 The sectors include electricity generation, also electrification as well as energy efficiency. So that means energy plays a crucial role in decarbonizing at a global scale. Maybe Chi-Wan can help me. And when it comes to the energy sector, I'd like to sort of share with you some of the recent observations that I have learned, which gave you a bit of a sense of where some of the argument is coming from. From the sources of energy, of course, the big thing is putting in a lot of renewables and carbon capture and precipitation of some of the fossil fuel generators. That's what is happening in the supply site. This is, you must have been taking some class on forecasting of prediction, but we always be very careful with the forecasting. This is the outlook of the solar installed annual capacity by IEA. And guess what? They have missed big time. This is the forecast and this is actual installation in 2020. So on the supply side, big things are happening in terms of decarbonization. On the demand side, I live in a place, a beautiful state called Texas and I'm just going to tell you something just happened in my name. Electrification is happening in big time in terms of both replacing the existing demand from non-electrical to electrical, as well as adding new kinds of electrical demands such as internet data center and cryptocurrency mining. This is the peak demand of ERCOT just in the past two years. You look at this number and it's like, wow, are we in a developing nation or what? 2021, 75 gigawatts. 2022, 80 gigawatt. 2023, this past summer, I was on TV a lot because every time they broke a record, they had to interview some kind of experts that they broke like 20 times just in this past summer. 85 gigawatts. So electrification is happening in big time. And on both ends of supply and demand, I think there's a very big important aspect to capture the decarbonization effort which people may have not talked so much about but I think it's very, very important. It's that energy efficiency. We need to do the same kinds of job with less energy and that applies to both supply side and demand side. This again is a very interesting study of the US energy consumption. Ever since the 1970s, some of you were there, you probably remember those days on the oil crisis where they put a big emphasis on energy efficiency and you look, the projection of energy consumption was way curved down. Precisely, it's not because we did not grow. The US economy has grown quite a bit. The population also has grown quite a bit. But guess what? The rate of energy use is not growing nearly as fast as what they have projected, largely thanks to energy efficiency. So to summarize, on the supply side, on the demand side and on the energy efficient side, there's a big things that need to be done at scale, at speed to do the decarbonization. We as, I see myself as a technologist, we are fortunate to be also in the era of massive digitization. By digitization, I mean the preparation of sensors, controllers, communication capabilities and hardware like power electronics. And they are digitizing both on the software side and hardware side, if you wish, that is transforming all three segments of the discussion we just had. Of course, at the end of the day, if you ask your neighbor or if you ask your mom, you say, why should we care? What do we really need at the end of the day when it comes to electric energy? I think you would not disagree with me that what we want at the end of the day is that we want to have affordable, equitable, clean and resilient electric energy services. That's what average customer cares about. But if you stop and think about these four attributes, they are not very well aligned with each other. You want affordable, maybe they're not very resilient. You want affordable, maybe they're not very clean. Our job, our community's job, the electric energy systems community's job is to really to harness the power of digitization, both in terms of software and hardware and trying to hopefully bring the four attributes a little bit closer to each other. That I think is our community's job and many of you are doing fantastic work pushing that agenda forward. Now within this scale of discussions, what are some of the scientific challenges when it comes to the electric energy systems? So I want to kind of illustrate some of the scientific challenges in the next couple of slides. On the physics side, so electric energy system is a beautiful interplay of physics and the markets with human deeply engaged in a loop of multiple timescales. So on the physics side, I would like to argue that perhaps the challenge, the fundamental challenges is that we're seeing a lot more of unknown physics and unknown models that is proliferating in the grid. And then we have to deal with a lot of these high dimension time varying intrinsic coordinates as we go in terms of dynamic system speaking, right? And that is giving rise to a lot of challenges of understanding the near real-time behavior of the grid, but also in the same time giving us a lot of opportunities to try to do the sort of state of the art kind of data sciences to advance the resiliency and the reliability of the grid. On the other hand, when it comes to the market, we almost, 80% of the US is covered in one way or the other some form of a wholesale electricity market like California, like Texas. So when it comes to the market, if I were to summarize, I think the key challenges scientifically speaking is how do you really understand the behavior at large scale of human in the loop decision making? Not one, not two, but millions of customers behaving together. And also how do you deal with this, you know, big elephant in the room, which is the renewable intermittencies, right? Through some kind of a market mechanism design and incentive designs. And when you are putting the physics and the markets together, you come up with this gigantic modeling or in the new fancy word, digital twin challenges, right? Which is I will summarize into two pieces. One is the large scale high fidelity modeling. And the other is the flip side of all these cool things on digitization has to offer, which is a cyber physical security, right? So I'm going to walk you through some of the work that we have been doing over the years on each segment of the three pieces that we just talked about. On the physics side, on the market side, and on the digital twin side, okay? Let me first start from the physics. We all love renewables. California, it's full of solar energy. Texas is the wind energy capital of the country. Great. But in place where I live, you know, we don't have to invent new problems. Just new problems are just coming to us and I just need to sit and listen to what my colleagues are talking about. And there's something very peculiar that's especially happening for high penetration of renewable areas like California, like Texas. Something perhaps not many people here talk about, but I think it's very important to be mentioned about is the dynamic instabilities, right? And largely these are caused by, you know, lack of coordination of some of the advanced power electronics controllers with very varying operating conditions. Just to give you some example, there was some 550 megawatt solar plant oscillating at seven hertz in California, causing some short-term operational challenges in the California high-end zone. There was this New England observation of 1.3 hertz of these are electromechanical dynamics that is happening on the grid all the time, right? But those which are not so well controlled or not so well maintained may have caused some magnified oscillation. And that is something that power engineers don't like very much, right? We need to try to get rid of that. And then pictorically what happened is that in a place like Texas, you have a lot of the renewables concentrated on the per handle area in the West. And these wings are being shipped over to the demand centers in Houston and Dallas and so on through long-distance transmission lines. People spend billions of dollars building this, what they call CREF project, competitive renewable energy zone project, which is aiming to bring this gigawatt and gigawatt of wind power from West Texas to the east part of the state. But because of this oscillatory behavior that we are showing here, system operators have to limit how much of power they are able to deliver over these long-distance big transmission lines to putting some analogy, not all of you are power engineer. It's like you're building an eight-lane highway. But sorry, because the cars are just driving zigzagging, you only allow four lanes to run on the cars. What causes this oscillation? It's caused by, we're gonna talk about that in just a minute. It's sort of a fundamentally, it's caused by some residence issues on dynamic systems. So I don't know about you, but as engineer, when I see this, when I see gigawatts and gigawatts of transmission capability being underutilized, I want to solve these kind of problems, right? So what can we do? Thanks to the investment from places like the Palm of Energy, this country now or North America now is well immersed with something called synchrofacers, which has this GPS synchronized capability of high resolution sensing of bulk high-voltage transmission grids. These kinds of oscillations, presumably could be one of the use cases where such sensors could be utilized to number one detect, number two locate, and then number three, hopefully correct some of these problems, right? So thanks to the partnership that we have with system operators, we're able to obtain some of this real-world data and try to start to formulate and think about the problem. At the beginning, I thought it was kind of an easy exercise. You want to find out where the biggest oscillation is and try to identify and then try to fix that. I thought that was an easy job as this data coming in, but it turns out it's a little bit harder than I thought. I'll give you an example of why it is harder. This is something actually happened near you. There was this incident that happened in the west part of the electric grid. There was a generator up in Alberta. You guys know where Alberta is, right? Up in Canada. That is creating some 20 megawatts up down, 20 megawatts in power engineering terms. It's okay, not too big. 20 megawatt of power swings. And for whatever reason, folks operating the California oracle intertie, which is a thousand miles down south, was observing at that time, 10 times magnified, 10 times magnified oscillation in the form of 200 megawatt or so fluctuations on that intertie, which was a very important intertie to ship the power from the north to California, right? At that time, these two places are 1000 miles apart. So it's like your intuition of something called oscillation is sort of a, somehow not very compatible with what happened. It's like if you throw a little stone into a pond, you see this ripple and you would have assumed the source where it hit the pond would have been the highest magnitude, right? But this is somewhat counter that intuition because apparently the source is a thousand miles away from where you see the largest, the magnitude of oscillation. And that is, I think answering your question about the mechanism. This kind of mechanism is caused by something we actually all know and we learn about in high school physics, something called a resonance. You remember this is a discussion about the army not allowed to walk in synchrony on the bridge because there was this incident in World War I that blew up one of the bridges, right? That precisely turns out, mechanism is speaking, what happened, right? So now the challenge is, if you think about a large power grid as some kind of a complex dynamic system where you have multiple inputs and you have multiple observations, this observation being the single phasers. What would be the best way for us to number one detect and number two, the most importantly, locate where the problem was because if I can locate where the problem was at least I can do something, right? So merely looking at the magnitude of this new advanced sensor is not enough. So we got to do something perhaps more than that. So that's the problem statement for this issue that we have been working with and we're caught about. We kind of got stuck, you know, we didn't really make too much of a progress for almost two, three years, you know, just don't know what to do. And then my collaborator, Pierre Kumar, give me a totally different source of inspiration. He's like, okay, have you thought about the computer surveillance video camera in your house? Said what? Said, oh, you know, this surveillance cameras that you may have installed in your house is continuously, you know, taking videos, shoots and trying to do something, right? If there's someone coming close to your house, they will trigger some kind of alert system. So what's behind the thing is that the computers are continuously doing some kind of a so-called background and foreground separation, right? They want to separate the background and the foreground. So what, of course, they cannot do it by manual. They have to do it through computer. So what is going on in the computer? What is going on in the computer is that this background, what is the mathematical term of a background? The background is something that does not change much over time, right? So that is something we call low rank, right? It's a low rank. And what is a foreground? Foreground is a person walking by your house. What's the mathematical term for that? You assume that at any given time, at any given time, there are not too many people walking around your house, right? So I guess the mathematical term for that is a sparse, right? So what is happening is that the computer is trying to separate a streaming video camera into something called a low rank matrix and something called a sparse matrix. Now, you know, in its original format, this is not a very easy problem because it turns out the problem formulation is an NP-hard problem. So computers don't have a good way to solve it. But thanks to a lot of the progress in the computational side, we have made great strides to actually reformulating the problem as one that is being convexified and can be solved very, very efficiently through algorithms such as augmented Lagrangian multiplier-based approaches or many of you may have heard of something called a robust PCA, a robust principle component analysis. So it's a very efficient, convexified algorithm that is very good at decomposing this matrix from a low rank one to the sparse one. Now you may say, what in the world are we talking about? We're talking about the power system oscillation, localization, why in the world you suddenly draw me this surveillance video problem? I don't know. We just took a leap of faith because we had done some earlier work that actually suggests that for a large power grid, very complex power grid, the single phasor data actually exhibits some strangely low dimensional behavior, even with all the noises, all the communication failure and everything. It's a very low, low rank behavior. So we kind of just thought, okay, let's try because presumably this resonance problem might be caused by the low rank power grid. So if we can remove this troublemaker, then the rest of it would still be kind of an intuition obeying, and then maybe that would give us a hint of where the problem of the source of the oscillation is. So we don't know. We just took a leap of faith, a typical data science approach, just collect this real data and we try it out. This is a problem that we tried out on a large system where there are quite a few counter-intuitive behavior. By counter-intuitive, I mean that the source of the malfunctioning generator is far away from where you observe the highest amount of magnitude of oscillation. These are exactly the kind of problem that captures the Alberta-California behavior, right? And then to this algorithm, surprisingly, out of the 44, 43 are completely correct. And even the wrong one, if you look at this, they are electrically almost, in other words, even the wrong one is suggesting of sufficiently narrowed down search space for localizing the problem. And if you're a power engineer, then say, electrically speaking, these two generators are almost identical to the grid, right? So then we took this to ERCUT, and then we tried it out. It worked pretty well. So we were very happy. We say, oh, great. I talked to the senior VP of operation and I said, Woody, go and implement this. We can help you solve this problem in Texas. He patterned my shoulder and he looked at me. He said, no, you're a good guy, I like you, but sorry, we can't do it. I said, why? He said, how can I trust this? That's a bunch of data scientists trying to play some data, and then you wanna convince me this is something I can do in the power grid? Sorry, no way. And that got us another two years of thinking. Trying to justify why it actually would work. So that would need a bit of a theoretical underpinning and model-based analysis. And again, we were kind of inspired by a conversation I had with late Professor Sanjoy Mitter, where seems to know everything, every book in the world, and he gave me this very old book by Wiener on engineering cybernetics. And there was a chapter about residents. And we took some inspiration and then we found out actually, the very easy, very interesting discovery in the sense that for any dynamic, linear dynamic systems, you can always write down the response from oscillation behavior into a residence component and the residence-free component. The residence component is the one that is giving you the trouble, right? The magnified magnitude. And turns out that no matter how many sensors you have, that matrix would always be at most rank two. And rank two was because you have a real part and imaginary part of this corresponding mode, right? So that gave us a bit of a theoretical justification of why it actually would work because we do observe in this prior work that these power systems, a residence is very low-rank behavior and low-rank properties. So we were able to justify that in the case of Texas. And then I'm pleased to report that this now is being implemented in the operational planning team, this particular technique. And then using this, in addition, we further have developed some corrective control mechanisms in which then whenever you have located such an oscillation, you can do some very quick things by control. So it means like energy storage and batteries. And that would mitigate a lot of these false oscillations. And by mitigating that, you're effectively raising the transfer capability. So the four-lane car becomes the eight-lane car road again. So effectively you are boosting up the transfer capability through gigawatts and gigawatts of power. So I think that is a very good example of where the physical dynamics lies. And then some of the data science tools can be effectively utilized to solve some of the physical problems. And along the way, we collect a whole bunch of real data. But we are again faced with another very interesting challenge out there where most other domains may not have such a problem. The electric grid is under the protection of the Energy Act 2007 on CEII. Critical energy, electric information, infrastructure information, right? So a lot of these things, if you take a step back and think about it, it's largely data science driven. You need to do some kind of machine learning tools. And the first thing that any kind of data science or machine learning tool would require is that it is requiring massive amount of training data. But if you talk to any power company, you say, I need a massive amount of training data on your problems, dynamics. Sorry, we can't give it to you. And Professor Swearing was in that control room last week where we had the lucky of having something like that, but most other places don't have that. So the first thing is, how can we generatively scale up a huge amount of real eventful dynamic data that is of use to the broader data science machine learning community while we are respecting the underlying physical characteristics of the grid, right? So that's another very interesting challenging out there. And we're making some headways along those lines and making the headways again is thanks to the fact that we are leveraging the multi-time scale as well as the differentiated level of access capabilities of different kinds of model of the grid. So at the very fundamental level, you can think about the power grid is driven by, just for the sake of generality, we can say there's a steady-state model that is governing the steady-state or quasi-steady-state behavior, like the megawatts and kilo-watts that you guys talk about. And also this dynamic time scale, let's say just focus on the electromechanical side that is governed by a set of electromechanical transient dynamics and you have measurements on both sides. On the steady-state side is covered by this power flow, which is algebraic equations. And on the dynamic side is covered by a set of dynamic ordinary differential equations. And you have measurements of different time scales. And also you have access to the physical model of different time scales. And I would say in terms of the level of difficulty of obtaining them from the real world, perhaps the least difficult one would be those quasi-steady-state measurements, like the megawatts and the load profile and so on. The most difficult one that you can obtain for the real world is perhaps the dynamic model, right? The dynamic model, sometimes even the grid operator don't have a good knowledge of that. So the question is how do you recognize the fact that there's something easier, something harder and you wanna piece together them and create a generative approaches to scale up a huge amount of useful dynamic data for your training purposes? So we took some inspiration from sort of the state-of-the-art generative model approaches, like GAN, many of you must have already used the generative adversarial networks for a lot of different purposes, but that itself will not be able to capture the underlying physical dynamic part. So how do I create a GAN that actually can follow some of the physical dynamic question that people like Professor Stackovich would say, oh, this looks like real data, right? So we actually took a very interesting combination of combining this vanilla version of the GAN with this very new idea of neural ODE's, which is a way of utilizing the neural networks to mimic the behavior of a differential, ordinary differential equations. So what happens is that we obtain things that is easiest to obtain from the real world, namely the load profiles. We create a synthetic voltage data, okay? Because the steady-state power flow model is relatively easy to obtain. So we create a synthetic voltage data that will be then used for the initial condition for the transient behavior that is governed on the bottom side. And on the bottom side, we'll use the very limited real-world data to be the trainer for this neural ODE. And then we will massively scale it up so that it looks actually real for your follow-up research purposes. So at the end of the day, this is what we created. So I want you to turn on your power engineering hat. You look at the left side and right side. You probably couldn't tell much of a difference. In other words, the right hand side is completely made up by the AI, if you wish. But it actually, you know, I would say, ah, it looks like a generator-induced oscillation. Right? I would say, oh, it looks like a generator failure or something. Right? So it is to the extent that it's almost realistic to a real-world problems. And we are able to demonstrate that in the Southwest Power Pool, which has about 1,500 nodes in the Oklahoma gas and electric territory. So that's another large-scale study that I think is of great interest to the community right now. We also have a lot of other ongoing projects on the physical dynamics. We are speeding up the toughest problem in the simulation, the electromagnetic transient process through some FPGA and ASIC-based design through a DOE project with some colleagues in computer engineering. This is something where I spend a lot of time now is that we are trying to understand the low-voltage write-through capabilities of large flexible demands. So we have been dealing with the supply side of low-voltage write-through over the years, but never have we put anything on the demand side. But thanks to our recent efforts, we're actually testing it out in the lab, trying to understand and create the world's first grid interconnection code for low-voltage write-through for large flexible demand. We're also doing something on using the so-called neural Lyapunov approach to obtain the certified region of stability for network microgrids. So network microgrids are a very promising solution for a lot of reasons such as resiliency. And we're also looking at, because in Texas, you know, Texas has more pickup trucks. They're all the other 49 states combined. So we are looking at the impact of heavy-duty vehicle. By that, I mean the class six, seven, and eight kind of vehicle electrification. And how does that mean to the grid? So that's on the physics side. I'm gonna also walk you through some of the research that we're doing on the market side, primarily focusing on understanding the human side of the behavior. So we all know one of the toughest cookies when it comes to engaging demand-side flexibility is your home and my home. Because in the United States, the residential customers is consuming about one-third of the total electricity in the country. And the retail customers, most of you, including my house, are still paying fixed kind of rate electric bills. Maybe a more advanced version is the two-tiered pricing, peak time and off-peak time. That's it. But California is changing the peak time from 5 p.m. to 7 p.m. because of the solar. How do you reflect that? How do you reflect in a real-time basis? So we have been kind of doing quite a bit of work along this line called energy coupon, which is a retail level customer behavior experiment that we did over the past seven years on trying to engage retail customers to participate into demand flexibility but keeping your existing rate structure. So it's a carrot, not a stick. So the carrot here is dynamically generated individual targeted coupons. So this is actually the app that we developed. So this is your, a lot of the research is behind the scene, but what you see is this. This is the blue is your baseline consumption. Then if you go down to the yellow zone, you get three coupons, you go to the green zone, you get, sorry, two coupons, and you go to the green zone, you get five coupons. And you might ask, why coupon? Why not direct money? Because it turns out that on a per kilowatt basis, the monetary saving you would have contributed to the utilities is very minuscule, 50 cents here and 20 cents there. If I give you 50 cents to change your behavior of your home appliances, you might say, sorry, I'm busy doing my homework. I don't have time, right? So how do you elevate that? There's actually this Nobel winning idea called the prospect theory, which says that if I give you $1 versus if I give you a coupon or a lottery ticket, which has a 0.01 chance of winning $10,000, which one would you take? It turns out that human beings, when it comes to a small expected payoff, we are a lot more risk taking than risk neutral. So by just doing this one little thing, plus the dynamic target of real time supply and demand situation, we want to test out this hypothesis that can be more effectively engaged residential customers through these kinds of coupons, right? So there's a whole bunch of research behind this, a group of researchers spend a lot of efforts and there's actually even a spinoff of knowledge on this. At the end of the day, will we be able to find that? Yes, this approach is very effective for targeting the active customers in reducing their peak time energy consumption compared to the state of the art technology at the time, which was actually, there was a beautiful study by Professor Frank Warlick at the Witteschool here, who have done a very nice study on understanding the behavior of critical peak pricing in Southern California. So what they found out is that in order to get equivalent of one kilowatt hour of reduction, the utility have to spend about 3.6 times of your retail rate to entice that customer to change their behavior. Whereas using this coupon dynamic target and real time information, we were able to show that this approach could have changed the per kilowatt hour cost of engaging the customer for the utilities by almost seven times to about half of your retail rate, right? So we did that, I mean, a lot of work. We have to literally like knock at people's door and get these experiments done in Houston area, but it was very hard. At the end of the day, five years later, we collected 550 customers. I mean, we were happy to publish a paper about these 50 customers, but then this Gates venture, the breakthrough energy people came to us and said, okay, can you do it for the whole United States? How many people do we have in the United States? 300 million. It took us five years to do 50 experiments. How in the world can do 300 million customer experiments? No way, right? So what do we do? We got stuck again. Again, we are turning our help for generative AI. So we said, okay, how about, maybe we can utilize some of these ideas from Houston area, try to understand this sort of a distribution shift from the behavior, because the customer in Houston is not going to be the same as customer in San Francisco, right? So we try to understand that sort of characterizing the distribution drift and using this conditional variational auto encoder, which is basically a different version of a conditional gain. Somebody may have used conditional gains to actually scale it up. And we are pleased to report that we actually deliver this demand response dataset to breakthrough energy, which is now available for everybody to use for the state of Texas. So that is not 50 anymore, but millions of customers on that scale. So I think it's a very interesting learning experience of going from the real world experiments and then scale it up to a much larger through the power of generative AI, right? We're also doing a lot of other interesting work on the market. This is something we have just begun in the summer, but I think it's going to be potentially quite game-changing. You know, Texas and California the same way, we are really getting through a tough problem, which is we need to build a lot more transmission lines to connect what is renewable to the demand centers. But building a transmission line in this country is a 10-year exercise, if not longer. So what do we do between now and the next 10 years? So we have been kind of playing with this idea of thinking about not only from the hardware sense, but also from an operating sense of changing from this sort of a peak time scheduling philosophy into something about average power scheduling by sticking properly strategically pairs of energy storage on some of the congested lines across the network. The end result is that you will be able to effectively reduce the need for build a lot more transmission line while increasing the network throughput significantly. And we all know thanks to experts like Professor Tui, the energy storage cost is going down like this. So the business value is getting there quite significantly. We're also understanding big time on the crypto mining's impact on the grid. The White House recently released a report about the carbon footprint of crypto mining. Texas is now 25% of the US crypto mining and about 10% of the global crypto mining just in one state. So it's about two gigawatt of a crypto mining now in Texas. Two gigawatt is no small number anymore. So what exactly is their impact on the reliability on the carbon footprint and on the electricity market price? The key takeaway message is that it turns out that location matters a lot. And that is in direct contrast to what the White House report was reporting. So there's a lot of detail on our high resolution data that is needed to characterize that. Also some new ways of constructing market clearing mechanism without the prior knowledge of the uncertainty distribution. We have also done quite a bit of work on the so-called scenario approach which give you a risk knob if you wish for the ISO, the independent system operators to design the market clearing mechanism as they see profit. And we have also done a sort of a, we kind of took inspiration from open AI gym. So we created an open grid gym. So those of you who are ISO savvy and AI savvy must have played with open AI gym for reinforcement learning course project or research topics, right? So how do we kind of introduce the state of the art reinforcement learning algorithms into the side of demand and distribution systems and lower the barrier of entry for a lot of the machine learning experts into the power market. So we created this bridge called open grid gym which allows the state of the art or AI methods to be readily integrated with software such as open DSS. Those of you who are in power systems knows open DSS. So that's on the market side. I'm going to talk just a few minutes about a very important topic which I think we actually been doing quite a bit on which is on the slip side, the flip side of the, all the cool things that digitization has to offer. And that is the cyber physical security. The large California now has more than a million solar panels. Every solar panel is equipped with something called a smart inverters, right? What does smart inverter do? Smart inverter give you the capability of maximizing the sun power to the grid and also allows you to remotely control and monitor what's going on on the panel. So how do you make sure that smart inverter does not fall under the hands of somebody malicious? Would these smart inverters be introducing some of these false oscillation behavior like what we discussed earlier? I don't know. So collaboration with some colleagues at A&M, we're doing quite a bit of work. We have actually a fairly large DOE project on this topic of defending, cyber defending all the solar panels in the future distribution grids. An idea is through something called dynamic water marking. So dynamic water marking, as you know, if you take a $100 bill, you, everybody here has a $100 bill, not really in any mark. But you look under the light, you see something embedded in the money, right? That's called watermark. What is a watermark? It is something that's not very easily deliable. It's indelible, and it can differentiate authentic money from the fake money, right? Here, the power grid is not a static system, it's a dynamically evolving system. So we are introducing dynamic watermarking. So the key idea is that you do the regular control as you would have anywhere done on the solar inverter, but on top of that control, you add a minuscule level random signal, very small random signal, but that random signal is going to propagate entire network coming back to the sensors, observing something we engineers all know about called transfer functions, right? Which is a dynamic input output relationship between what is entering here and what you're receiving on the sensor end. And then we have worked out the theory, is that a simple example of single input, single output system, but the key idea is that it turns out this is actually a necessary and sufficient condition for authenticating these kinds of attack by doing these two tests. In theory, you have to do it in a limit operation, but in practice, you can do it on a moving window horizons. So that's this $6 million project that we are in the midst of doing, trying to do that. And I guess instead of walking you through, I'll just play a video. There is a dangerous cyber vulnerability to our national electrical grid network. To solve this, we have been working on this important problem to find realistic solutions. Algorithm development for cyber attack detection from Texas A&M, corrective control development from MIT, computational algorithm development from Argonne National Lab, and validation in the lab, actual microgrid testing from IIT and Texas A&M. Then we successfully detect potential cyber attack scenarios through our proposed solution. We are implementing this general purpose cyber attack defense methodology to center point solar farm this year. They are located in Evansville, Indiana. One is called Oak Hill Solar Farm and the other one is called Volkman Solar Farm. We pioneered a real-time secure monitoring system for detecting attacks called dynamic watermarking. Here's how it works. Solar panels generate electricity, which is stabilized and inverted, then supplied to the grid. The sensor measurements communicate with the controller through the network, which is vulnerable to attacks. That's why we inject our dynamic watermarking signal into the system. Our software continuously monitors system behavior. Once an attack is detected, an alert is instantly sent to an operator. Our team brings expertise in the fields of power electronics, power systems, and cyber physical systems. So this is on the cyber physical side. So I guess I'll just try to wrap up things. I guess I spent a lot of time talking about various different aspects, but if you were to stop and think about what we talk about, I think grid is a critical vehicle for getting us to the decarbonization. And when you are working on a grid problem, you have challenges on the physics side. You have challenges on the market side. You have challenges in large-scale modeling side. And all of them are absolutely fantastic problems to work on scientifically and having a major societal impact. So we have been also doing a bit of a societal impact through policy outreach. We wrote this article actually before, if you look at the date, we published a jewel article, but also we had a companion piece on the hill. There was a predated the passage of the bipartisan infrastructure package because we felt that it's very, very important to study these issues. We're being requested by the Public Utility Commission of Texas, so the new chairwoman, her name is Kathleen. She's very worried about Texas not having enough of a supply for the next summer. So she's asking us to study the energy efficiencies. What in the world can we do in the next six months to 12 months to get Texas through another major extreme weather situations? So we're working very closely with ERCOT and with the chairwoman's office to get this done. On the blockchain side, we just organized this workshop bringing together the world's largest Bitcoin miner, such as Riot, and the system operator, so A.G. is one who is speaking, he is the head of the large flexible demand task force at ERCOT, and the priority power is one of the largest suppliers of schedulers, the qualified scheduling entity for this large demand. So trying to bring these different groups together, talk about the technical issues at a stake level. At the federal level, RAM actually came down to visit us in April, together with some other friends around the country. We are actually working with National Science Foundation to define what is important on carbon neutral electricity and mobility, and hopefully some good funding will come out of this area. And we're also helping the DOE's solar energy technology office in the three weeks time to define some of the research agenda on solar power AI at DOE. And we've also done quite a bit of just public service to do a lot of media interactions on issues surrounding the grid. So I guess I'll stop here. I'll just say that, this is a fascinating area of research and the fact that you are here is a great testament that the electric grid is a very, very centerpiece of the decarbonization and sustainability effort that many of you are pioneering for. So I applaud your efforts in pushing this agenda together. Thank you. So BPP is fairly loosely defined terms. So I suspect what you mean by BPP is the capability that you can coordinate a group of responsive demands to be dispatched, right? That's the ultimate goal is that we would like this to be beginning like BPP. So think about if you are a BPP operator, you receive a distraction from California ISO, okay, reduce five manuals. Now you work out your formula to this aggregate to cut from the 1, 2, 3, 4, 5, 10. So you should be scared to reduce this one intersection. How do you get that done? You have two ways. One is, I directly control them. But some people don't like to be directly controlled. So this is another way that you can compile. But hopefully I can entice you to the wire targets. So the way of achieving the BPP. So I want to share with you a story about that. We were doing this experiment in a neighborhood in Houston called Spectris. Not a country, but it's a neighborhood concept. We just blindly knock at people's door and get everything changed. And the end of the day when we studied this, we found out some very strange behavior which is totally, I have never expected it. It turns out that you saw that we have an active group and an inactive group type. And then you look into the active group which are the ones who are in the background. It turns out that their social economic status, they are not necessarily the poorest people on the street. They do share one thing in common. They are mostly immigrants. And mostly Indian and Chinese doctors. And because they live with a lot of Asian immigrants and they're usually middle-class immigrants. Because these immigrants with mentality, even though they might be making half a million income, they will still try to get this $5, $10 worth of football. So that's my long answer to the question of on the active side, we did not design that at the front, but we did realize that there's some social economic aspect. Yes. Are there also agents in the immigrant related to the intermediates of the local energy when it's all or for what extent we might be linked with them? So the physical cost is a combination of intermittent but not the type of intermittent we're having, but it's more of intermittent than it's much faster. So if you don't, everybody here is Indian, right? So you know something called Fourier analysis, right? So you take a wing power or solar power output, you run a Fourier analysis when you get out of it. You get a frequency spectrum, right? So what you see on the spectrum, especially that the intermittent is for the variation is across the board. Both in the minute time still that we are talking about and also in the millisecond time still that people typically don't observe. That's one. But all fundamental issue actually has to do with this controller attunement. So all these controllers, three single wings or a single solar power has a power electronics in order. And they are all controlled. And they are all tuned according to their local behavior requirement. It's like sometimes the left hand is fighting with the right hand and that's the cause of this force transmission. It's a very interesting point, yes? I have a question on the market side. You mentioned that the market does play a role in the development of the system. I think that the market design is fundamentally integrated with the availability of digitalization. So give you a concrete example. Texas this Monday just issued a new request for proposing for 3,000 megawatts of cataclysmic elements. And in the proposal, they request three things that can be qualified. It's natural generation, demand response, and energy storage. Coming back to the long motion of our VPP. If you can have the right kind of digitalization that can really engage the demand response in a wide way, then you can actually make money on it. We should go ahead and look at the emphasis because we are actually right now with just the 3,000 megawatts that share such capacity. So I will say the market has a direct relationship with the distribution community. That's just a one concrete. That's a great question. And we are still exploring the potential use case of this generated data. But I think I firmly believe that a lot of means of that generated data. One of the direct use case that we have used so far is for training the fault classification algorithm. I'll give you an example. I have a student who used to work for Korean Tech Club, the Korean Electric Power Corporation. And they used the drones to fly around, to take images of the poles, electric poles, and look at their insulators if they have failed or not failed. So you go out, take pictures, look at this insulator failure or not failure. How do we know it's failed or not failed? They use human intelligence. This 30-year experience power engineer sitting there, oh, it's a failure, this is not failed. How much can you steal out on that, right? And you develop some AI or whatever, full algorithm to say, oh, I can detect this for you, separate the 41s and not 41s. But then you say, okay, how many data samples do you have? 300 data samples over here? Is that enough? No, sorry, that's not enough. So one of the direct use for this generated approach is that we generate a whole bunch of these, they send both images. And then I speed that to that AI algorithm to do the last picture. And we have a paper that we presented where we're actually at the last 2020.