 Hi everyone. Welcome to the SmartGrid seminar. This quarter we have invited Stanford postdocs to share the research with us. Our first speaker is Dr. Nicholas Astier, who works with Professor Charles Kostak, Professor Ron Rajagopa and Professor Frank Warlock. These are the similar presentations for this quarter. It's my great pleasure to introduce Nicholas, who's been a postdoc here for almost two years now, and we'll be going back to France and at the end of the academic year. He's on leave from the French Energy Regulatory Commission, which Russ is a real bonus. He certainly knows his academics, but he also has an appreciation for the regulatory environment within which electric power operates in many countries. He has a PhD from Toulouse in southern France, southwestern France, one of the strongest programs in economics in Europe actually. His PhD was on the economics of modern electricity markets, which makes it particularly relevant for our discussion today. So let me welcome Nicholas, who's going to be speaking on the topic of distributed generation and the need for network expansion. So Nicholas, take it well. Okay. Thank you, Shari. So thank you very much, Chimua and Shari, for the nice introduction. So the topic of today's talk will be, so today's talk is joined with, is some more joined with Professor Rajagopa in civil and environmental engineering and Professor Warlock in our economics department. And the topic we're interested in is the impact of distributed generation on the need for network expansion. So when we talk about distributed generation in the context of the HECD grid, being for now of rooftop PV. And so each time a project that involves rooftop PV makes it to the news. So for example, here this project in New York City, which involves a lot of rooftop PV that made it to the New York Times, you typically get this kind of statement that says this project will help the distribution grid and kind of alleviate the stress you might have on the electricity grid in general. And this kind of statement makes a lot of sense because rooftop PV is located, well, directly on your roof, where you do consume electricity. So you do expect that it has to provide some good benefit because it's going to decrease the amount of load you consume. And so if that's actually the case, that's a huge deal because if you think of how much you pay in your electricity rate for the distribution grid, it's a very large amount nationally. So this chart by the EAA shows the capital spending or spending in general in distribution grid in the U.S. for the major utilities. So that's 70% of the load in the U.S. And so what you see is that it has been increasing and the order of magnitude is tens of billions of dollars per year. So if the above statement that when you install rooftop PV, you decrease the need for grid investment is true, that's a huge deal in the sense that even a few person improvement in how much capacity you need in terms of grid capacity can very quickly escalate to, I'd say, a billion dollar or a billion dollar a year in terms of savings for red payers. So that's a big deal. The issue is from an academic perspective, that's when you actually try to assess to which extent the statement is true, so that to which extent distributed generation, such as with the observer, takes pressure off the distribution grid, it's very hard to converge on an answer and it's a very highly debated topic. And if you compare estimate you would get from, let's say, some consulting firm or some project developer to estimate you would get from some papers in literature or from other consultant, the estimated benefits can vary by one or two or three magnitude. So there is no consensus at all whether this statement is true. And what makes this debate complicated is that there is no real, as there is a very big lack of empirical evidence that actually looks at power flow and an actual grid at a large scale and try to really make this assessment of what has been the impact of adding more and more distributed generation to the grid. And so that's where our research comes in is like we managed to get access to very detailed data for the French power system that enables us to tell some interesting thing about this question. So the outline of the talk would be first on background to make sure that everybody has the same concept about power systems and understand what we're talking about. Then I work this with the approach and how it fits well with the data we have. A little more technical discussion on how we tackle our question from an empirical perspective and then result and discussion. I'll save some time at the end for Q&A and discussion. But if there is some pressing question, there is a Q&A feature and I hope Charlie will warn me if I'm missing an important question. So I guess if you get interested into power system at a very high level, the first picture you will encounter is this one or the version of that one. So that one is from the Department of Energy. But the first thing you will learn is the power system is divided into four pieces. You have generation here, transmission that distributes the generation from power plants over large areas. And then once you get closer to load center, you stay voltage down, distribute to a distribution grid here in Green and you reach end consumers. So the reason why you will always find this picture is that and why historically the power grid is organized like this is for the most part that in this piece here on the generation side, there is huge economies of scale of building very large power plants. So since we're going to talk about France, let's take the example of a nuclear reactor in France. So this would be a 3 gigawatt plant. So to give you another magnitude, 3 gigawatt is like 3% of the maximum load ever recorded in France. So that's a huge amount of electricity that won't be consumed locally. You really have to distribute it over large areas. And this is what the turbine looks like. So there are two turbines in this power plant because you have two generators and these things are huge. They actually have a name because that's specifically designed for the station. And to give you an idea, this red cycle here would be a house. So these are very big assets that produce a large amount of electricity. So you need a transmission grid to distribute it over large areas and then you step down the voltage to reach end consumers. So given this organization, power flows are very straightforward. You generate electricity here, you transmit it over the transmission grid, once you reach, so you transmitted high voltage to decrease power losses, but high voltage can be dangerous if you're too close to it. So once you get closer to consumer, you step voltage down and you distribute it to the end consumers. So you get this top down power flows from generation to end consumer, which is very straightforward. And basically all the architecture of the grid has been planned with this mindset and this operation of the power grid. What changes with distributed generation is that even though you do have some economies of scale, they are much smaller. And so if you take the most extreme example of a sort of panel, this would be the power plant itself. I mean, you will need some inverter and a few cables, but this is enough to produce electricity and you can literally hold it in your hands. So this is not this kind of huge scale assets that we historically had with power plants. And because they are smaller, you can connect them directly to distribution grid. So you would put rooftop solar here on the roof of an end consumer, but there are different distributed generation technologies beyond roof solar. So for example, a small wind farm would also be connected to the distribution grid or a larger PV farm also may be connected to the distribution grid. But you also have like small hydro, like a small dam on a river that is a few megawatt will also connect there or small thermal units such as cogeneration or a waste, a facility burning waste, for example, would also be connected there. And so now that you have started injecting some electrons in distribution grid, this very simple picture of power flow going all the way down from generation to end consumers start to be brought a little bit because you have both generators and consumers here. And so what power flow look like here is less obvious than it used to be. So the first, the natural intuition is, okay, power flows are not straightforward, but since you decrease the net consumption, because you generate where consumers are, this has to be good. And even though we don't really know how power flow evolved, this has to be good. The picture is not that straightforward because you don't dispatch distributed generation and for distributed generations such as wind or solar, which are very intermittent. So this is, for example, the output of a random wind farm in France for a couple of weeks. Output is very intermittent and not dispatch. And that might vary across technology. So you don't really know whether when this facility produces electricity, whether when this facility produces electricity, you actually need it. And since power grids are kind of designed to be able to supply electricity in the most stress moment, like these high demand scarcity events, you have no guarantee that it's actually this distributed generation units actually generate when you need them locally. And that's an empirical question. And so the purpose of this work is to tackle this empirical question and to look empirically at what has happened in France and to see whether distributed generation has actually produced electricity when we most needed it. And we look at five type of technologies. We get PV, wind, small light renewable and renewable thermal. We get access to detailed data that I'll describe later. And because the data is detailed enough, we kind of are able to draw statistically robust conclusions. And so in a nutshell, what we find is that we do find very contrasting impact of distributed generation on the need for network expansion, depending on which technology we're looking at. And in particular, we find PV and wind to be pretty much not very helpful during peak there. Contribution in decreasing the highest load level is very small negligible for PV. So we don't expect on average on the population of substation for France a large impact. And we also observe an increase in the variability of the loads as seen by the transmission grid due to the additional PV and wind generation. And if anything, this might actually increase operational costs. So at least from the data we get, we don't find strong evidence that wind and PV have a huge impact decreasing the cost of distribution network. So that's quick. I'll try to work through how we get to those conclusion in the rest of the talk. So what we want to assess is the impact that's adding distributed generation to distribution grid as on average on the need for network expansions. So the best way to talk to this question is to actually look here at this substation that is at the interface between the distribution and transmission grid. And so that's exactly what we do. We kind of focus our attention on distribution substations which are this red cycle here. So if you take real picture and not sketch drawings and you take the French Ecstasy grid which is the case study we are exploring. So this is France. This is the transmission grid which was in blue before. And if you zoom in a given area, you would start seeing the sub transmission grid in purple. And the unit of separation we're looking at would be these dots here which are distributed substation. So those are facilities where you have a couple of transformers to step down voltage from the sub transmission to distribution grid. So this is the point where we observe load. And if you zoom further to the area that this substation is supplying electricity to, you start seeing the distribution grid. And so this distribution grid is of course consists of consumers but also of distributed generation. And so in this case, for example, you would have a small wind farm here that is connected directly to the substation which is here. You would have a wind farm here. And if you zoom down to the low voltage grid, you will also have sort of PV on end consumer houses. And so what we observe is this substation here that aggregates all the consumption and generation from user downstream it. So all this wind generation, solar generation and consumption from these houses and businesses that are here. And the good thing is that we observe over 2,000 substation for friends at an hourly granularity. So we have the hourly load curve for this substation and we observe them over 14 years. So that's the first piece of data we are using. So concretely what it looks like is if you take, for example, a week, you would have the load at this substation for each hour. So that looks like this, for example, for a given one over a given week. So that's a lot of data points. That's over 250 data points, 250 million data points, sorry. And so we kind of structured the pieces of data we are interested in. And so we focus on two fundamental concepts in the power system literature, which are the the low duration curve and the round duration curve, which I'm going to explain right now. I see there is a question that I cannot see it. For some reason I cannot see the Q&A. So first, the low duration curve. So the idea of the low duration curve is a concept that is used quite extensively by power system engineers, because it's very convenient for practitioners. And the idea is very simple. From this row data that I showed you, that is hourly net load levels, what you do is you would sort them, so in that case in increasing order. So I do that for a week, but in practice you would do that for one year or several years. And the reason you would do that is that when you plan for an electricity grid, what you care about ultimately is to be able to supply reliable power most of the time with a very high reliability. And so what that means is you want to plan the grid to be able to reliably supply customers for the high load events, but maybe not all of them, because like planning for everything is costly. So basically what you would choose as a planner is you would choose some reliability special p-hats. And having ranked these hours in increasing order of load enables you to quickly assess how much capacity you need. So here if you say I want the reliability of my grid of p-hat, you can move all the way up to the load duration curve and that tells you how much capacity you need for the grid. And so we can build such a load duration curve for each substation and each year. And that's what we do as a first block of data we use to track how the use of the electricity grid has evolved over time. And to give you some intuition, I'll get back to that later. But so now if you imagine that you used to have a load duration curve, which is a blue curve here, and you add some distributed generation you need that produce a given amount of energy on average. Well, as a grid planner, the impact of this unit is very different depending on when it produces. So if you imagine that it produces predominantly in in hours during which load was already low. So if it shifts the lowest quantiles more, well, you will indeed have a shift downward of the load duration curve with the capacity savings to get here pretty low. If by contrast for the same amount of energy produced, you shift the peak hours more, then you get much, much larger capacity savings. So you really care about when distributed generation produce relative to how high electricity consumption is in different hours for the substation. The second object we look at in this work would be is Audiograms. So it's a more recent focus in the power system literature, which is actually pretty much driven by things we are observing in California. So it's so far it's mostly a system-wide issue in the sense that it's this idea of the California dirt curve. So if you're not familiar with it, it's the idea that once you install a lot of solar PV, you get a lot of protection during the day, but that sunset production decreases and load stays high. So you need to ramp up a lot of gas turbine to make it for that. And you have to make sure that your electricity system actually enables that. So what we are looking at is the same concept, but at the distribution rate level. And so what we do is we take the Audi net level of the substation. We take first differences. So we just take one hour minus the previous one. So for example here, load is increasing, so we have a higher ramp upward. After that, load is decreasing, so we have a low ramp. And by analogy with what we do with the load duration curve, we also compute this, we also keep track for each substation in each share of this ramp duration curve here that gives you the distribution of ramp you have seen for this substation by increasing magnitude. So here, for example, this would be the highest ramp downwards. Here's the lowest ramps upwards and in the middle hours during which load was relatively flat. And so at the distribution rate level, you don't really have this system-wide adequacy issue, because it's at end of the IAS level. But you might imagine that if you start having very volatile flows, that means that your transformer setting will have to change quite frequently. That means that it might be harder to do phase balancing. So I mean, you kind of want to keep track of this metric also at the distribution level to get a sense of how it's evolving in the future. And so what is a good evolution and what is a bad evolution for the ramp duration curve? So again, if you take in blue the ramp duration curve for a given substation that you had historically and you start adding some distributed generation units, what you would like to see is some clockwise rotation of the ramp duration curve because that means that the highest ramps got smaller, both in absolute value, both for the positive and negative ramps. And if you see a rotation in the other direction, that's not a good sign because that means that you basically increase the magnitude of the highest ramps you're seeing locally. So that's what we keep track of in terms of use patterns of the electricity grid. Now we have to combine that with distributed generation. We have to be able to say this ramp duration curve has been observed when there was that amount of distributed generation connected to this substation. The good thing is the universe of power plants in France is publicly available since 2017. So we actually observe each power plant and for most of them we observe them at the individual level. So it's one observation, one plant, and the dataset reports which distributed distribution substation is part of this distributed generation unit connector. So we can, for all the green parts here of these pie charts, we actually observe perfectly to which substation the given distributed generation units connect. So there are also these red and yellow slices here for which we need to do some matching because for privacy reasons the smallest unit, so if you have a rooftop PV on your on your roof, you wouldn't appear as a single observation in this dataset for privacy reasons. So those are aggregated which makes it harder to guess where they connect in the grid. But we can form very reasonable guesses about which substation they are most likely to connect to and we actually run a bunch of sensitivity analysis to make sure that it doesn't affect the results. So I refer you to the paper if you have a question on that. But long story short we are able to observe very granularly the universe of distributed generation units and to which part of the grid they connect. And the last piece of good news is that in France the amount of distributed generation collected to the grid has been increasing quite a lot over the period of our studies so between 2005 and 2018 we've seen a very large increase in the amount of distributed generation to the point where today in 2018 there was roughly 28 gigawatts of distributed generation that's about the quarter of the historical peak load of the country which is both very significant and it's also still pretty far away from the kind of install capacity you would need to get like very very deep penetration of renewable kind of system because the total energy generated by sewage generation is still the small share of the total energy generation. So that's exactly it's not being significant but they are still a long way if you are really willing this technology at a very very large scale as we set some targets for. And so to sum up the final dataset we assemble is this combination of grid usage on the one hand so this low duration curve for and renderation curve for each substation in each year which we can match to how much capacity of each technology was connected to each substation in each year. And so having assembled these datasets what we are interested in is assessing the impact that adding a megawatt of a given technology has had on average on the load curve and renderation curve seen by the substations so if I get back to my example of you have an historical low duration curve you add distributed generation and it shifts down so what we're trying to estimate is basically the difference between these two curves so if you add one megawatt say of PV and it happened to shift on average low duration curve like this the difference between these two low duration curve would be this object here that we collect quantile impact function basically tells you for each quantile of the distribution of hourly net load like how much this content has been shifted by the addition of one megawatt of this technology so in that case it tells you that I think one megawatt has had a large downward impact on the lowest quantile but not so much on the peak hours so you're not very happy if you see a picture like this and by contrast if you estimate a quantile impact function that look like this and that tells you that for one megawatt of distributed generation you get a large shift downwards in the top quantiles but not so much in the bottom quantiles then you're pretty much happy because that tells you that this distributed generation technology elevates the stress on the power grid in the hours where it needs it the most so you do what you do hope to have this this downward sloping quantile impact functions and for rent it's the same intuition it's just what what is interpreted as good news is slightly different so again blue would be the historical rent duration curve and purple how it changes once you have added a one megawatt of a given distributed generation technology and so if you tend to make extreme rent worse and so you shift the duration as a rent duration curve in that direction you end up with an upwards sloping quantile impact function and that's bad because that tells you that the extreme rent both positive and negative are exacerbated by the addition of this megawatt of distributed generation and what you're hoping for is more downward sloping quantile impact function because it tells you that the addition of the distributed generation you need tend to decrease the extreme rent you observe on average at the substations uh so in terms of econometrics how we estimate that is is very is much more straightforward than the equation may feel like but it's basically just we regress we keep track of each quantile for each substation in each year for either the rent duration curve or the duration curve and we regress that on the insult capacity adding a substation fixed effect and near fixed effect to control for trend over time and specific condition at the substation level and the coefficients who are interested are these betas here which tells you the impact the average impact of technology on the q quantile of either the low duration curve if that's what's this q keep track of all the rent duration curve if it's around duration curve and so if you stack up this coefficient for all quantiles for given technology so if you say like for example t is p and you you stack up all this coefficient for the main quantile you kind of rebuild this quantile impact function i show you so that kind of keep tracks of what has been the average impact of adding the distributed generation on average on the load curve or rent duration curve faced by served by substations and so for the impact on the load duration curve what you expect is this coefficient to be between minus one and zero because well they produce electricity so they have to decrease load and they cannot produce more than they install capacity and so if you get a coefficient of say minus two for pv for the 0.5 quantiles it tells you that for the average on the population of substation in France adding one megawatt of pv has decreased the median hourly level at the substation level by 0.2 megawatt hour and you have zero interpretation for the rent duration curve if you expect them to be between minus one and one because again you cannot ramp up more than you cannot induce a highest rent than just shifting from no prediction to full steam production and so if you estimate for example a coefficient of so beta point 75 wind of 0.01 it means that on average over the population of substation in France over the period we look at adding one megawatt of wind as increase a third quartile of the distribution of hourly run by 0.1 megawatt all right so two sets of results one sets on the load duration curve and one set on the rent duration curve so for the load duration curve i'll start with a non-renewable thermal so this think of those as a backup diesel generator or gas fired per generation units and so the quantile impact function we estimate for those is kind of a uniform shaped downward with perhaps some more production during peak hours which is good news for the grid and also makes sense with the fact that that for example gas fired per generation units in France are incentivized to produce more during the winter and that's when peak happened due to a quick heating so this this curve makes a lot of sense and tells you that well adding a non-renewable thermal distribution generation as on average help the distribution grid in terms of how much capacity you need in in the long-term equilibrium if you start adding renewable thermal and small hydro to the picture well the venues is now it's a post-sloping so they tend to decrease the lowest quantile more than the upper quantiles but you still get quite a significant contribution in the peak hours so it does help a little bit it's just it tends to decrease load more off-peak than on-peak and and the real disappointing picture we got from the estimate we we got was for wind and PV where we we got this like very concave and increasing functions where it tells you that if we take PV to start with you get like a negligible statistically zero contribution to decreasing load during peak hours which is not super surprising in France in the sense that we're reaching peak in the winter when it's very cold and in the evening so it's very unlikely that the sun is shining then but I guess what what's more surprising is that it goes pretty deep in the quantile so you really have to get pretty deep in the load duration curve to start seeing an impact and what was more surprising was perhaps for wind because wind generation is higher during the winter so we kind of hope that it would have a higher impact on peak load it does have a statistically significant impact so it does help a little bit decreasing the need for capacity investment but it's it's a very small increase so one megawatt increase will on average yield a 0.004 decrease in the 99s quantile so that that was somewhat disappointing and the other interesting feature we get from this graph is on the other side of the distribution in the lowest quantile where we see this very large impact on the bottom quantiles so tv wind but also small hydro and renewable thermal they tend to to decrease the lowest quantile of the load duration curve which historically didn't matter that much but starts to matter a little bit and will matter more and more and the reason why that is is that as these bottom quantiles decrease so the net load as seen by the transmission grid becomes negative so locally there is more generation than production so you start having backfitting hours so power flow flowing from distribution to the transmission grid and if this phenomenon amplifies at some point this reverse power flow start to be actually higher in absolute values and the peak flow you historically had in the other direction and so they start being a concern in terms of how much capacity you need for the distribution grid because you want to make sure the distribution grid is able to deal with these very high local generation hours and so the substation where you observe this kind of phenomenon historically so the substitution for which the peak usage of the substitution in absolute value was reached during which was reached during an hour where power was going from the distribution to the transmission grid used to be very marginal it was like less than 1% of substation in 2005 and that were like small mountain area substation where you had some idle and very little load but we can see that this has been increasing steadily and as of 2018 you get about 8% of substation where you actually reached their peak usage during the backfitting hour and this trend is very unlikely to stop because if you look at the number of substation for which a negative that could happen at least one hour during the year so there has been at least one hour where the generation was higher than local consumption it has increased also steadily and now about a quarter of substation like this so this feature that you depress the lowest quanta significantly that used not to matter that much is going to start mattering more and more because you would have to either invest in the grid or go tail or store but you have to do something about it the second set of result is on the distribution of audio ramps so again ramps are this variation in net load level over the course of an hour so that kind of capture the variability of the load profile that the given substation is supplying and so what we find for several technologies in small idle is pretty much no impact or there's no statistically significant impact which I don't tell us much what's more surprising again and quite disappointingly is that for wind and PV the addition of more megawatt of wind and PV generation tends to rotate the curve in in the bad direction if you wish but then that it tends to on average amplify the magnitude of audio ramp you see you see locally so for if you take the 99s quanta or first quanta adding one megawatt of wind or solar tends to at least we estimate that it increased on average over the population of substation the extreme the most extreme ramp by point 15 megawatt again that's sometimes for a small level of penetration that that that should be largely manageable the question is like when you start reaching very high level to which extent does it do that affect the operation of the distribution grid and create some need for expenses in terms of operations and to give you a sense of what's in practice so that that's not a demonstration that's pure and a good evidence but if we take the the substation I use as an example to explain what the low duration curve is and what the random duration is so the blue curve I showed you was a given week in June for the substation in 205 and now the substation has added over 20 megawatt of wind and PV and that's what the load was in 2018 for the same week in June and so you kind of reach through this feature that the impact on pick has not been significant and you get this negative hours of network and this very large string in in power flows but that's like just to give an intuition of what are the concrete impacts okay so hopefully the main takeaway that I introduced at the beginning of the talk are a little bit clearer now so the purpose of this work was to assess empirically to which extent different distributed generation technology have decreased the need for network capacity in France on the population average for the country over the course of the past decade and we find like that the impact is very different from for different technologies and in particular for wind and PV we find very little capacity to benefit in the sense that the impact on the top quantiles of the distribution of early load levels so the white part of the deterioration curve is small negligible and at the same time they tend to increase extreme ramps so if anything they might even be more likely to increase the cost of distribution networks if we do nothing about it we don't observe cost hierarchy so we're not saying they do is if anything taken as face value without any mitigation strategy that's what most likely to happen and and we don't find such extreme behavior for the other distributed generation technologies so the the ongoing following work is is obviously on mitigation because what we're finding is is is adverse impact which doesn't mean that the technologies are not desirable per se it does mean that they have this impact and we should deal with it there might be a lot of good reason to install them and they may decrease emissions they may be cheaper they may create jobs I mean we're not we're not saying it's a bad thing to install distributed wind or PV we're just saying that we observe that they create this change in use patterns of the grid and it might be a good idea to get a deeper look at it and so what we are investigating right now is to which extent battery storage can help mitigate that and other that compare for example to other traditional approaches such as just curtailing except generation or trying to get more flexible those these kind of things