying if you like to keep its temperatures at the set points and the blue is the low set point the red is the high set point and the lines that I superimpose on these graphs are my model, my simple model for how the building is consuming energy and the difference between basically the blue line and the red line on these plots is flexibility, that's how much that's the reduction I can expect from changing the temperature set points and so things I like about these lines is well they also capture how things are changing with respect to the outside temperature which is something I was interested in they give me a sense of uncertainty in the building response so when you look at the dots they're not exactly on the lines but this is something I can measure I can calculate a confidence interval around these lines to say well how much do I expect the change to be and so we were able to make plots like these so I'm just going to talk about the right part of this slide where the top plot is showing you flexibility benefits so the percentage so this is the percentage impact of the increase in the set point on energy for cooling and again we saw some pretty big impacts and maybe this is not so I guess to put things in perspective when we started this off there were actually quite a few people who managed the campus who were telling us there will be no impact you will change these set points and nothing will happen so this was nice because well for the other people who thought this is going to make a difference well you can kind of see and to some extent you'd think duh you know I increase the set point the building is consuming less energy but actually and I can talk a bit more about that if there are questions the way the building behaves is not so simple to model so saying if I increase the set point the building will consume less energy that's kind of duh but how much and how does that change with the outside air temperature that's not so trivial which is why these sorts of experiments have value and then so let me jump through this these are sort of conclusions that I think I've talked a bit about okay and so getting towards the end there and then I'll I'll have some time for questions and so there are still a lot of things we're working on it in many ways we've really just scratched the surface so I wanted to give you a sense of you know what snacks what are the sorts of questions we're still trying to answer and one of the things one of the ways we'll try to do this is we're continuing to run experiments we'll start in roughly a month I think for the 2022 season so we want to know how do buildings respond to different set points so not just the two-reference increase faster changes how about if I only write part of the building at once this is one of the things that I mentioned earlier I'm very interested in to be able to disaggregate the impacts so I talked a tiny bit about but the buildings responded differently so understanding what are the drivers for that the heating the HVAC systems it's kind of in the name are trying to do both ventilation and cooling and there's some competition in the in the requirements between the two so understanding more about that what are the so that that's also something that we're very interested in what are the costs of doing this kind of demand response and how can we measure the costs so thinking about occupants oops how can we scale so another piece where we want to learn more about is how to scale the results so I said there's roughly 150 buildings on this campus we are only going to be able to test a subset so we need to have some sort of way of saying what do we think the results would be on the larger population and at most you know the best we're going to be able to do is some sort of educated guess but even that educated guess is going to be pretty valuable both on this campus and in other places how do I take those decisions that I have now inside the buildings and put them into my central decision-making problem how can you know how can I better understand what's happening especially on the shorter timescales which as we start moving towards real-time control problems we're going to need that that the daily timescale won't be enough for those kinds of problems and how do I think about uncertainty and I'm not going to go through this and let me just say well we have in addition to great students that I have there on the right we also have some great faculty supporting us and bringing a lot of expertise on because as I hope you saw we have a lot of different questions around these that are you know one person is not going to be enough to do this and okay and this is my conclusion so I'll stop here the last thing I wanted to say because I do you know and I hope it was apparent from everything I've talked about is none of this would be possible without the people that manage this campus you know sort of the invisible people that no one ever sees that are actually you know running the system and who are well they have a vested interest because they want to keep making their system more efficient but there's actually a lot of staff time that's going to running these experiments and without this sort of partnership this really good partnership that we have with them none of this would be possible so I want to acknowledge them thanks very much probably have time for a few open questions before we do that I'd like to turn over to Melissa for a brief announcement yeah just a quick announcement for the attendance sheet I put a sheet outside of this classroom so when you're leaving just check your name and that's how we'll be tracking attendance for the rest of the quarter so any questions for Jacques I know some of you will meet with him afterwards let's start here then all these yes well first I just want to say this is like super super interesting I think like you know the first thing you learn when you learn about energy utilities is that you know really there are only levers on the on the supply side if you meet demand and you know this seems like it could be once taken to scale just like an incredibly important piece of improving energy efficiency overall my question was kind of about that that path to scale so like I'm taking Stanford as like a relative outlier and like the level of control that it can sort of achieve and that it's willingness to like invest in these in these research projects right now like do you have any sort of rough estimate for like the relationship between what upfront capital investment is required to like achieve these systems versus like the energy savings over time and then I guess sort of the second part of that question is this sort of like what what funding structures do you think potentially work obviously you know government provides a lot of money for home weatherization like are there are there potential like things you could do with with government funding here I guess that's a little broad okay so so I have a number but you kind of caught me off guard so I remember if it's 20 or 30 percent but but there's I think so let's say 25 on average but that's the the fraction of US commercial floor space that has what are called programmable thermostats you know so the sorts of controls that we have in the buildings here that's actually already pretty wide spread in the US so one of the things that is I'm sorry so you're saying software is the real yeah so so so here the so exactly software is the word I was looking for thanks John you know that that the real investment we're talking about here or at least that we're trying to talk about here is software one of the sort of guiding design decisions of this program is we're not installing any hardware so you know so so there's nothing that I'm installing in the Sanford buildings that was not already there so yes Stanford is a place that has more money than most and so there are quite a lot of smarts that are already installed but not you know not not so far-fetched I'm actually talking with a campus University of Michigan you know so I'm talking in different places that don't have you know as we're you know not even just universities that that wouldn't necessarily have as much money as Stanford but that do have these smarts so in terms of you know investments I actually think well by design we're trying to see how much can you do if you make no investment and so the in the academic literature what this kind of work is called supervisory control so that's so what what people in the buildings industry have been wanting so some academics have been wanting to do for decades is and the the key word usually here is model predictive control for buildings is rip out the the software systems that are running the buildings today and replace that with much smarter systems kind of like what they have in process engineering and chemical engineering the idea being you know we're going to do much better the idea here is not to do that at all is actually to say can I connect to the systems that are already there just build a software overlay and with that software overlay try to do better so you know it won't be as good as if I changed everything inside the building but it's something I can do much much faster it's much it's much more scalable it's much more so you know in terms of repeatability which I think you know and that kind of goes to your question about cost and that's another big design decision in the way we're running these experiments every single building is different and every single building has a different set actually it's not one software system set of software systems managing the what's in there but almost you know virtually every single room is controlled by this you know what's the top of the dead band what's the bottom of the dead band you know the this way of controlling the rooms that's almost universal so that's why you know that's the control decision we're making you know almost by design I'm saying I don't want to control the fans in the building and the vents you know all the different parts because even just internally at Stanford that would be too much work you know like the we were able to write software to talk to so I said something like a thousand zones across nine buildings each of the buildings we're working with is completely different if I was trying to rip out the software systems I wouldn't even you know be done with the first one so and maybe just like you know one last thing on the governments you know I'm hoping these things were paid will pay for themselves they're not you know like the the idea is not is really not about investments oh jeez yeah mine was about the zoning the zoning the zoning oh okay sorry so yeah let me say a bit more about that so so there was always so there already was a system to go in and change zone by zone it's just given so well one so these systems aren't used so much and so basically what one thing that you they could do is you know you can they can go in the system and say you know I'm going to change quorum one or four or you know whatever this room is and then the next one and the next one but you know like there isn't so part of this is just a scheduling system to be able to talk to those rooms independently and well there there is a bit of a and actually we're that's something we're we're working on right now there is a bit of a challenge with network issues so anyway I think that goes to the cost is the amount of information that the building systems can take once is actually pre-limited so you know so there's some work on the software side there you know just bandwidth issues because the the the systems are pretty old and so you know so so so pre-existing so it's basically we added a layer on top of pre-existing software so but but we didn't add controls we we just made it possible to use the controls one last question up here I have a good question about data generation process and you mentioned that you know if you could get binary variable to move to continuous or if you could have more stationary precise controls it would allow you to sort of build both more better experiments but also create more useful insight and I'm curious sort of what's the barriers for that and sort of how much violence do you might need to have from operations and what are the sort of conflicting priorities for that that particular objective so on the data generation side yes well we're very data poor you know so so you know so so if you think about you know like a lot of machine learning type problems we have very little data here because you're right you know or I think that that's what you were saying the the operational sequences in the buildings matter a lot you know so it's like the plots I was showing earlier we had basically 60 to 70 points per you know for the summer to train our models you know and there's sort of you know and then there's like the weather dependence there's there's what was the set point control and if I start making distributed set point changes I add much more degrees of freedom so you know the the number of features that I have that relative to the number of observations is never going to be very good you know so so one of the things one of the constraints here if you like on the on the data learning side is how do you build models that can learn from real data but from limited amounts of real data and that's that's a real constraint we're just about out of time I can't resist adding these are a great set of questions one thing that we we did have a seven hour two years ago about the LA 100 study by the mayor's office they from what I understood they've kind of done part of a lot actually quite a bit of your supply side but we're just starting to think about the demand side the group that did a lot of their analytics was at NREL the National Renewal Energy Lab so the they do I think they would confirm this they do have pretty good data so they may be more data rich than most settings so if you haven't talked to them they'd probably love to talk to you can I comment on it both LA and NREL the good question here about how do you do this elsewhere working with the NREL group their big fear that they would love to replicate what they did there this is a little bit broader but this would be a great way to accelerate the impacts what they're already doing there's they're doing kind of the whole list of things you guys worked out so LA got a lot of resources but I won't go where they got the money from but it was a very well funded so the question is if you go almost any other city in the US you don't have quite the resources but you can learn from LA on what work you can go out with your your analysis say here and hand and convince people they don't really need more money and if they do need a little bit of money it's going to have a very good rate very high rate of return so I think this really is a way forward because it's IT stuff mostly and not thermo stuff although it controls thermo stuff it's pretty easy to do it fast and just the way you described it or at least get not all of the benefits but a lot of them immediately before you need to replace the buildings which does take a long time anyways thanks so much that was a terrific talk.