 Okay, good afternoon. Welcome to the Smoky seminar. Our speaker today is Professor Lin Bu from Malkia University. We just talked about developing digital models for distributed energy resources. Our next seminar is in two weeks, May 18th. So now next week, two weeks. So we have two more presentations on low flexibility and demand side management. So if you have any questions for those who participate online, you should use the UQ&A feature to send us your questions. They will be addressed at the end of the presentation. So a quick introduction of the speaker today. Professor Lu is a professor in the ECE department at North Carolina State University. This is a traditional powerhouse in power systems. There are a few traditional powerhouses. This is one of them. She is a fellow IEEE and has over 30 years of experience in electric power engineering. She received her PhD from RPI in 2002 and from 2003 for 10 years. She was a senior research engineer with Pacific Northwest National Lab. This is in the state of Washington. Dr. Lu's research includes no modeling and control, energy management systems, renewable integration, microgrid modeling and control, real-time last-scale co-exemination systems and metadata analysis. She has published more than 200 papers. And without further delay, I'll let Dr. Lu share her slides and start our presentation. Yeah, I can share. So it's a great pleasure for me to come all this way out with you, the future engineers, policy makers, hopefully. That's on you. So that's why I take all the time and then travel this far. And then I hope that the time will be well spent. I know that most of you do not have solid power systems background. Rather, you probably have a mixed engineering and maybe law or economics. So I'm not going to talk too much about the technical details. I will introduce this power system digital twin as an environment that you can use to do your own research for a city. So our work is mainly sponsored by Department of Energy. And then we've worked with the New York Power Authority and Strata Solar, who is a solar developer, and then municipal utilities. The reason I mention that is because our research is built on the real system of auto and data. And then we work with these utilities and solar developers closely at the real data sets to build the system. So those are the students involved in the study. We have been working on this for about six years now. So it's a collaborative effort. So here's the outline of today's talk. And first, I'm going to give you an overview for the parts platform we developed. I may be focused on the design considerations and then tell you why among so many parts systems simulation systems. This approach is unique. And then I'm going to talk about these key supporting technologies. I won't have time to go over all of them. So I'm going to focus on three aspects. One is the data and the scenario generation and the control and energy management system model. So first I'm going to introduce this platform to you. Here is the structure of this platform. You can see that we have a model from the bottom, which is every building, every household, and then we attribute them up to our distribution system. And then that includes the PV farms here at a five megawatt level. And this is 100 megawatt level. And then these are just tens of KW. You can think about this as rooftop PV systems. And then this is a solar farm out in the field. And this is a large field solar farms. The reason we model them in different ways is because these are distributed down in the bottom, very small in capacity, but the large in quantity. And then they integrate into the power grid to the distribution system. And then for these large, like 100 or 500 megawatt solar farms, they directly integrate into the power transmission system. And then they can provide larger grid services, for example, black star or regulation service directly by themselves. But for these distributed energy resources, you have to aggregate them all together. You have to have an aggregator for them to be able to provide services. So we model those in the real time simulation platform. So a lot of you probably haven't heard of it. It is a real time simulator. So real time simulation is different from the traditional way of simulating grid. One second elapsed on this distribution system is actually one second in real world. Why do we want to do that? Because when we build this system, we want to make it a digital train so that we can connect the equipment to this system instead of passing this equipment in the field. In that case, we need the simulation system to actually mimic what will happen in the field. Meaning one second elapsed in the system is actually one second elapsed in the field. So that's why we use this real time simulator on the platform to build our model. Hardware in the loop is HIL. You're going to see that a lot as well. It means that from our system, we can connect to, say, a solar panel inverter, which is an actual equipment. We can also connect that externally to a control system, not necessarily a piece of hardware, but a software. We have a control energy management system on my computer. I can link it to this hardware in the loop simulation system and then take the data out and decide what to do and then feed the control command back to the system. That way that system model in this HIL test bed will actually not only be able to model what the system will do, but also model how the response will be. Meaning if I take action, the system will know that and then it will react to my action. So this is the idea behind that. This is a structure of the test system. We have a piece of hardware. We build one test system there, but our grid is large. You know that transmission system have thousands of nodes. Distribution system also have hundreds of thousands of nodes. So how we solve this ability issue, we can put one system in one core and then we can model them in multiple core. For example, in our case, we model micro grids or these solar panels are inverters in one core and then the distribution system on another. And then we have our collaborators. This is a national lab. They model the whole New York transmission system on another court. And then this is our another collaborator in South Carolina. This is in Washington and this is us in North Carolina. So our system model different parts of the system, but we can connect the system together through BDN link or a simple file transfer. For example, we have this Texas UT Austin, we can send our command to their panel and then their inverter to control that inverse action. And then that action will be feedback to us and then we know how this action would be as if in field. And then we can beat this action, the result of it to say in and out and they are modeling transmission level impact. So by doing so, we can have an integrated test system. So from rooftop panel all the way up to a transformer to a distribution feeder and then to the transmission line. So this is our idea. So we can distribute our models of different simulation platform and then we can connect them together as a network of digital twins to model larger system. So we've expanded the scalability by doing that. So what is the fundamental challenges we're trying to resolve when we paid us like $6 million. What we're going to deliver to them. First of all, what is the current system kind of do that we can do the current system in lack of adaptability adaptability is very important because you have one model developed. And then usually utility won't update that until a year later or maybe several years they use the same parameter for that system. And then there's not shots. So they won't adapt to the actual field measurement. What we're trying to do is we want to not only make this a system realistic meaning that they can accurately represent the physical system behind you. The past system behavior because you're using historical data to drive that, but also you run it in online and then your real world you can collect the measurement. We can use that measurement to change our parameter so they can adapt to different operation conditions. So what is the response? Responsiveness means I can take contractions in the past. We have a simulation test system. It runs based on the fixed control setting. But right now what we're trying to do is that we connect our test system with the control system. Whenever we responded to a real world scenario we take an action. This test system would get that action open the breaker or raise the load or based on the first drop the load that responsibly need to be modeled. So we need this system to be scalable meaning we cannot just model a small city because this whole regional grid included tens of cities or hundreds of cities. So we need to make this scalable and then capture this uncertainty and variability, meaning in real world, there are a lot of measurement arrows. If your simulation system you have this really clean sine wave, then when you design your control circuit, it will not have any arrow because it doesn't really model the noise. So a real digital twin should be able to model this uncertainty and variability in real world as well. And the last, when we have all these models developed, we need to make the model more compact because we cannot rely on extensive computing platforms. For example, we pay 70 K just to get one or two corset high. So if you want to build a larger system, that means you have to spend millions of dollars just on the computing platform. So we need to strike a balance between accuracy and complexity. So this idea is very similar to training AI for self driving vehicles. You won't ask a human to sit in that vehicle and drive around the world and trying to train your AI. The majority AI will be pre-trained in a simulated environment where you can model the cars. They are driving here and then when your model is more responsive, that means this car see you, it will take an action. You need to be able to model that. Otherwise, you take a picture here and then you take the position here. If it's a static model, it's a snapshot and then these people walking on the street will not move. And then when you train a model using these fixed objects, it won't be responsive. So that's why we wanted the whole thing, not only be mimicking where exactly this position is, but when time goes by, this object will move around as if there is a human moving around. So I'm using this as an example. These are the measurements you took from the real world and then you will model these human beings in your fictitious world, making them move based on their reaction. I see a car coming. I won't stand there making it hit me. I will probably walk and see the signal and then walk or not. Just make that decision as well. So here's the design consideration for our platform. First of all, we use the past data. We will focus on the future. Ideally, we only need one set of parameters that based on the past, I can accurately focus on the future. However, in real life, your model of parameters will only be accurate for certain operation conditions for certain set of actions. So what will happen is at this point of time, I might take an action and then I use the fixed parameter. You'll see this is what will happen in real world, but your model system probably will go down like that. And then the model in the real world will start to depart from each other. Then if you're still using this model to make a decision, then you won't make the right decision. So that's why we need to make it adaptive and responsive. Because when there's an action here, I either change this model so you can respond to it. Or if I don't have a responsive mechanism, I change my model, making the parameter change, reflect that change. So at this point, detect the arrow is greater than a threshold. Then I'll go back, use this real time measurement to decide what would be the model parameter I will use from this point after. So I can make a model adapt to the new situation. So this is the design philosophy. You can see that for those of you who have very little power system background, I'll tell you how the major component to build a power system model would be. First of all, you need to model typology because power system is a huge network. This is a transmission system. A regional power grid at least will have 2000 boxes to reflect all these generators, equipment, and all the load here, you will have an aggregated load. But each load here is actually representing another thousand nodes, which is the distribution system. This particular node probably will be San Francisco. And then here you probably have tens of meters out there. So in the old days, what you would do is that you will first decide what would be the typology. And second, on this typology, these will be generators, these arrows will be load. What are these models would be? And then you parameter write them. In the old days, these are single phase models because in transmission level, ABC faces, they're mostly balanced. So I can use one phase model the rest of the phase. However, in distribution system, you supply single phase load. On that page, you can have tens of loads. And then on phase B, you have maybe 20 rows. And then on phase C, you have five rows. So they're unbalanced. You can see there's a different color represented in the phases. So they use different type of models. So traditionally, they're modeled separately. And then we also will model steady state. You can see that if you have these slow changes, that's steady state model. And then the dynamic model is you suddenly have a generator draw up line, frequency will change suddenly. And then you're from one steady state moving to another steady state. That's how we call the dynamic model. So in the old days, we used to fix the set of parameter determined by case norecal data. So in the new, what we propose right now was the advancement of technology. Compared with this existing state of art. This is our advancement. So in all the days that we use the worst case scenario, only a few snapshots. So for one year, you probably study 10 cases 10 snapshot peak load scenario. That's it. And then you're going to go to the parameter or model parameter. They update once a year or even longer. But nowadays we have a lot of distributed energy resources. What that means to the system. We need to have a lot of distributed energy resources into the distribution system. They are really together. They will have a significant impact on transmission. So now we need an interface between transmission and distribution. The transmission need to not know one load for the whole day. It needs to know that when time passed by every certain minutes, gave me an update of your newest load. What is the status of your distributed resources? So nowadays the first thing is we need to model distributed energy resources. And the second thing is that the transmission distribution need to be modeled together. Because there are so many distributed resources. You need to draw it to distribution that their aggregate impact start to affect the transmission grid operation. And then the next change will be steady state and the dynamic. They need to be modeled together. Why is that? Because these smaller distributed resources, when there's a fault happening here, a lot of them maybe shed offline. And then that dynamic response will decide what will be your next steady state operation point is. So we need to model the dynamics in order to know what's the next steady state operation point is. So the last piece is we need not only time series stimulation, but the dynamically calculate and validate the model, whether or not they can mimic the actual world of behavior. So we need to dynamically calculate and validate the model parameters using the measurement data. And then the system typologies. We cannot lose only one typology because when we have this many renewable in the system, they can be taken offline because the wind is not blowing, the sun is not shining, they don't generate it. And then you have to dispatch some other generator from remote and then change the model to do that. So if you use only one typology, model the whole scenario is obviously not going to cover all the original addition. So we need more scenarios to design control so it can adapt to all these kind of different operations scenarios. So the last piece of change is right now we have communication networks. So you not only need to model the physical system, but also the control system. Like I said, you have a vehicle. You drive on the road, you need to make a decision based on how many people are crossing the road to decide if you slow down, wait for how long for them to pass. Same thing here. We need to have a control system based on the actual grid operation addition decide whether or not I need to redistribute our power flow to compensate for, for example, one area. The wind is not blowing, I need to shift power from this area to this area. Then that kind of response need to be made here in real time. And then if I decide that is going to happen, I need to send back the command to the hardware and the system to this digital train to say hey, now we need to increase the generation on this part to compensate for the deficiency in this part. If I send the command sending to the system, then they will mimic the consequence. We have three or four plans. So we need to decide which one is the best. So this part of the simulation need to run faster than real time. You design ABC, and then we model the consequence with a plan B is the best plan, because it's the least cost plan, the least chance will cause congestion. Or the least cost for this energy price to jump up in this area. Then we can execute plan B. So that's how we use this kind of digital twin. We can make better decisions because we can take the data and then we understand what will happen. And then we tell the grid what we should do. And then once you decide that in the simulated system, you can send the command to your actual system through the communication to do a better control for your real time system. So that's the design consideration. So now it is a high level consideration. I'm going to give you a few examples to see how this is actually implemented. So this is just a comparison for the existing model in our model. If you're interested in, you can dig into these references. So the first thing we do is we model the dynamics from microsecond, like the switching, but these inverters, they're switching at microsecond level. And then they are very fast single phase. And then we can model them using phaser model, which is maybe second. So when we model these different system together, they are asynchronous simulation because we run them on different platforms. This is possible. But we need to add an interface to communicate with each other. So this circuit will know what happened in the macro grid. And then the macro grid take the information from the whole circuit, decide what is next mode. We will have to consider the dynamic response inside different systems and then simulate them in an integrated fashion. And then second, we need to model the control. You can see here, this device level controller and system level controller, they have an interaction. For example, if I'm a VR aggregator, then I talk to all these different resources, like grid forming hybrid energy system, which have perfect control. We can have demand response. Some of them are on some of them are off. So the off lines may not be able to use at this moment. Similarly, we have these grid following DRs. Some of them, the communication link may be down. Some of them may not. But we need to model this device level controller, how they can interact with system level controller. So this allows us to do that. Once we have this kind of simulation, we can trim machine learning agent. So this is one application for our digital training. We have this environment. We simulate high fidelity models that when we take an action, the environment will tell me whether or not there is a congestion. There is a voltage violation. There's an over circuit over those circuits. Then it will be me tell me you use this control. I'll punish you. You use this control you solve the problem. So the first thing that we can enable is for you to use this environment to train your machine learning agent. Because when you have a simulator, then you can tell this environment what do you do instead of going to the actual grid. And you can use this control because that will be cost a prohibitively high because nobody will let you train your agent. Because you can cost a black house, right? This is the first use of it. And then the second use of it is that we usually do some like a separate tax. Because we can model not only this but also the communication links. What we can do is that we can actually model how each this place you can inject some fake data to bring the system down, especially if you have distributed energy resources. Then you have meetings of these resources you want to control what if some of these people control the certain link and then put these malware on this link and trying to bring your system down. So this is one thing we do we use reinforce some learning based approach trying to design the attacker or the defender. Again, this requires to use this environment because this environment will send us data and then depending on how you set up this scenario you can attack either this uplink or the downlink. This is a monitor link and this is controlling. So you decide what you do. And then for example, our student is studying this is his thesis you can reference to it. We can attack the, the battery unit, because if we can attack the battery unit, we can make you believe that my battery is fully charged, but the reality already these charges up. Why this is important, because I can target a three o'clock at the computer battery. And then after that, you have no opportunity to charge it anymore because PD is out. So for the rest of the day, how to shut down. If this happened to be a system supplying critical load you can imagine, I let you believe that you at a charging level higher than you expected what will happen the rest of the day. So this is another use of this high fidelity simulation system, we can model several tasks, because we can model the consequences, as if it's in field, but in field nobody will allow you to do this kind of study because you just cannot inject these fake data to a real system. So this is another use case. Let me see. So, when we do the this environment, we can design different microgrid energy management system. This is our system. What we are doing is that we try to see if we have a different microcontrollers how I can control this circuit. And then the next thing we will do is actually. You can see how at a system level we can coordinate these two microwave. You can see that when we coordinate that if I have more resource, I can supply a larger area like this area to is enlarge an area one string. So when we do this, we can not only design individual microgrid controller, but also design system level coordination. We can also do like a man who will emerge, we merge these two together. So, giving us this kind of high fidelity simulation system, we can not only design the bottom of this but also we can test that in different kind of scenarios. So this is how we use these models. So this is the design philosophy, and then how we can use that. And then the next step I will talk about some data related work, because we have this test system. We have to make sure it represents enough scenarios how we are going to do that. In reality, utility have this data set, but these are classified and proprietary information. But if we want to train an agent, we need a huge amount of data. If the utility won't want to give me that data, what could I do? And usually what they can say, they will give you say, hey, I'll give you 15 feet or to start with and then use that development methodology and tell me what to do. But what the problem of that, when you use only 15 feet or so when thousands of feet or so out there, then this data, even though you can use it to develop an algorithm, this is not going to be generalizable, because if you use this the feeders here, but you don't know if the typology change, the load change will happen. So how we are going to solve this data deficiency problem, especially in this world that everybody is holding tight on their data set and refuse to share it. So what we're trying to do is we're generating the static data set. We use real data as an input, and then we develop a machine learning algorithm. We use a database model, which is used a lot in image processing. We use this realistic data feeding into this generator model, and then we generate a deep-fade typology, deep-fade load profile, so that we can represent a different kind of operation conditions. But they are generated from actual data sets. Like if I learn from 1,000 phases, and then I can use that to generate people's face and then ask this machine to recognize their face, whether or not that's human or which part of the country they're from. So again, we call this our person model usually requires typology and a component model and then parameters. So today I'm going to briefly touch how we generate these synthetic typologies, because you can see that it will use only one typology to do the work is obviously not sufficient. So how we can use this actual typology, generate hundreds or thousands of typologies, allow us to try our developed method will work on different type of typology or not. And then next, I'm going to talk about how to use a one-set, like say a hundred or 1,000 smart meter data, generate millions of smart meter data, reflecting the actual user behavior. So here is our actual feeder. This is the HV system. A lot of research are done by HV system, which is standardized system assimilated everything. But you can see from actual feeders, the shape and then the location is very different from a test system. So how are we going to handle that? One way of doing that is that this is an actual feeder. You can see it has a lot more detail than actually test system. So one way of doing that is directly taking this actual feeder, reduce them to different type of models and then use that typology. So this typology has the advantage of its realistic because it's directly from the actual data set. It is an actual representation without typology. The disadvantage of this approach is when you have one type of typology, you only generate so many associated typology with it. It doesn't have the diversity of it. So the approach we are applying is using this image processing way of doing that. We use the generator, the discriminator. We take maybe 50 feeders and then we took even the parts of the feeder and put them in the generator. We generate a typology. We use the discriminator to say, hey, this is the actual typology. This is a fate. So by doing this competitive learning, the generator eventually learn how to generate a typology, this discriminator will say yes, this is an actual. So by doing that, we can use a limited number of actual typology, generate thousands of typology, have a similar characteristic. But it can extend the typology range to a different domain, it's even the magnitude. That way we can use this to do a lot more study. You can see that this is our deep-fated typology. So using these typologies, we can generate different shapes and then for different branches. It's like a tree. We take a hundred tree. We look at their trunk, their branch, and then from that we learn to generate the shape of the tree, similar to these 100. And we can generate a thousand types of trees. And then use that 1000 of trees, we can train our agent to see if they behave all well in this 1000 instead of only merely 100 trees. This is the idea behind that. It doesn't have issues if you directly borrow these image processing software. First of all, this is an actual physical grid. When you directly generate a tree, if that tree is okay, you cannot climb back time. But in the parses model, if this model does not meet the physical requirement, you cannot use it. So you need to do post-processing. First of all, you need to ramp our flow. So see if this voltage is being limited, if this line parameter actually set up right, so you won't have huge voltage drop, which, in reality, will never happen. Or you need to monitor the mode class. Because if you train it too well, you can see eventually load class, and then you want to generate a diversified apology. So in part system, although we can use some existing state of darts to image processing and speed processing, but we do need to use our domain expertise to fine-tune the model or to check the model to make sure they work in the actual system as well. So the next step is the component model. After we generate the apology, the next step is to give these loads different shapes. Before that, we actually need a high-resolution data. However, all the smart media data is 15 minutes. What's the advantage of that? This is the 15 minutes data. This is 30 minutes data. But in reality, we want to model the dynamics. We want to model the sharp changes. So we need to know the high-resolution data. This is very similar to image processing. You have a blurred image. You want a sharp image. How are you going to do it? You train an image learning model to see how you can compensate to this and then put the detail back into this blurred image. So what we're trying to do is that we take 30 minutes, 15 minutes data. We try to recover all these details. I won't touch the technical details because we have a YouTube video here. You can check out this name of the student or go to my website. The YouTube video tells you about the details. But I want to convey this idea. Why we do that? Because in our great models, even though you want to study, for example, the demand response. Overall, the energy may be okay. But if you demand a response, hang on and off hundreds or millions of devices off, off, off. You will have a sharp, large change field. And then that could bring the system frequency down. So that's why we need to model not only 30 minutes average of power consumption, but also 1 minutes of power change instead of energy. We need to power ramps. That's why we need to do this super resolution. We did a lot of our fine tuning. First of all, we need to talk about how we can directly borrow that and then it will work, but actually it won't because the loss will be different. So we do a fine tuning. So the fine tuning idea would be this is the ground truth. This is our actual system data. This is before we polished this load profile. You can see that it has a lot of sharp increases, which in reality will never happen. So we designed the shape loss function and then design this run function to force this generated load profile to have similar one choose that type of over. We call it that we shape that the upper run. And then we have this envelope to be similar. So that way we will be able to generate a very similar load profiles and then varying the ramping character. So when we use it to simulate demand response, we can do that. So the other thing about that is how to make a load more realistic because you model demand response. What people used to do is they'll take a load shape. This is a 15 megawatt circuit and then every node on the speeder, they take the same load shape and then they just normalize based on the transformer capacity. For example, this transformer is 100 KVA. It's only just divided this by 15 MVA and then times 100. So now let's do that to the same shape. What's the disadvantage of that? Everybody have the same shape. They will actually have a, everybody will peak together. Everybody will go to the valley here. What's the problem with that? It does not have the diversity, right? Because in different parts of the circuit, in reality, they look like this. Part of the circuit probably peaking here, part of the circuit probably have multiple peaks, and then this probably particular load will peak here. So in order to model that diversity, we need to generate a group of load profiles together so that there are not only temporal coordination, but geographical similarity. Meaning, suppose that we are in the same neighborhood. We see the same weather, right? So our peak time probably the same because if we have this really hot afternoon, our load probably will peak in the same time. But if you randomly draw these load profiles on different area, they won't have that temporal correlation, right? And then similarly, if we live in the same neighborhood, our square footage is same, and our living cabin problem will be same, right? So you will have similar load profiles. So we want to capture that. So what we do is we use a model, but we don't generate one at a time. Similarly, if you generate cat, you can generate one cat at a time. But if you pull all these cat together, they bear different characteristics. What if I want to generate a family of cat? What do we want to do? We want to have a group of character reflected. For example, there's a cat mound, this is a baby's bed. They bear similar characteristics. You can put different kinds together. That's why we developed a multi-node gun to solve this issue. So what we do is basically, we face a very interesting issue. First of all, we call this load profile to image. Meaning, suppose we have 100 load profiles. We use color to represent the high and the low. This is a 1kw, this is 15. And then if we think about this one line as this, you squish them, but use the color to represent the power. And then we generate a group of load profiles like it generates a color patch. So by doing that, we will transfer these load profiles into a color patch. And then using that, we can use the state-of-the-art gambes model to generate a picture. And this picture will be decoded into the load profiles. So one interesting observation we have is for a cat, if they're not beyond the same family, you can use a visual inspection, say, this is a fake. That is true. But for these load profiles, it's very hard. You need to be a professional, like working in this area, knowing what this group level characteristic is. So for a general audience or the user, they won't be able to do that. So what we're doing is we use this deep learning-based classifier, which we machine to tell us whether or not these loads are under the same transformer. So you can see nowadays, power systems in here, they're not just learning power system stuff. Instead, we do a lot of machine learning. Actually, my students find a job in Facebook, Uber, and Apple because they use these machine learning-based methods to automatically generate these profiles. And then they use this machine learning-based method to automatically generate the parameters. So this is right now how the power engineers will do. So I'm not going to talk too much. I'll briefly touch the parameterization as well. For parameterization, you can see that we can have very sophisticated models, but they have different type of loads there. What could we do with it? Because if we have this many parameters, we need to populate these parameters. So one thing we can do is it usually load the desidation. So what we're trying to do is we take this profile, and then we decompose them into different transformer. And that would be the possible load of all they have. And then for the buildings, if you have a smart meter measurement, what will be your HAC load? Why we need to know that? Because a lot of the non-responsive resources is HAC or water heater, like air conditioning or water heater. So we take the smart meter measurement, we tell the user at this moment of time how much HAC load is there. And then these are the resources you can use to beat into the market, say, hey, I'm going to provide this much of a demand response. So we call that a real-time parameterization. Because we take your smart meter measurement, you need to decompose them into the HAC load, PV load, EV load. And then we tell you what to do with it when you manage the demand response. So this is more about the parameterization. The next thing that you can see that we try to not only do this on individual buildings, so this is the individual building, and then we decompose what the HAC is. We also try to do them on aggravated buildings, like 10 users. What would be the HAC load for these 10 users, 15 users and 100 users? So once we take the smart meter data from a building, or take the measurement from a transformer, or take the measurement from a whole feeder, we'll tell the operators, here you have this many air-conditioned load you can use for demand response. And then they will decide, okay, based on this, I probably can use 20kW. And then which user I picked. So if I have these users, I know each individual, I can see at this moment in time who's contributing the most, and I will pick that as a response resource. Instead of, I pick 10 users, some of them, the HAC load here is actually zero. Then I pay them the money, but they didn't actually provide any response. So these kind of load desegregation are very important as well. This is the number one application that actually wants from us, because they want to see the efficacy of their current demand response programs. So the last one is more on the energy management. For this piece, when we develop the energy management, what we are trying to do is a per face demand response. I think a lot of you probably will be interested in that, because this is a distribution feeder. And usually when you have a load, they're not supplied by three phase. They're usually phase A, like the red and green is phase B, and yellow is or brown is a phase A, B, C. They have different kind of colors. So when we do demand response, we do per face demand response, meaning each demand response resource will report to me their phase, and then they're magnitude. Why we need to do that? Because for us to do this a microgrid control, you can see that the three phase can be very unbalanced. This is phase A, load, phase B, phase C. What's the problem with that? In a circuit, if I have phase A really, really high, my voltage will be distorted. And then my protection relay may not work. That's the first thing. And second, the power quality may be an issue. Some of the motor loads may not function. So if we do per phase demand response, what I think I have two more minutes, I'll just cover this part and then we'll start to get into Q&A. So what we do is different from most of the demand response we consider per phase, meaning right now, even though I see there's no need, I have enough power, I don't need to respond to that, but I want to reduce the phase A load so that it will close to phase B and C to remove that imbalance. And then the other thing we can do is to remove the arrow between the focus. Here, the focus is here and then my actual load is there. It's under focus. So if I do demand response, I can remove this condition like if I have under focus or over focus, I can use demand response to compensate for that. So you can see that this is the actual load in a real system. You can see that for group A, this is a phase A load and then this is a phase C load. So A and C are very close, but phase B is almost like one-fourth of these two loads. And then this is the second group that phase C and phase A are close, but phase A and B are close, but phase C is lower. So in reality, this is often the case. So if in this case you have different load group inside this feeder, what's the problem? If you have microgrid and then I need to supply different load and then I need to maintain phase balance. If this is my power source is at, then I need to pick up group four because this group four or groups two and three would probably help me balancing the whole load. But if I only pick up one and two or one and three, then I probably won't have a balanced load. So this will be unfair to certain loads. This is another example. If I have a power source at two, this one is balanced. Then I only want to pick up another load group that have balanced load. I don't want to pick up grid A because if I pick it up, if I don't have demand response, it will make my load very unbalanced. So to group A, it's unfair. So that's why we want to use demand response to remove this imbalance so that all the group can be served equally. So this is another thing that we are considering why we do load this location because then we can have a real-time monitoring of these demand response resources. We can control the phase balance amount of these loads. I think I'm going to stop right here because I don't want to cover too many technical details. I don't think that's useful to you. I'm going to go to the conclusion directly. Today I focused more on the digital twin-phase study. The digital twin-phase study can cover study state and dynamic response altogether. And then it can model the communication links. By doing that, we will be able to let you test your algorithms. And then we also can control this communication delay. So when we do a load aggregation study, we use this communication delay to see if we have an arrow. Will you still reach your control goal? So model the separate layer is very important as well. And then we talk about the high fidelity digital twin to compare, to use a field measurement to quantify our digital twin. So that it's responsive and then it's also highly accurate. It's not like a static model, use it all year long. Instead, we update our model parameter every hour or so to cope with the new operation condition. So the challenge is still on the data acquisition side. Because each time when you have a limited data set, you need to try to use a synthetic way to augment the data. It generates a lot more synthetic data set. But if the synthetic data set is generated from a very small group of data, then your synthetic data set can only reflect the characteristics of the data given. So you still need to get more diversified loads to start with, even though you generate a lot of synthetic data, but you need to pay attention to the data you ingested with the limitation of that. Okay, I'm just going to end my talk here. A lot of these are actually the technical details. We are still going to have YouTube videos. So you can just go to my website and check the papers. And then if you're interested in certain machine learning technologies, you can just go and find the YouTube video. That should give you more technical details. But I think that high fidelity digital training will be very useful for you to do machine learning trainings. And then also try your policy to see how the policy will work in this real world. Because it can project into the future by setting this digital training to different renewable penetrations or different control mechanisms. So there are a lot to be done. But I hope that if you're interested in this research, we'll collaborate in the future. Now I'm happy to answer the questions regarding the modeling side. What do you do? I don't think this group will do the model yourself. But I think you will do a lot of data processing. So what do you see that for these type of test system, if it's out there, how you're going to use it? I think we've studied a lot. We work with the utility engineers. And operators use this a lot because they want to know when they implement the new kind of technology, what will happen? But so far, we don't have a lot of policy because we use it. Absolutely. Yeah, this is very interesting because we do have this kind of work right now is ongoing. I did mention that. This is our test system. So what we're trying to do is we actually have access to the data from over 100 households and then from 10 Tesla charging stations. So we're trying to use this synthetic way of generating those. So based on this 100 household and based on this 10 Tesla charging station, we want to generate a lot more charging profiles similar to that. So we can populate into here because each this distribution be the supply of thousands of loads, right? We only have 100 households won't be able to study their behavior. So that's why we extrapolating them to see if we generate a synthetic data set, can we populate that these say in the future with 100% of the penetration. Then how does the feeder is going to model that whether or not it's going to overload our transformer here or overload our line here. So that's the way we use this synthetic data for as we collect a limited side of the data, because right now the data is very smart to go. You just have 10s or 100 available for one year right now. I don't know in your study how many you use. I think that's a good point. I think that's not like this cause we've done two separate projects. That's the one using the networks for generation. We did that 2021. And then also with another project using like about a couple of three million sessions. I don't create a statistical based model and look at the different driver groups and try to be able to generate like each other profile as well. Yeah, I think there's a lot of energy between what we're trying to do as well. Yes. It actually is another parallel method. You can use statistics of me. You can decide that the probability when you're charged when you're not charging and about that way. So, yeah, yeah, I think that's another way of doing this kind of synthetic data set. But I don't think the character that was our superior because they can capture, as you said, those spikes. Yeah. Yeah, the main reason we use generating method is also because it's automatic. We started with the same method, but you see when you generate the 1000, like you say, then you have to set up the parameters. Right. If you make the machine learn that and then generate that, it will be far more efficient in the future. You want to generate more. It's more like when you're trying to generate like 10 fake human pictures, you can draw them. Right. But it's very slow. So if you let the machine learn that good enough, it suddenly can generate a million faces. So we, the reason we use that is the automation part. But I think that's a very good research direction. So any other one month share your research and how it is going to be used.