AI for Power and Energy Systems

Summary

The Climate Change AI lecture delved into the pivotal role of artificial intelligence in the power and energy sector, highlighting its significance in climate change mitigation, adaptation, and sustainable development. The lecture provided a comprehensive overview of how electric power systems operate, the integration of machine learning in optimizing these systems, and strategies to enhance power grid resilience in the face of climate change. Insights into the socioeconomic influences on power systems, the necessity for reliable data, and the importance of interdisciplinary collaboration were emphasized. Through case studies, the potential of AI in various operational facets of power systems was showcased, emphasizing the need for responsible and equitable deployment of technology.

Highlights

• AI plays a critical role in modernizing power grids and enhancing their efficiency. 🔧
• Machine learning aids in forecasting demand and renewable energy supply, crucial for grid stability. 🌟
• Case studies demonstrate AI's ability to map solar energy and optimize grid operations. 🗺️
• Innovative methods, like reinforcement learning, are being tested to improve grid management. 🧠
• Responsible AI deployment considers data privacy and equitable access to technology. 🔒

Key Takeaways

• AI is revolutionizing power and energy systems, offering solutions for climate change mitigation and adaptation. 💡
• Understanding power systems requires a blend of technology, policy, and societal insights. 🔍
• Machine learning optimizes grid operations, making them efficient and resilient against climate impacts. 💪
• Case studies reveal AI's potential in mapping renewable resources and improving power grid management. 📊
• Interdisciplinary collaboration is crucial for impactful AI deployment in energy systems. 🤝

Overview

The lecture by Climate Change AI explored the intersection of artificial intelligence and energy systems, focusing on AI's capacity to drive advancements in power grids, enhancing their efficiency and resilience. The session outlined the fundamental operations of electric power systems and the strategic application of machine learning to optimize energy distribution, manage demand, and integrate renewable sources efficiently.

Discussions highlighted the necessity for a multidisciplinary approach, melding technology, societal needs, and policy insights to address the challenges of decarbonizing energy systems and ensuring robust grid infrastructure against climate variability. The role of AI in forecasting, grid management, and enhancing operational capabilities of power systems was critically examined through insightful case studies.

Key takeaways from the lecture emphasized the transformative potential of AI in the energy sector, underscoring the importance of equitable and responsible deployment. By utilizing innovative machine learning approaches, the power industry can achieve sustainable energy solutions, ensuring both environmental and socio-economic benefits.

Chapters

• 00:00 - 01:00: Introduction This chapter, titled 'Introduction,' appears to be the opening segment of a larger body of work, likely a book, paper, or presentation. The transcript begins with a greeting, indicating the start of a session or discussion. However, the provided text is incomplete, ending abruptly after announcing that there will be a short wait. This suggests that this chapter may serve to set the context or stage for what follows, possibly covering the purpose, objectives, or agenda of the ensuing material.
• 01:00 - 14:00: Why Care About Climate Change in Power and Energy Systems? The chapter introduces the importance of addressing climate change within power and energy systems. It highlights the impact of climate change on these systems and discusses why stakeholders should prioritize sustainability and resilience in energy infrastructure. The introduction ensures participants are ready to engage with the topic.
• 14:00 - 28:00: Strategies to Decarbonize Energy Systems The chapter begins with an introduction, expressing gratitude for the large audience attending the live session. The speaker is prepared to dive into the topic of decarbonizing energy systems.
• 28:00 - 50:00: How Electric Power Systems Work The chapter titled 'How Electric Power Systems Work' introduces the lecture led by Chris, the lead teaching assistant. The focus is on AI applications in power and energy systems. It highlights Climate Change AI, a global nonprofit committed to promoting significant work blending climate change and machine learning insights. It addresses the audience as summer school students already familiar with these concepts.
• 50:00 - 78:00: Machine Learning Applications in Power Systems The chapter "Machine Learning Applications in Power Systems" started with a discussion on community engagement, encouraging participants to join the community platform, sign up for the newsletter, and engage through the chat and Q&A features on Zoom during lectures. The emphasis was on maintaining an interactive and inclusive environment, although not all questions might be addressed.
• 78:00 - 105:00: Case Studies and Key Considerations for ML in Power Systems The chapter introduces Professor Pryod Donti and Dr. Simone Suo Phobe as key speakers. Professor Pryod Donti is affiliated with MIT, co-founder and chair of Climate Change AI, and has earned a Ph.D. in Computer Science. The chapter likely covers insights and expert perspectives on applying machine learning (ML) in power systems, focusing on real-case studies and important considerations.
• 105:00 - 114:00: Conclusion and Next Steps This chapter highlights the professional backgrounds of notable researchers in the field of AI and public policy. It emphasizes the achievements of a Carnegie Mellon University graduate who was recognized by MIT Technology Review in 2021 as one of the top innovators under 35. Additionally, the chapter introduces Dr. Simone Suop Phobe, an applied research scientist at Microsoft's AI for Good Research Lab, detailing her academic journey through Columbia University and Stanford University, as well as her role at IBM.
• 114:00 - 141:00: Q&A Session The chapter titled 'Q&A Session' seems to take place within a research lab context located in Nairobi, Africa. The session begins with a warm introduction by someone named Chris. More detailed information about the speakers is available both on-screen and online. The session begins with an expression of readiness to start.

AI for Power and Energy Systems Transcription

• 00:00 - 00:30 welcome everyone we're just going to wait a little
• 00:30 - 01:00 a few minutes to make sure everyone else wants to join is able to join
• 01:00 - 01:30 okay great um thank you everyone we have so many people here here joining us live um so is I'll start by introducing
• 01:30 - 02:00 myself my name is Chris um I am going to be the lead ta for this uh today's lecture um on AI for power and energy systems and um so let's just get started um so climate change AI uh hopefully as you all know is global nonprofit working on catalyzing impactful work at the intersection of climate change and machine learning um you as summer school students you already have
• 02:00 - 02:30 um probably seing the website join the the the community platform hopefully sign up for the newsletter so just a few housekeeping items and please use the chat for General comments and ask questions using the Q&A feature on Zoom um we will be looking through these questions throughout the lecture uh and asking them to the to the speakers um yeah as unfortunately we won't be able to necessarily answer all questions but we'll do our best uh to to
• 02:30 - 03:00 do that um and so really excited today to be introducing two amazing speakers professor prian and Dr Simone suo phobe so um yeah so Dr Professor priod donti uh is a professor at MIT um she is a co-founder uh and chair of climate change AI received her PhD in computer
• 03:00 - 03:30 science and public policy from Carnegie melon University and as the recipient of the MIT Technology reviews 20 2021 35 innovators in a 35 award and Dr Simone suop phobe uh is an applied research scientist in the Microsoft AI for good research lab she did her PhD at Columbia University um and also holds a master's degree from Stanford University and also works as a research scientist at the IBM
• 03:30 - 04:00 research Africa lab in Nairobi the more detailed parts of their bios you can read on the screen um and find online and with that uh we'll switch it over to uh our speakers today thank you Chris for this uh warm introduction and uh we'll get we'll get started here if it's okay
• 04:00 - 04:30 um um Can can everyone see my screen okay perfect um hello everyone and thanks for joining us for this uh lecture on Power and edges systems it's my pleasure to be co- lecturing today with a prer and I'll give her a minute
• 04:30 - 05:00 to to say hi before we jump into it yeah hey everyone super excited to be here with you today and just because it didn't get set at the front thanks also to Chris and Ruth our awesome Tas for today yeah um so uh we'll get started with the lecture um I'm sure everyone is well aware we'll be focused on uh Power and energy systems today and the goal for this lecture um is a few things one we want to be able to contextualize what
• 05:00 - 05:30 um climate change mitigation and adaptation means for the power and energy systems sector and be able to call out um why sustainable development strategies are also as important um for this sector when we think about um climate related strategies um to do that though we need to present a background on how electric power systems work um and share with you guys either what are the opportunities or what are some things to keep in mind when applying ml to the power and energy systems um
• 05:30 - 06:00 sector we will then go into some detailed case studies to look at how ml was used and how we should be framing ml related um questions for True impact in this space and finally we want to share from our personal experiences what some potential entry points or next steps might be for you and also hear from you on um how you will take this moving forward so so with that in mind we'll
• 06:00 - 06:30 start off with um why why should we care about uh climate change in the power and energy systems uh sector when we look at emissions um as reported by um the ipcc 2019 um 2022 report um the numbers that we see is that the energy Supply sector contributes about a third of CO2 emissions um as compared to sectors now
• 06:30 - 07:00 depending on how we do the accounting meaning that if we look at say the building sector and the enery used in um building materials that accounting might shift these proportions a little bit but the general idea here is that the energy Supply sector is a huge contributor to the emissions that we see and what do I mean by energy Supply sector this refers to any infrastructure or any equipment that is used used to either extract uh
• 07:00 - 07:30 transport transmit convert energy so that um energy related Services can be provided this can be related to um what we're familiar with which would be Electric Power Systems but it can also um account for foil systems either natural gas cooking fils or even Heating and Cooling um systems that may be familiar with so now we know that this sector is a huge contributor and maybe there are opportunities for us to to
• 07:30 - 08:00 implement a few things that can help mitigate um um climate change impacts for the sector another thing to keep in mind here is when we look at um low carbon sources in the space and we would see that for electric for the electricity sector low carbon sources account for about a third um almost 40% of um energy sources whereas when we look at the total energy sector low carbon um sources are actually still in the minority so we still have a huge
• 08:00 - 08:30 proportion of energy related uh services that are not from low carbon sources and again this is an indication that there might be an opportunity for us to consider what some opportunities are to um increase this proportion um before we move on though it is important for us to call out that low carbon sources are not equivalent to renewable sources yes there overlap but there's a clear distinction between uh both of them and in this particular case this figures are
• 08:30 - 09:00 highlighting at least um the combination of low and renewable energy so again coming back to my initial question why should we care about uh climate change in the power and energy systems um sector first things that we that we observe is that there's an increasing pressure on power energy systems as a result of climate change what that means specifically is that one we are seeing
• 09:00 - 09:30 an increasing number of climate climate related events this can also mean increasing the intensity and the variability of those events we we're experiencing um a lot of more heat spells cold spells um and this can also lead to a change in the demand pattern so for example in the Pacific Northwest where I live um we're experiencing this summer we had very very hot uh temperatures and again this relates to
• 09:30 - 10:00 these two points of one variability in climate related events but then this also shifts the demand for heating and cooling in the Pacific Northwest where we didn't previously have um Heating and Cooling built into um our our um homes the second reason is that the P Energy System sector is transitioning is Shifting and what I mean specifically is that we're shifting from the traditional if I say higher
• 10:00 - 10:30 um carbon sources to low carbon sources and what that means is that these carbon sources say in the case of renewable sources might be weather dependent and with changes in our climate this can actually impact directly impact the amount of energy that can be produced and by consequence supplied um to Consumers again in regions that are also um incorporating more say renewable or low carbon sources where were we're expanding our infrastructure either
• 10:30 - 11:00 reinforcing our existing transmission system or due to Growing growing global population increasing access to Regions that previously didn't have electricity so with this increase in the infrastructure um or the grid infrastructure or power infrastructure we are more susceptible to climate related um events so we need to actually think about how do we make uh the grid as a whole or how we provide Power and energy as a whole more um I
• 11:00 - 11:30 say robust and resilient to um climate R the other point that's not necessarily brought out in the previous slide is that um we've seen that income and economic growth moves in lock step with um electricity consumption this figure here shows if you look at Vietnam in the 90s or South Korea in the '90s and how they've been able to uh grow grow their per capita income with time this has
• 11:30 - 12:00 also moved in lock step with how much electricity has been consumed in these economies so what this might suggest is that income growth is directly tied to electricity usage and while we are um enforcing or or pushing for strategies that mitigate um climate risk in the power and energy sector we need to make sure that we do it in such a way that we're not prohibiting economies from growing we're not prohibiting people from um gaining income and being able to live um decent decent
• 12:00 - 12:30 lives so what um that might mean in terms of uh climate um adaptation and mitigation strategies for emerging economies versus um maybe developed economies the implementation might be different and let's say for industrial economies we would have to maybe focus more on reducing emissions while maintaining the um levels of development that have already been achieved the progress that has already been made but for emerging economies while we're
• 12:30 - 13:00 either maintaining or reducing emissions we still need to push for an increase in the development um levels that are seen in those in those economies um so now that hopefully I've been able to convince you why we should care about it um priia I think will hop in and and talk a bit more about mitigation adaptation absolutely and so um I think suzo gave a really or sorry Simone gave a really nice overview of um where um
• 13:00 - 13:30 why energy systems are so important for mitigation adaptation and sustainable development and so we're going to use this time now to talk through how we actually make progress on these objectives how do we decarbonize Energy Systems how do we make them resilient and robust to climate change but before we tell you we actually are curious to hear what your thoughts are so um I'm going to give everyone a minute or two to maybe submit in the in the zoom chat um and and your thoughts or answers to
• 13:30 - 14:00 this question on the slide which is what are some strategies to decarbonize Energy Systems
• 14:00 - 14:30 I see in the I see in the Q&A there's a question about what's a good source for energy consumption on the country and subcounty level subcount level I assume that's referring to data data sources for that um either pre or someone do you have maybe some suggestions for data sources on that we um pre if I may briefly answer we do
• 14:30 - 15:00 provide a few resources for um data at the end of this lecture while it might not necessarily be specific to that question um but at least maybe that might be a starting point we do provide some resources at the end of the lecture yeah and I'm not 100% certain but my first guesses would be the like International Energy agency or our world in data as potentially to sources that might have it but I'm not 100%
• 15:00 - 15:30 sure awesome so thanks for all the answers coming in in the chat so I see various categories of of actions here that that you know make a lot of sense so one of them is thinking through kind of how we actually Supply Power right make it more renewable stop using fossil fuels and so forth um another has to do with actually the um efficiency of of how we actually provide various sources
• 15:30 - 16:00 of power um and then another has to do with um let's see actually you know changing how much we consume and how we consume via demand response and such and so these are really great these are hitting the nail on the head and what we want to do in the next few slides is provide a little bit of a framework to to actually um kind of conceptualize the different ways that you can actually decarbonize Energy System so uh Simone if you'd mind going to the next slide so um the basic idea is that when we think about decarbonizing Energy
• 16:00 - 16:30 Systems the approach in different countries and different contexts will depend on the specific National circumstances but there are some commonalities and approach that are shared between contexts so uh next point and the a particular conceptual framework that we can use in energy related sectors is all called the Kay identity and this breaks down greenhouse gas emissions basically into the multiplication of a few different terms that each IND indicate strategies that
• 16:30 - 17:00 we actually have to reduce greenhouse gas emissions and if you go to the next point one nice thing about this identity is it actually connects things that you can do on the supply side as you're actually providing power and providing energy to places um and on the demand side so what you can do with respect to for example Transportation buildings industry and other sectors that consume energy and so we're going to walk through this particular um aaya identity term by term next okay so the first first term in the Kaya identity that we're going to break
• 17:00 - 17:30 down is this uh term service per population if you could go next what this refers to is the idea that when we're consuming energy we're not just doing it to consume energy right we're trying to actually get something or do something and the question of how much of that thing how much of that service we consume per person reducing that reducing consumption is one lever that we have to actually reduce greenhouse gas emissions so to make that more concrete next Point um we're going to
• 17:30 - 18:00 use the example of um passenger vehicles so trying to connect some things you might have heard in the transportation lecture to things that you're going to hear in this lecture in this case in the in the case of Passenger cars the service that we might be talking about is the number of vehicle kilometers traveled so how far you actually get to travel in a vehicle in this context next the what reducing consumption might mean is reducing the number of kilometers that are actually driven for
• 18:00 - 18:30 instance I as an individual could move closer to work so that I don't have to drive as many miles or kilometers to work every day and there are systemic changes that can also be made actually changing for example how we build our cities to actually allow people in general to kind of live closer to where they work so that the the the kind of total need for the number of kilometers to drive goes down you can also increase the number of passengers per trip and per vehicle so even if you keep the number of vehicle kilometers driven by a
• 18:30 - 19:00 particular car the same if you divide one car trip by one person versus by two people you can also right reduce this term of service per population and um the kind of General strategies that these imply one is that you can do kind of individual Behavior changes but also and importantly that systemic changes and structural improvements can really change how much we consume per capita so going on to the next term in the Kaya
• 19:00 - 19:30 identity this is the term energy per service so imagine that we've decided we do want to travel a certain number of vehicle kilometers how can we actually reduce the amount of energy that it takes for me to get that again remember in in some sense as a person I don't care about energy I care about service so if I can get the same service for Less energy that's improving efficiency and that's pretty great so in the context of Passenger cars um the uh if
• 19:30 - 20:00 we go to the next Point um there are a few things that can be done here so one is improving vehicle efficiency so you hear about for example k um uh kilometers per liter or miles per gallon fuel efficiency of cars you can actually improve that via kind of technological change you can actually drive more efficiently in ways that improve the number of uh kilometers you're getting per per unit of fuel and you can also switch to other modes of transportation things like bikes which are much less
• 20:00 - 20:30 energy intensive per distance traveled than Passenger cars in general what this kind of leads to is a combination of efficient end use Technologies so the technologies that are actually um providing this service like the car making those more efficient but also more efficient generation technology so if I can actually for instance um when I'm transporting power to your electric vehicle actually reduce the amount of power that's lost along the electrical
• 20:30 - 21:00 lines that's also a way to kind of improve the efficiency here and so the last term um in the Kay identity is um greenhouse gas emissions per energy and this is one that I think a lot of your answers in the chat got at which is that to provide a certain amount of energy we have a choice as to which energy source it comes from does it come from fossil fuels or natural gas does it come from Renewables and basically this lever is by switching to clean energy we can reduce the
• 21:00 - 21:30 greenhouse gas emissions that are associated with each unit of energy that is used and so uh in the context of Passenger cars this can involve strategies like switching to battery electric vehicles that are powered by a clean power grid and of course that's very important you need to Green the grid alongside switching to battery electric vehicles for this to work um or another option is to switch to Alternative liquid fuels like electrofuels solar fuels or hydrogen fuels that are produced in a lower carbon Manner and that uh emit fewer
• 21:30 - 22:00 emissions when actually burned and the general strategies implied by this are a combination of electrifying and switching to lower carbon power so move to electricity and make the electricity cleaner as well as replacing fossil fuels with clean alternative fuels and um kind of this is a in some sense a summary a conceptual summary of things that you can think about doing in energy related factors again reducing consumption improving efficiency and
• 22:00 - 22:30 switching to clean energy but in looking at each individual component we also shouldn't lose track of the larger picture which is ways we can drive reductions in greenhouse gas emissions directly and so um going to the next point we also have access to economy-wide strategies to reduce total greenhouse gas emissions and that involves things like targets regulation standards and Investments across the the economy strategies like carbon pricing which directly try to um try to to cap
• 22:30 - 23:00 or to tax uh the amount of greenhouse gas emissions that a particular entity um might might produce and then they might take some of the levers implied by the right side of the equation to to try to actually make that happen and of course some carbon dioxide removal as well um so far um PR has been able to call out how um we can
• 23:00 - 23:30 decarbonize um but if we can take a minute to having considered decarbonization how can we adapt um Power Energy Systems to climate change and please feel free to to drop your thoughts in the chat maybe we can take a a minute or so to do that um what would be some strategies to adapt our current um systems to to climate change
• 23:30 - 24:00 and while folks are dropping uh questions in the chat um so there's a really great question from an attendee about you know the Kaya identity um has
• 24:00 - 24:30 a a population term uh and so this is basically using a mean value for each person um and of course different people around the world consume different amounts so the way I would take the Kaya identity is yeah so in some sense if you were to use it to actually compute greenhouse gas emissions you would have to use averages across the the terms across the world um I think of it more as a conceptual identity which is to understand in any given context what are the different kinds of things you can do um and of course um as also emphasized
• 24:30 - 25:00 in the kind of intro on sustainable development the strategies in different places will need to differ significantly people in developed contexts consume much more than those in developing contexts and as a result we shouldn't for example treat the reduce consumption term equally um across all contexts so um that's a great question and yeah I would think of it more as a conceptual way to break down the strategies rather than something that we use to literally calculate the greenhouse gas emissions in reality
• 25:00 - 25:30 thanks thanks for that um addition to prayer um I see that the chat is quite buoyant but a few things that uh stand out to me here are around grid modernization um I also see some some points around improving resilience or increasing resilience um in infrastructure and assets um I see some um comments around around uh better
• 25:30 - 26:00 forecasting um all around all of which are touching on on very import important strategies for for adaptation so um starting off with the first one that I want to call out is increasing robustness and resilience and for those that made this point U maybe um they were reading our minds a little bit but yes this a a a huge point and the difference here is that robustness here is the ability for the grid to handle um events while resilience would
• 26:00 - 26:30 be the ability for the grd to um regain um operation after such such an event so here for example um when we say being able to to be robust means accommodating correlated failures this can be in the case of extreme weather events how well are we able to handle um such situations
• 26:30 - 27:00 and resilience here would be can we quickly repair um the infrastructure that is down and come back after after such events in reality um I think both of them play well with each other where we have this uh somewhat cyclical process where we uh prepare and build robustness into the grid such that during the event we can handle it as best as possible and then in enabling quick recovery and learning
• 27:00 - 27:30 from those um um observations so that we can be better prepared against the next um event another way to is um accommodating changes in in supply and demand yes we do anticipate that um changes in whether will impact um energy production or impact consumption but um are we able to better handle those changes in demand um that might occur due to extreme Heating and um do so with
• 27:30 - 28:00 low carbon um sources um the third Point here is around building adaptive um capacity and what we mean is that um while we increase um access and reliability as we've seen these are strong drivers for economic development and can Empower people to be able to better respond or better adapt to changes in climate so while we push for um Solutions or approaches that directly impact the grid
• 28:00 - 28:30 also ensuring that we provide um energy or electricity at levels that people can actually use to to gain economic independence is is also as in as important um having provided that context I'll hand over to prayer yeah so so far we've talked about some in some sense macro level strategies to to really think about how you um decarbonize an energy system or how you um adapt it to climate change
• 28:30 - 29:00 but in order to really truly understand the nitty-gritty of these strategies we we actually want to do a little bit of Deep dive into specifically how Electric Power Systems again remember one type of energy system but an important one um how Electric Power Systems actually work and so um if we go to the next slide and we're going to um you'll notice a bit of a pattern so we're going to ask you about things before we actually tell you about them so um the the question here if you could answer it
• 29:00 - 29:30 in the chat is what is an electric power system as people start typing answers in the chat I think this would be an appropriate time to bring up a question from the Q&A which is which comes from Ali mhud how can we integrate smart grids with complicated Energy Systems in third world countries
• 29:30 - 30:00 yeah this is a a great question and I can take a little bit of a stab and then Simone if you also have some thoughts um in some ways actually um it can sometimes be harder to integrate smart smart grid Technologies into developed country infrastructure than in developing countries depending on whether you're thinking about retrofitting existing systems versus building out new ones so for instance if you are kind of creating a new energy management system or creating a new
• 30:00 - 30:30 micro grid you have the ability to integrate sort of the latest and greatest Technologies in terms of the software systems that are used to manage power grids as well as the sensing Technologies in a way that retrofitting doesn't necessarily allow now technology costs can be more expensive with some of these installations and that can create a lot of inequities and so kind of um uh kind of uh Equitable financing and and kind of um Finance transfers become a really important part of this so for from a financial perspective there's a
• 30:30 - 31:00 lot that needs to be done to basically redistribute um financing to enable this but from a technical perspective actually in some ways um uh kind of outfitting new or newer infrastructure with um Smart Technologies can actually even sometimes be easier than retrofitting existing Technologies um I do Echo um PRS perspective I think the other thing we we want to keep in mind is um what are some of the the questions or challenges
• 31:00 - 31:30 that we want to be able to solve when um implementing smart B Technologies in emerging economies I think that can also really influence the design of such systems um with the goal of uh easier um integration but also reducing the cost um for such devices so um I think I would try to approach it in a more um holistic way in a sense that if we could know a prior what we want to be able to
• 31:30 - 32:00 answer with such Technologies then we can design um technologies that better suit the environment rather than see picking up um technologies that have already been predesigned I might not necessarily be suited for the Contex someone there also look to be two questions in the chat about um climate change adaptation and disaster risk risk management I don't know if it is makes sense to take them now before we move on to this next section um sure if you don't mind calling out the question or
• 32:00 - 32:30 not in the chat sorry in the Q&A so one of them is how much can forewarning actually mitigate the impact of disasters and the other one is um extreme weather causes lots of disruptions do these strategies take into consideration the history of predictive models of extreme weather patterns ac across the globe oh and you're muted yeah um I think with the first um this the first one I see is around extreme weather causing a lot of disruptions
• 32:30 - 33:00 does does this take into account predictive models I think we do need better predictive models for um weather and also tying it to what the impacts are going to be either on like um generation and I think we do have a case study actually that talks a little bit about that how do we move from um weather um models to what that might mean for planning in Power Systems so yes um that is is a strategy and should be taken into account um what we may not
• 33:00 - 33:30 have necessarily discussed are um how to frame that question understanding where the boundaries might be and how we actually use it for um impact within the space um the next question is how much can for warning actually mitigate impact of disasters um so I have some thoughts yeah you it depends on the nature of the
• 33:30 - 34:00 disaster and it depends on the amount of forewarning so for example if you have a sense that more likely certain parts of the infrastructure over the next 10 years then you can actually build up and retrofit your infrastructure to be more resilient to the impact of hurricanes um by kind of um enfor reinforcing certain lines for example in order to then kind of make sure that that part of the system is stronger when it comes to kind of nearer term warnings uh there are
• 34:00 - 34:30 again some cases where forewarning can help uh not a climate change related example but um in India um during the covid-19 pandemic there was a moment in time when the government said hey can everyone shut off their lights at this specific time of the day and sort of step outside and clap to um give um uh thanks to all the healthcare workers who are um you know working very hard in the context of the covid-19 pandemic um and that uh action you know caused or could
• 34:30 - 35:00 have caused a very sudden dip in electricity consumption and if a power grid operator has to look at that and see that in real time and scramble in real time to figure out what to do that's very different than if they have a little bit of warning ahead of time to tell different power generators hey um you actually should probably ramp down or shut off or start to shut off your power um because we will have a dip in generation at this particular point in time so are strategies you can take both on kind of planning Horizons and in
• 35:00 - 35:30 real-time operations Horizons based on warning not everything can be fixed with a warning but some things definitely can awesome so I'm gonna move us on to uh this particular section so thanks to everybody who contributed into the chat about what an electric power system is and um I think a lot of you have some very um great ingredients in there so I'm just going to move on to to the next slide um and kind of spell out our conception so when we talk about
• 35:30 - 36:00 Electric Power Systems we're talking about the systems through which uh first uh click um our electricity is um produced um and then if you don't mind clicking Simone um and that that's our electricity generation infrastructure that electricity generation is then transported along lines and electrical equipment um to where it's consumed by endus consumers like us and in the middle there this transmission and distribution portion um
• 36:00 - 36:30 transmission refers to the part of the electrical system that is transporting power at very high voltages very long distances and distribution refers to the part of the grid that is transporting power shorter distances at lower voltage an analogy to think about why this is the case is to think about Road Transportation where often we have a combination of Highways and Service Roads you take a highway in order to go a very long distance um very quickly but
• 36:30 - 37:00 if we always went that speed near where we lived a lot of people would get hit so we don't want to do that and so nearer to the final destination where we want to go we exit from highways onto Service Roads which are roads where we go more slowly in some ways less efficiently in order to get that last mile or last few miles to our final destination in a way that's a bit safer and so transmission distribution are similar high voltage is less safe but is
• 37:00 - 37:30 much more efficient for transporting power and so our transmission systems take that large um take power that large distance very efficiently Distribution Systems take it that last stage a little less efficiently and while that last slide presents in some sense a one-way idea of how Power Systems work um from get power flowing from generation to consumption um Electric Power Systems are are rapidly changing in various ways so
• 37:30 - 38:00 first um we're starting to see bidirectional Power flows which means that we don't just have big power generators sending power towards consumers we also have things like rooftop solar panels on um on homes or batteries that are connected to the distribution system that are actually shooting power up the other direction and so we're starting to see a power grid that's operating not in a one directional power flow but using by directional power flows we're also seeing whereas before
• 38:00 - 38:30 power grids were often operated in a centralized way where we had power grid operators or a market dictating how large power generators would actually produce power we're starting to move to a lot of non-centralized control paradigms so for example before electricity demand was viewed as inflexible it's a given the power grid has to deal with it now we're seeing strategies like demand response where electricity consumers can actually
• 38:30 - 39:00 change when exactly they're using power based on the needs of the grid and that is in some sense a form of of decentralized or distributed control and there are actually two different paradigms for uh non-centralized control distributed control basically says we have many different devices that are trying to mirror what would happen centrally so maybe we tell them what to do but we tell them what to do um we give them some information to act in a way that's coordinated decentralized
• 39:00 - 39:30 control basically says all of the devices kind of independently decide what they want to do and we might send things like Crisis through a power Market to incentivize them to have certain Behavior But ultimately each device is going to decide independently and based on its own kind of self-interest um what to do so there's some different paradigms there um and then we also have a distinction between on versus off- Grid setups so for example in Islands or rural areas that are very far away from a centralized
• 39:30 - 40:00 power grid it doesn't always make sense to um try to kind of put connections from a big centralized power grid to a farther away place because it can lead to certain inefficiencies and um as you transport power a very far distances so um for instance some places might have off-grid Micro grids which are smaller power grids that are independent um in in certain cases or there might be smaller grids that are semi-connected to to a larger grid and also in some cases homes might have personal power sources like a solar panel and a
• 40:00 - 40:30 battery and then um last but not least um you know given that we're in the climate change AI summer school many people may also have on their minds the impact that Ai and data infrastructure are also having on Electric Power Systems in particular we're seeing due to kind of recent advancements in and increases in use of of of generative AI technologies that there's been a lot lot of build outs of data centers and infrastructure in places where this wasn't previously anticipated and that's
• 40:30 - 41:00 led to things like um extensions of the lifetime of coal plants that were slated to retire now this is a situation where planning is helpful right if you know uh how your electricity load might change due to AI or due to other factors you can kind of plan to service that using renewable energy we also should make AI more efficient um but in the near term at the very least this is having an impact because in some S Electric utilities aren't prepared to deal with the the growth associated with this and
• 41:00 - 41:30 so this is something we should keep in mind as well so when thinking about um kind of how a power grid works it's um important to understand that power grids are physical systems and so there are various kinds of physics and constraints associated with power grids that need to be satisfied at every single moment that you're operating a power grid so for example um at a high level there always
• 41:30 - 42:00 has to be an exact balance between the amount of power that is produced and put into the power grid and the amount of power that is consumed and so at every single moment in time you have to maintain this exact balance between power input and power consumption now we can kind of um go even a step further and write down um kind of a set of mathematical equations that character ize uh in some sense the the the key goals we have when actually
• 42:00 - 42:30 operating a power grid and we're going to use a an optimization problem called AC optimal power flow as our way of kind of showing in some sense what it would look like to ideally operate a power grid so the idea is that in an AC optimal power flow style setup we have a power grid operator that's going to tell all of the different controllable power generators on the grid what their power and voltage set points should be and the
• 42:30 - 43:00 goal um is to do this in order to next Bull it um meet power consumption so again kind of power production and power consumption need to match and when we say power consumption that means a combination of true power consumption so for example the um the load coming from you know lighting and Heating and Cooling and such minus the power that's lost along the lines as it's transported and also Min um accounting for for example rooftop solar which we often
• 43:00 - 43:30 think of if you're a centralized power grid operator rooftop solar production is often thought of as is negative consumption from the perspective of the centralized operator in addition you want to make try to minimize costs so minimize the cost of providing power and you also want to satisfy various constraints associated with um how power flows along the electrical grid um and how you how much you can change power consumption so um this is characterized by again an
• 43:30 - 44:00 optimization problem called AC optimal power flow which is on the slide and so the kind of terms in this optimization problem have to do with the first term is the objective which is minimizing the cost of power generation the next term has to do with various equality constraints which I'm going to skip over the third term has to do with the fact that um power generators have limits in terms of the minimum and maximum amount of power they can produce
• 44:00 - 44:30 and also lines have limits in terms of how much power they can actually transport so we have bounds on these quantities codified within our problem and then the last term is our power flow constraint which basically uh dictates that the amount of power flowing into any node on the power grid has to equal the amount of power flowing out of the node on the power grid and this ends up being a function of power generation power demand voltages and the resistances or admittances of the
• 44:30 - 45:00 lines and then for those who are familiar with optimization and the concept of Duality which I realize is not everybody so don't worry about that but for those who are curious the Dual variables on this last equation actually are the prices of power on the power grid so there's a relationship between this optimization problem and what you would see in a power market so that provides in some sense an ideal ized representation of how power grids work um but um in in reality power
• 45:00 - 45:30 grids are you know more complicated and so I'm going to give maybe a minute to folks to to maybe drop some thoughts in the poll what are some ways in which operating a power grid might actually be more complicated than what we just showed um in the meantime uh there's a great question in the Q&A from ma
• 45:30 - 46:00 matthus are there any data sets or collections that would be recommended for someone jumping into energy forecasting be it grid graphs asset specification generation or consumption Etc yeah so um some data sets that exist are um for example in the US context the um major um uh independent system operators or Regional transmission operators like pjm kaiso Etc will
• 46:00 - 46:30 publish aggregated information about electricity supply and demand when it comes to um asset specifications in the US context there is some information from the Environmental Protection Agency that tells you the amount of Power Generation Um at an hourly level from any um fossil fueled units above I think 25 megawatts um so you can get some granular data there grid structure specifications tend to be very protective um because um of uh safety or Market
• 46:30 - 47:00 reasons um so even in the US which I'd say has some of the more open grid data um in the world um grid topologies and grid asset specifications tend to be hidden and so we tend to have to use synthetic test cases provided by entities like itle e um I'm not as familiar unfortunately with the data landscape in other places around the world um but um it uh unfortunately I think is not even as open as that um
• 47:00 - 47:30 elsewhere thanks Bri all right so thanks for the the answers in the chat so there are lots of really great answers in there um that um I'm not going to get to kind of summarize or read through all of but just if we go to the next slide I'm going to throw up a bunch of text and basically indicate a few ways in which reality might be more complicated many of which are captured in the chat so one of them is that the
• 47:30 - 48:00 optimization problem I showed AC optimal power flow is actually kind of expensive to solve it's a it's a non-convex um problem with nonlinear constraints and so power grid operator use approximations in practice um we showed what happens when you schedule power at one point in time but you actually have to schedule power over multiple points in time so you have to decide which power generators to turn on and off and also how much you're going to change that power generation from from one moment to the next um many quantities on power grids are uncertain so you don't
• 48:00 - 48:30 have a perfect picture ahead of time of electricity demand or electricity uh generation and so you both have to make those predictions and also make real-time adjustments based on what actually happens on the system and there are all sorts of you know outages more nuanced physics that happen on the power grid as well as um some nuances with respect of power pricing the um uh location the the prices the Dual variables on the power flow constraint that I showed before those are in some sense wholesale prices um but for
• 48:30 - 49:00 example consumers face retail prices some power generators are compensated outside of the market to incentivize them to keep running and provide kind of capacity to the Grid in the case of extreme events um and there's a lot of power that's not actually procured through markets a lot of power um more than 90% is procured through Power purchase agreements which might be bilateral agreements between a generator and a large um consumer so um while we H yeah while hopefully the idealized
• 49:00 - 49:30 picture provides a little bit of a mental model as to how grids work um of course when diving into this topic it's important to understand the ways in which those models might be slightly incorrect or slightly limited um so I think uh one of the last things I want to talk about before I I hand it over to to Simone is while we've talked about power grids as a technical system in real reality they're a sociopolitical technical system which is
• 49:30 - 50:00 to say there are lots of entities and lots of policies that actually govern how power grids work so first of all um Power grids are quote unquote natural monopolies because power grid infrastructure tends to be in some sense the it's hard to build two sets of power GD infrastructure in one place there tends to be one set of power GD infrastructure one entity or set of entities that manages it and as a result these create natural monopolies where it's it's not kind of that many competitive entities can come in to manage the system and as a result there
• 50:00 - 50:30 these grids tend to be managed either by public entities or by private entities but entities that have really really tight regulation on them in addition there are lots of different stakeholders that come into play regulatory commissions that are setting highlevel requirements and parameters system operators who are responsible for maintaining lowcost and reliability utilities who might actually own or manage the actual physical equipment suppliers demand aggregators who are working with multiple consumers to help
• 50:30 - 51:00 Aggregate and make their demand more flexible and consumers and and others who are are participating in the market so again understanding the stakeholder landscape becomes quite important and then some other considerations to keep in mind one is that different Power grids in different parts of the world um have different assumptions um if you could click to the next bullet um on whether consumers get 247 reliable power so um in for example the us there is an assumption that whenever I flip on my switch I should be
• 51:00 - 51:30 able to turn my light on whereas for example in some parts of India there are scheduled and regular power outages that are um communicated potentially ahead of time and potentially not but there isn't necessarily always that same assumption that power is always available um in addition um Power um utilities and system operators are incentivized financially in different ways around the world to um take or not take actions to modernize the grid so for example many entities um are the way
• 51:30 - 52:00 they make money is through something called a regulated rate of return which is for every dollar or unit of currency that they spend on a physical grid infrastructure Improvement they get some kind of rate or percentage back but in general around the world this tends to apply to physical improvements like putting infrastructure in the ground it doesn't tend to apply to Software related improvements and so actually financially there often is not an
• 52:00 - 52:30 incentive um among power grid operators to modernize their infrastructure from a pure Financial perspective although of course as grids get more complicated there's an implicit incentive that a lot of this more fancy software is needed to deal with that so with that um we've kind of spent most of this time giving an overview of you know Power Systems and energy systems in general um and I'm now gonna um kind of flip over and I forgot I'm I guess not handing it over I'm continuing
• 52:30 - 53:00 on but we're going to flip over to to doing an overview of where machine learning actually plays a role here so we're going to give a really uh quick overview of various categories where machine learning plays a role across operations planning Innovation policy and markets and data management so first diving into power grid operations and we can go to the next next so one category of ways that machine learning plays a role in power
• 53:00 - 53:30 grids is by helping us to actually understand and assess the state of the power system as it stands so this might be the current state of the system so Based on data I bring in what are the voltages what is the structure of my power grid I might actually not know this or are there any outages on the power grid and we also might predict something about the future state so prediction of future renewable energy Supply demand emissions and so forth and where machine learning uh can play a role um is in actually helping to to
• 53:30 - 54:00 estimate these quantities where the kind of pro of machine learning um against other methods is that it's fast it can use many different types of data you know sensor data image data and learn correlations between them and as a result it's actually been a really powerful tool for near-term and real-time predictions the challenge though is that machine learning methods tend to require consistent data so that I have you know similar data coming from my sensors at every point in time and in
• 54:00 - 54:30 situations where sensors are noisy where there might be missing measurements that's sometimes difficult um and sometimes machine learning methods also struggle with longer term Trends because they are learning relationships in present data that may not always generalize to Future data so in those cases rule-based systems physics based systems optimization statistics can play an important role as well in terms of a specific case study um one example is the uh example of electricity now casting um so for
• 54:30 - 55:00 example the nonprofit open climate fix worked with the UK power grid operator National Grid ESO um to improve their electricity supply and demand forecasts by by a few few times in terms of reduction of um error um and they did this through a combination of more advanced deep learning architectures as well as well as by cleverly combining multiple sources of data like time series data satellite data and the predictions from physical weather
• 55:00 - 55:30 models so okay so we know the state of the system now how do we actually manage the power grid so um one kind of strategy we have is to actually better um handle how we actually schedule controllable power generation in a centralized uh power grid so recall this problem of AC optimal power flow that we just talked about and the goal is that by if we're able to solve AC optimal
• 55:30 - 56:00 power flow or other problems to manage power grids in real time um and at larger scale that allows us to better integrate time varying Renewables which since those vary over time having realtime Solutions is helpful as well as improving the robustness of the power grid and reducing waste um the challenge is that existing solvers for AC optimal power flow are really slow and so they don't sell fast enough to deal with the real nature Renewables and they don't solve at large enough scale to deal with the fact that we have many more
• 56:00 - 56:30 Renewables and distributed devices coming onto the power grid and so the approaches that are used are you know a combination of optimization and machine learning and there are various approaches that both try to directly speed up the AC OPN normal Power flow problem using machine learning um and you can go ahead to the next bullet um and so that's things like predicting constraints in AC optimal power flow that might be redundant or learning better starting points for optimization solvers or even learning full end to-end approximations for AC
• 56:30 - 57:00 optimal power flow and the tutorial um for for this week on Power Systems will actually deal with this first bullet of machine learning for speeding up acopf and there are also uh various techniques that try to use reinforcement learning um instead of solving the centralized optimization problem in a standard way actually use reinforcement learning to suggest what actions to take on the power grid in addition I to optimizing power grids in a centralized way power grids can also as we mentioned before right be
• 57:00 - 57:30 operated in a distributed or decentralized way and machine learning can play a role in helping to actually control those D distributed resources like solar inverters like batteries like flexible loads um where the need is that we need um control strategies that are you know fact flexible scalable robust physically feasible so they have to work well and they have to satisfy the constraints of the engineering system and so the approaches that are taken are
• 57:30 - 58:00 often a combination of machine learning and control theory where machine learning can often come up with very powerful and well performing strategies for um distributed control but where the challenge is that machine learning techniques don't tend to come with robustness guarantees or provable guarantees which if you don't have those things can mean that you black out your power grid so there have been a lot of works that try to combine machine learning and control theory in order to deal with this and so the general idea
• 58:00 - 58:30 is that if you have um you know uh kind of some control theoretic uh uh controllers that have some provable guarantees and you have machine learning methods that can learn from data then you can do things like take the outputs of the robust control model and the Machine learning model and choose which one you're going to use at a given point in time or mix between them um or you can even embed control theoretic constraints into the design of your
• 58:30 - 59:00 reinforcement learning models in the first place one thing we forgot to mention is on these slides there are some citations in this kind of box letter year format and so um the slides will be shared with you after and there's a references section at the end in case you want to follow up on any of the citations in here and um kind of the last uh category of of applications I want to cover is predictive maintenance and efficiency Improvement so this idea that when we're operating a power system sometimes things break sometimes things are
• 59:00 - 59:30 inefficient and if we can detect those inefficiencies or outages ahead of time or in real time then that can help us operate the system more efficiently and so um machine learning is used in this context alongside um to to analyze large amounts of data on the power grid um in order to figure out where there might be efficiencies um in addition to approaches like manually inspecting the power grid and tradition signal processing approaches and so some examples here include detecting methane
• 59:30 - 60:00 leaks in natural gas infrastructure so as natural gas is transported in pipelines and in compressor stations um from where it's extracted to where it's consumed um there have been efforts to use a combination of satellite imagery and even sensors outfitted on the natural gas infrastructure in order to actually detect where there might be anomalies or where you might actually see methane leaking um using the satellite imagery in addition there are lots of applications that similarly try to detect anomalies in solar power
• 60:00 - 60:30 production wind turbines batteries or also identify places where there might be theft um of electricity on the power grid by basically analyzing and trying to find anomalies and patterns in the underlying data so having talked through machine learning for operations in power grids I'm going to hand it over to Simone now finally to H talk about some of the other ways that machine learning plays a role um so the next um area that we'll be
• 60:30 - 61:00 looking at is the planning area um though subsequently will cover um Innovations policy and uh data management so the goal with uh planning in um par systems here is either the design or construction of new infrastructure or reinforcing the existing backbone it could be transmission backbone distribution um backbone so here with our planning you want to be able to incorporate low carbon sources
• 61:00 - 61:30 like I mentioned reinforce your existing connections and ensure that you have um High reliability in terms of service provision for the consumers and some ways that planning is usually done or some um approaches that I use for the different uh dimensions of planning can be uh multiobjective optimization so where do you place your resources where do you place your assets um how do you um design the lines where should the lines actually go depending on where your your customers are um it could also
• 61:30 - 62:00 touch on um estimating demand so doing manual surveying to see how much demand might be in the area uh it can also mean consensus building in sense what systems do um the consumers actually want to have is it a micro grid is it uh a centralized Grid in the system so there are multiple dimensions of planning and some examples that we see that um uh machine learning has been able to support has been one in helping um understand grid topology so mapping out
• 62:00 - 62:30 um lines and um solar wind infrastructure to understand how you can provide services from um renewable energy um the other aspect has been in demand estimation especially for new consumers and capturing the uncertainty that might occur when estimating the potential demand that a new customer might have um on the grid um another application area that we see um gaining some um traction for machine
• 62:30 - 63:00 learning applications is in the Innovation space and in by Innovations what we mean is either in the development of lowc carbon um energy sources or more effectively um improving um energy storage and efficiency or um sequestration of um greenhouse gases um here in the case of uh of uh CO2 sequestration even energy storage we
• 63:00 - 63:30 find that um typical approaches are for like human guided experiments sometimes in a wet lab sometimes through simulation but we're seeing machine learning increasingly being used to help discover um new viable um battery chemistries for example that can um lead to um fewer experiments being run and finding the right um set of of chemistries that lead to um better battery Behavior battery performance there's also some work that
• 63:30 - 64:00 we've seen being done um where for uh nuclear fusion uh reactions um sometimes the uh generation of plasma can cause uh disruptions due to just the inci in the plasma so there's work being done in predicting where disruptions might occur in your reactor as disruptions can cause damage to your equipment so they want to be able to predict where those disruptions might current deploy uh disruption Management Systems ahead of time to prevent those disruptions from
• 64:00 - 64:30 um kicking in so that's another dimension where uh machine learning has also been called on for Innovation purposes um another aspect also in terms of um the power grid has been in how we design um policies regulations and also how we monitor the impact of these policies um on markets and um uh any Downstream from there uh this has typically taken place
• 64:30 - 65:00 a form of either policy analysis uh Market studies um doing Market design but um now we're seeing um machine learning being increasingly called to supplement uh some of these approaches so uh there's been some work in understanding um how um Trends in uh the solar um Power space have shifted even analyzing um the patterns and this actually reveals gaps of where um the market
• 65:00 - 65:30 could either go or where technology didn't necessarily um do well in addressing some of the challenges that we face so this can influence how we design say solar policies or um solar power plant um policies that can improve um adoption of low carbon low carbon sources um another area too has been in leveraging reinforcement learning to set Market prices um here there's um some work
• 65:30 - 66:00 where um they use reinforcement learning and deep neural networks to determine the right prices that could reduce the pick to average ratio and understanding how um different um either micro GDs would respond to different prices it's a it's a good um way to understand the impact of different um prices on um the market and how to adjust the policies based off of um Behavior um another error too is in uh data management we're increasingly um
• 66:00 - 66:30 observing that utilities energy providers are handling more and more data either through the deployment of smart meters and and um better compute that's available uh but not as much um work has been uh put forth to Advan um data usability so we're seeing U machine learning being called to facilitate that cleaning compression condensation and processing um as opposed to doing manual
• 66:30 - 67:00 cleaning or the traditional say data compression strategies um ml is being used for example in the case by um calist Cooperative where they do uh record matching between data sets um on us electricy data safe work and eia um that are usually hard to match um independently so these are posted um independently but can be hard even though they might be referring to the
• 67:00 - 67:30 same power plant the same um generators um it's hard to match them so they actually develop a model that can help match them so that there's more cohesive view of the of the landscape um so we've tried to give you a brief run through on like the different areas within um uh P systems where machine learning um is being utilized but we we will go into some case studies a little bit later on in
• 67:30 - 68:00 this presentation um for now we want to take a step back and and ask you um what do you guys think are some important things that we need to keep in mind when we use uh machine learning in the end Power Systems what are some things that are um important for us to either implement or consider and feel free to to um drop your answers in the chat
• 68:00 - 68:30 just wanted to raise a question that's posed in the Q&A um this comes from Jeff soich which asks what is the typical reaction time that is needed in optimizing power flow for example how quickly does a generation facility need to be able to reduce generation in order to avoid the dangerous imbalance so this is a great question and the maybe unsatisfying act answer is it depends so to give a little more
• 68:30 - 69:00 qualification on that um in uh the us we we operate the power grid at a frequency of about 60 hertz and so if your your frequency changes based on how much Supply demand IM balance you have so if more Supply than demand your frequency is higher if you have more demand then Supply your frequency is lower um and if your frequency is off off by kind of a a you know a fraction of a Hertz or you know a 100th or something like that then
• 69:00 - 69:30 you have a couple of seconds um to get that back and they're often um kind of automatic um uh mechanisms installed in power plants to detect when there's a small mismatch in the frequency where each power generator has then like a little bit of a a fractional responsibility to help make up that difference um but if you have you know a very large um frequency difference like you know it hopefully never gets this
• 69:30 - 70:00 extreme but you know 55 Herz on a 60 HZ system you definitely don't have a second you you have much much much less than that and one thing that's challenging is that um there's a concept called system inertia which comes from the fact that conventional power generators have spinning components that are synchronized to the frequency of the power grid and if there's a mismatch in supply and demand it actually takes a little bit of time for those spinning components to speed up or slow down um whereas with um you know solar power um
• 70:00 - 70:30 and kind of other sources that don't have that that spinning component um they're not providing that buffer to the system and that means that in in systems with larger amounts of um kind of non-spinning power generation mismatches in supply and demand can lead to much faster swings in the frequency of the system so what we're actually seeing is that as we move towards more renewable energy based systems the amount of time we have to respond is even less and that's driving some of the the need for kind of automation because we're now
• 70:30 - 71:00 starting to talk truly subse scales where it's really hard for a human to respond that quickly um I think we have another question in the chat which says Beyond AI or perhaps combination are there any traditional application of network science and graph Theory and energy um space um I think for um planning there's quite
• 71:00 - 71:30 a bit of graph the that might be called up on there especially on how you do connections between um say medium your your distribution and transmission um infrastructure how do you connect um households actually how do you connect low voltage lines from households to your substations your substations um the feeders all of that um in an optimal way where you are minimizing um the cost and not just going um as the as the crow
• 71:30 - 72:00 flies so I've seen quite some work in terms of like graph theory for that um type of research um PR feel free to add to before you no that sounds great feel free to keep going good um also I given the comments in the chat I think a lot of um comments here align with some of our thoughts um one is around um the data reliability data access integration with existing
• 72:00 - 72:30 system data quality um domain expertise which is a huge Point respecting physical constraints cyber security all of these are are really great considerations um glad you guys have been able to um call those out um and to that point um what we're trying to get at here is that um grids electricity grids actually have um physical laws that they have to to meet
• 72:30 - 73:00 meaning in a simplest form it could be matching supply and demand at every point in time like priia mentioned um it could be respecting um the limits of the lines so they actually um operational and physical laws that have to be respected and even how like electrons flow throughout the grid and so when we call on or when we utilize machine learning in this space we have to make sure that whatever outcomes that we have are in line with how a grid can be actually Implement um operated rather um
• 73:00 - 73:30 and to do that there are few things that we need to think about as mentioned in the chat um when using machine learning one is around safety and robustness um is is the solution um going to lead to cascading failures in our grid we need to make sure that um whatever ml the ml model outputs we're able to respect um safety constraints for um the grid we also need to ensure that we're still um
• 73:30 - 74:00 robust the grid is still robust it can still handle uh disruptions and whatever Solutions we come up with um are the solutions feasible yes we might come up with a minimal cost um solution but does it actually respect um the line limits or is it requiring more power flow than the line can actually like hold so there this there physical feasibility of the actual solutions that are can be put out by the model um do we understand the drivers or the levers for that output
• 74:00 - 74:30 and do we have control over them so can we understand what um shifts the decision from one uh prediction to the other and we have control over that so for example when we're predicting um demand do we understand why the demand is being predicted that way is it what what um latent information within the data is they learning from to make that prediction and do we have the right knobs to be able to um control that um do the outputs from the models meet
• 74:30 - 75:00 regulatory standards um there are different regulatory standards depending on where you are and it's very important that the um ml models um respect those regulatory standards for safety for um um and robustness of of operation um there are already existing methods too for different um Power Energy System uh problems whether in predictive maintenance whether in uh uh predicting uh generation there's already some performance me um um there some
• 75:00 - 75:30 performance measures that the baselines are already meeting so the machine learning methods really have to provide a case for why there should be a switch to the amount method it could be um they're doing better than the state of the art there it could be that um they're faster to run um they require less time um but we need to be clear why um we want to be able to switch and what metrics we use to uh determine that the ml model is um a viable
• 75:30 - 76:00 alternative um as mentioned throughout this um uh lecture or discussion there's different um infrastructure throughout the GD beat the skater systems um the different smart meters of the consumer level so whichever um model that we're using can it easily integrate with the existing infrastructure or does it require overhaul um to um acquire new infrastructure and this can really be prohibitive especially for emerging
• 76:00 - 76:30 economies where our resources are particularly constrained we need to ensure that whatever outputs we have can be directly um integrated to existing infrastructure um is the is the model usable is it accessible um Can people can stakeholders who actually U make decisions say the uh grid operator can can they actually use the output from this model or use the model and understand what it means or is it just a one-time thing that um they can have
• 76:30 - 77:00 access to um another aspect is also uh preserving privacy um it's quite important both privacy for consumers but also for the safety of the grid um and prevention from like Bad actors um are we making sure that whatever um work um the ml work we're doing with ML methods are preserving privacy of of users or of assets is also very important um finally um while um there's been development in
• 77:00 - 77:30 a lot of ml methods um it might not necessarily always be the case to develop um never satisfiable data hungry methods that and I think this is where the opportunity for like physics informed um machine learning really comes into place like are we using the data efficiently to be for for the space I would say add some key considerations um in this
• 77:30 - 78:00 space um now that we've um looked at what to think about When developing ml methods we we will go into a few case studies to see how we can contextualize some of these thoughts within um work that has been done um out there so um starting off with this uh graphic in I'd see when we think about using machine learning um models or methods in P system and really any any domain the one of the first things that
• 78:00 - 78:30 come to mind is what data set do we have and what um method um are we going to use most times is what is the newest and brightest method even um that we can use given this uh really cool data set that we have but uh what we're trying to advocate for um especially if the objective is to have impact is to take a step back and think if we were to ideally solve this problem who would care and why should they care about it
• 78:30 - 79:00 so understanding okay what domain are we in how do we scope this project who are the people that will care if we solve this and for them to care what metrics are they interested in when the model is developed how do we analyze that um the outputs from this model actually meets the metrics that they care about can the model be deployed um is it usable and what are the impacts and this is really not just even like a one-time um analysis it's something that we have to
• 79:00 - 79:30 sort of go through um multiple times as we better understand or grasp what the objective of the project is and what impact might look like and um it is through this lens that we want to be able to look at some of the case studies and see um how we can think about it along this uh framework or this mental model so uh the first case study that we will look at is mapping utility um skill solar in India um again the reference
• 79:30 - 80:00 format you can um access the the actual papers given the reference from lecture but the idea here is that um India has a goal of um renewable or 500 gaws of renewable energy capacity by uh 2030 and net zero emissions by 2070 and here there was no clear way of understanding okay how um well are are we meeting that objective that we've set
• 80:00 - 80:30 for ourselves so there was no um comprehensive data set on like geospatial information for utility skills to across the country so it was really hard to evaluate um are we meeting the goal that um we've set for ourselves so one approach that um the authors could have gone about would be to try to collect this information from different utilities but you can um see how this can be somewhat laborious and uh time consuming to do this either from
• 80:30 - 81:00 a survey standpoint or um manually colle um collecting this information so they opted to do um detection of uh utility scale solar using uh satellite imagery here the objective was by detecting um utility scale solar they can figure out or better support um where to place transmiss um lines where they actually um invest in transmission to be able to support
• 81:00 - 81:30 great integration of these Renewables um you can monitor a progress towards the targets that one has set for themselves especially um over time and can quantify what the impact would be of um placing utility skill solar on different um land use types so who would care about this kind of work would be energy uh energy planners policy makers who set this target even um um utilities who are
• 81:30 - 82:00 planning out um asset expansion um like we've mentioned in in previous um slides so to do this they started off with a small data set of solar PV Farms that served as um labels and they use multi-spectral sentinel 2 imagery um to learn uh and train the models and what they did was a I may say human machine collaborative approach
• 82:00 - 82:30 where um they did some clustering to obtain weak labels um of where um scale solar might exist and then from these week labels they were able to then train a supervised uh segmentation model for detecting um these Farms um from their um work they report a standard machine learning metric so things that you do with segmentation say
• 82:30 - 83:00 like um intersection over Union prision recall given the the the way the problem was formulated and set up but then they also try to report um other metrics that decision makers might actually care about so where are these PV um or where are these um utility scale solar plants locator what is the size when were they deployed what was the impact on land use and land cover the are actually metrics that a decision maker can um can take and say okay this is how it affected um
• 83:00 - 83:30 um the economy or or affected um where we where we placed Sol the thing to know about to note about this work is um figuring out even though they were able to um detect utility solar where the boundaries of this of this work so there's a focus on PV footprint over power related metrics or yeah over power related metrics um so
• 83:30 - 84:00 rather than saying that um or even energy related metrics I'm saying that or the plant is operating at this um um capacity they can that's beyond the scope of this work so understanding where the boundaries are of the work can also help the work um the some deployment considerations were how this approach might differ regionally um how sensitive is it to panel siiz this
• 84:00 - 84:30 I Simone I think we're losing your audio on and off maybe your um headset is dying oh hello can you hear me yes I think we're hearing you on and off a little bit okay I'll hopefully shouldn't be Dy but I'll give it one last try and then I'll switch off if it does um so yeah sorry about that by the way guys um and this these are authors have
• 84:30 - 85:00 actually uh extended this work to also consider um wind um power plants um in what they've termed AS Global Renewables watch uh a downstream effect that we would need to keep in mind is what are the implications of sharing the um location for uh generation assets this could either be like safety implications um but it's something that you might need to think about and how they de
• 85:00 - 85:30 decimate or share this information with um different stakeholders the other before I move to this case is the audio better yes much better okay okay thank you awesome so the next uh case study that we will uh look at is one around climate model uh projections and uh downscale in um solar Radiance from course to final resolution um um
• 85:30 - 86:00 outputs here um the authors are interested in better supporting um planning with renewable planning of renewable energy production um given the given changes in climate change and um the argument that they make is that um with the availability of corser resolution um estimates it's harder to plan um for renewable energy production especially
• 86:00 - 86:30 as the climate is changing and so there's a need to have higher or higher resolution of final grain estimates to really capture the local differences that might occur when estimating the amount of uh renewable energy that can be produced in a location so um the lack of downscaled um data or higher resolution corser resolution um um data was the problem here and one way is just to use the historical
• 86:30 - 87:00 available data but then one is in capturing the changes that may have occurred um that are now reflective in how much um Power can be generated um the other approach is to use finer green climate models but those can be very computationally intensive so they proposing an enhanced way of um getting to um higher resolution um outputs um without um high computational cost the pathway to impact here meaning that why
• 87:00 - 87:30 this would matter why this would be important is that these um higher or finer grain outputs um can be used for Power Systems planning but can also capture the local differences that might occur that might be lost when you're dealing with a Coster resolution um outputs and who would care um system planners um um utilities generators um anyone more
• 87:30 - 88:00 less interested in low carbon energy adoption um would be aable candidate um to care about this um from a data set standpoint they use the course skill um climate model outputs of solar Radiance and um as the input of like course resoltion data and as the output or training signal they provide the fine grain V analysis um soar irradiance product so
• 88:00 - 88:30 here it's a St supervised deep learning uh model modeling approach where um they are more or less predicting with subp pixel convolutions a finer scale and then comparing it to the true one and that like I mentioned serves as the training the training signal um the metrics that that they use are um the rot square error which is um good metric for this but they also use
• 88:30 - 89:00 what they call structural similarity index measure and the idea here is that with the uh mean square error there was so much smoothing occurring that you were losing the local differences um that might occur and that was one of the motivators of the work anyway so they um explicitly put in this metric to be to capture um how what we're doing in the local differences that might might be present um for the boundaries of this
• 89:00 - 89:30 method it's really dependent on having good uh supervisory uh data and the data came mostly from um the us but when shifting to uh data poort areas then um the methodology might not necessarily be as applicable or might need to be modified to um adapt to that scenario and um while they did spatial um resolution work they didn't do a
• 89:30 - 90:00 temporal resolution work so can they go from a lower temporary skill to a higher temporary skill that might be another dimension that um the method might not necessarily um apply um also um though the objective was on um better supporting uh planning for um renewable integration um it didn't consider maybe an out of scope of the work it didn't necess consider the grid constraints given existing um infrastructure which is fine but um it's
• 90:00 - 90:30 just about understanding where the where the boundaries of the work might lay and how to use it um at the end um the questions that they' have to also answer is in terms of deployment is can people easily access um this data set or use it or use insights from this data how accessible is it for someone who is interested in um replicating such work to be able to um um access the outputs sorry and and replicate the work and
• 90:30 - 91:00 finally um what the implications would be for different uh stakeholders when um carrying out planning um for renewable energy integration I'll I'll pause here for priia to hop in on the next uh case study all right so uh we're gonna go through two more case studies and then um a a you know a few more quick closing thoughts and then we'll be opening it up to your questions so feel free to also be gearing up with some questions in the in the chat um or in the um Q&A if you'd
• 91:00 - 91:30 like so the next case study we want to talk about is reinforcement learning for uh topology switching so the basic idea here um next uh bullet um is that um we need to increase the utilization of Renewables on the power grid and also adapt to extreme events that might happen on the power grid but the challenge as we've discussed before next point is that um and you can keep going
• 91:30 - 92:00 um that existing power grid optimization methods are not Dynamic enough or robust enough or fast enough as we talked about before um and in addition even as we try to develop new methods there doesn't tend to be as much you know good simulation infrastructure that that really shows us uh what's happening on power grids and this means that we're not able to properly validate our methods or deploy our methods um or
• 92:00 - 92:30 develop them in a way that is really contending with the real factors of the power grid and so some possible approaches to deal with this include you know optimization methods control theory methods or data driven methods on the methodology side for the optimization and on the infrstructure side um next bullet the um kind of approaches involve trying to create better simulators or test beds or even um I I'll use the the buzzword digital twins um of the power
• 92:30 - 93:00 grid in order to help us actually uh develop methods and validate methods for power grid optimization and so the pathway to impact here is that we ideally want these methods to ultimately be used by the power system operator um either autonomously or or human in the loop where there's a control operator who's giving highle action and and guidance and where the machine learning algorithm is executing those strategies on a lower level um for example on a subsec time
• 93:00 - 93:30 scale and the stakeholders here involve a combination of power grid operators and for example software providers that that might actually um power grid operators often tend to procure software from various companies to actually help manage their system so working with those software providers to actually integrate some of these techniques into the software is one potential kind of um Pathway to deployment so one project that's dealing with this is the um learning to run a
• 93:30 - 94:00 power Network challenge or L2 rpn which is run by the French power grid operator um the Electric Power Research Institute and a bunch of other additional power grid operators and both academic and non-academic stakeholders and what this initiative is is is a series of challenges where they try to um validate the use of reinforcement learning to optimize power grids and where they provide a simulator um and if we actually go to the next bullet uh and the next one after that they provide a
• 94:00 - 94:30 simulator called grid to op which is actually a um reinforcement learning style simulator that tries to mirror some of the the things that are challenging about optimizing power grids um in addition to that um that when you have a simulator that simulator has to have some data actually infused or injected within it and so this involves a combination of what is the structure of the grid what is the electricity demand and Supply on the grid and the data sets here some of them are
• 94:30 - 95:00 synthetic and some of them are reflective of the the French power system and the Machine learning approach here is a variety of methods so through the competition series people have submitted a large kind of variety of reinforcement learning methods um almost all of which use some combination of pure datadriven learning alongside heris sixs about what power grid operators actually do today alongside some physical knowledge of of power grid physics and and how the grid actually
• 95:00 - 95:30 works now how do you know whether a method is actually good right when how are you know how do you know if you're optimizing the power grid well so kind of similar to some of the considerations we talked about with AC optimal power flow we care about the cost or the price associated with the solution in addition we care about the physical robustness of the solution so when I'm optimizing the power grid using reinforcement learning does the power grid actually survive does it stay up or does it black
• 95:30 - 96:00 out and in addition we care about the speed of these methods the reason that we're using machine learning in this context is because we need faster and more realtime methods which traditional optimization problems are solvers are struggling with and so we also need to assess the speed of the solutions that we're producing the uh kind of boundaries of the methodology ology is in in large part that a lot of these methods are still early stage so they still are the simulators and the methods developed on
• 96:00 - 96:30 top of them are still generally optimizing relatively small power grids and using synthetic data so they don't capture the full scale and realism of the system and also even though in some cases the goal is to try to use these human in the loop so the you know the machine learning algorithm suggests something and the power the power grid operator suggests something right now the simulation environment are not taking that human in the loop capability into account and there are lots of deployment considerations that are then you know needing to be captured there
• 96:30 - 97:00 different control rooms in different Power grids work differently um and also different Power grids are operated more centrally or more of a decentralized manner or in a mix and the existing simulators tend to um uh suggest that you're doing a kind of centralized power grid optimization and in terms of the downstream effect when you're using you know machine learning reinforcement learning on a power grid there are lots of implications for pricing right are
• 97:00 - 97:30 you investing a lot of money in sensors and advanced software and such and that might increase power prices um implications depending on whether you're thinking fully automated control or human of thee Loop control for for the jobs of system operators um reinforcement learning can facilitate a change in Market structures for example from fully centralized markets to potentially decentralized markets and a potential Downstream effect is also when you're optimizing a power grid and
• 97:30 - 98:00 making a power grid more efficient you're also helping make it more efficient for every single resource that's on the grid so that can benefit Renewables but that can also benefit fossil fuel generators so it's important to understand sort of which power grid optimization progress is facilitating Renewables integration versus which is maybe actually neutral um in this respect all right so a last example uh I'm going to give here is about actually using machine learning to create data um which
• 98:00 - 98:30 in some cases again seems a little counterintuitive because machine learning relies on data but there are certain situations where we maybe don't have access to the real data because it might be private because it might be protected in certain ways and yet we're having a publicly available data set is very helpful for research and for policy analysis and for other actions that allow us to decarbonize the power grid um so one example in the next bullet um
• 98:30 - 99:00 is uh or actually you can go on uh three more bullets so basically given this need to have public Energy Systems data um but where there are sometimes regulatory or cultural barriers to sharing this data or privacy issues one possible approach is to create syn thetic data either manually you can look at your real data identify what makes that data look like
• 99:00 - 99:30 that data so if I have a data set of um electricity consumption um collected via Smart Meters from consumers I can look at that and maybe say okay the data has a certain shape in general it it it has a certain Peak at this point in time and you can try to replicate that manually or you can try to use automated techniques to try to figure out the distribution of the underlying data and replicate that in synthetic data the other approach is to facilitate sharing
• 99:30 - 100:00 of real data and that can be through techn you know technical methods like adding noise to the underlying data in a way that maybe makes it realistic enough but maybe protects it enough um or also via policy or organizational changes sometimes data is not shared not because of any actual technical or soot technical reason associated with the data itself but rather because it's hard it takes effort to clean data to host it and so of course this is this is an important thing to consider as well and
• 100:00 - 100:30 the pathway to impact here is that um you kind of want to ensure that data ultimately is available for use by you know researchers decision makers by market participants many other entities and this can be either through public release or in some s in some cases controlled release where someone can request access to data and there's some process to verifying whether they'll use that data appropriately so one approach that has
• 100:30 - 101:00 gone the synthetic data route and the Machine learning route here is a project called Faraday by the center for Net Zero and what they're trying to do is actually take data sets of of of real smart meter data which they have access to due to um their particular business which involves actually working with customers to to make their demand more flexible they have access to a lot of private smart meter data um but they can't release it because of privacy and Regulatory concerns and so the question
• 101:00 - 101:30 is can they actually leverage that data set to um produce synthetic data that doesn't have privacy leaks but that does allow kind of research and the policy Community to to work with um realistic data and so what they do um from a data set perspective is that they have a proprietary data set of um over 300 million um smart meter readings um coming from octopus energy which is an entity that they're affiliated with um
• 101:30 - 102:00 and their approach is basically to train a specific type of um of um autoencoder model a conditional variational encoder Auto encoder to map this training data to a latent space which basically means look at the data and try to figure out a kind of um smaller dimens fixed dimensional representation that Capt the axes of variation in that data um and then having mapped information to that Laten space they train a machine
• 102:00 - 102:30 learning model on top of that to learn that distribution of the data and then they sample uh next bullet from that particular machine learning model um and pass it through a decoder um to basically say okay I'm now going to be able to sample new data points that look like my old data points in in some high level way so when you're starting to generate synthetic data using a machine learning model you have to be really careful that you understand what it means for that
• 102:30 - 103:00 data to be good because if you imagine this nightmare scenario where you generate synthetic data using a machine learning model and then you use another machine learning model to analyze its synthetic data and your data is bad and your Downstream machine learning model is doing some stuff that's imperfect that can be like a horrible game of telephone for those who played this game as a kid where you whisper a into your friend's ear and then the friend Whispers it to the next person and by the end of the circle the phrase gets very changed from what it was um you
• 103:00 - 103:30 want to make sure that you're not propagating Errors By kind of having a bunch of machine learning models in a row that are feeding off of each other's outputs um and so you really need to understand in this case of synthetic data how do I know that the data is good and there are a couple of different metrics that are used here um one of them is Fidelity so can I calculate some statistics about the original data and does the synthetic data have those same statistics does it have the same mean
• 103:30 - 104:00 does it have the same variance if there are physics properties associated with the data or those preserved in the real in the synthetic data the other metric is uh utility in the sense of um usefulness so when I use the synthetic data for a real life application does it actually provide reasonable insights or does it give me the wrong answer so if I use use the synthetic data to ask the question of how much solar power should I um install on my grid in the future
• 104:00 - 104:30 and I use the private data to ask the same question the answer shouldn't be so far off from each other because otherwise you actually have not produced a good synthetic data set for the purposes of power grid planning and then a third metric and an important one is privacy um you know is the original and private data exposed now the boundaries of the methodology is as much as we can write down these metrics first we actually have to get better about writing down these metrics
• 104:30 - 105:00 um in order to fully flesh out exactly what you need to measure to to understand these these aspects um but it's also an open question how do you actually verify that those metrics have been ensured um or how do you actually construct models in a way that ensure those metrics and there are some deployment considerations here as well in that there are trade-offs behind the above metrics you can imagine that if you have data that's very high fidelity it might be more prone to leaking
• 105:00 - 105:30 privacy in some sense the highest Fidelity synthetic data set is one that is exactly equal to the original real data set but then that's exactly the wrong move in terms of privacy so really thinking through for your particular use case what's the right trade off between Fidelity utility and privacy that is an important deployment consideration um both in the context of a particular use case and overall and of course there are various you know regulatory requirements various issues of trust um and um you
• 105:30 - 106:00 know questions of who can access and use the data in practice which you know some folks say we want to try to get these methods to be good enough that you can fully publicly release data others say we're never going to get there but maybe we can do controlled releases where people can request access to this data and then um a kind of Downstream effect of this is that if you make data fly publicly available that improves access for everyone but that means quote unquote improved access for both good actors and Bad actors and so it's worth
• 106:00 - 106:30 again thinking about the implications of that when doing data release so we talked through a bunch of different case studies here for um the use of AI and power and energy systems and before we kind of move on to our kind of closing here one thing we want to emphasize is that very you know far back in the intro lecture um of the summer school we talked about various ways and which responsible AI is a is an
• 106:30 - 107:00 important consideration to think about in the context of climate change related applications and this is true in the context of power and energy systems as well um when earlier Simone asked about what are important considerations to think about with respect to ml in Power Systems many of the considerations that many of you dropped in the chat hit on a lot of these considerations and so that that's really great and so just to recap really quickly from the intro lecture um we want to make sure first bullet that um we are
• 107:00 - 107:30 cognizant of the ways in which biases may affect the the data that we're using the models that we're creating and the context in which we're using those models thinking about things like Geographic disparity in data availability we also um want to think about next bullet um uh the importance of trustworthiness and accountability in our model so we've talked about ways that safety and robustness are really important when optimizing a physical system but also
• 107:30 - 108:00 there are various ways in which as we talked about interpretability and auditability are really important because if something goes wrong on the power grid we need to be able to understand why that happened um in addition it's important to think about the ways that our use of AI and machine learning in the Power Systems and energy systems context um interplays with various aspects of You Know Who Holds societal um you know clout and power and who already has access to money for example so it's
• 108:00 - 108:30 important to ensure that these methods are deployed uh kind of you know developed and deployed alongside and and and with a diverse set of stakeholders and that the way these methods are are deployed doesn't um centralize Power or or exacerbate existing issues of of colonialism so for example smart meter data in different countries has different privacy protections associated with it and so we shouldn't for example just train machine learning methods on models associated with countries that
• 108:30 - 109:00 have not put privacy Protections in place um instead we should we should really think about this issue in a more um kind of Equitable way and these considerations are not exhaustive um and they're a lot to think about so of course the the there is a huge importance to working with relevant stakeholders throughout the entire development and deployment Pipeline and this is why kind of through our case studies sort of throughout the introduction to Power Systems we try to outline who some of those stakeholders might be and for any project that you're
• 109:00 - 109:30 working on it's it's it's abundantly important to think that through right from the get-go to make sure that you're shaping your project with that consideration uh very deeply in mind okay so we're gonna wrap up here and then I think we'll have a few minutes for questions at the end but um we wanted to talk through kind of next steps and and talk through our Journeys into um this space as maybe an example of what those steps might look like um so my own journey into machine learning
• 109:30 - 110:00 and Power Systems is that um as a high schooler I got very you know excited and interested about working on climate change but then in undergrad I got very interested in computer science and at the time it was really unclear to me how those two things connected to each other luckily um a group of researchers in the UK had put together a paper talking about how AI would be critical for um Power grids um and I got very excited about that area and then initiatives like the computational sustainability
• 110:00 - 110:30 Network you know communities that really brought together researchers um who cared about the intersection of of computer science and sustainability um those kinds of communities were really valuable to me um and some of what kind of climate change AI is trying to provide is this kind of thing and so definitely if you're interested in this area I very much encourage you to connect with your fellow summer school attendees and other you know take advantage of this community that that's here um in terms of positions and collaborations I think you know both my
• 110:30 - 111:00 research and academic Journey um was was very helpful in in doing this work but definitely you know various internships um at um you know us National Labs that work closely with power grid operators as well as at power system operators like National gri ESO um those really shaped my perspective on the problems that are actually being faced in the industry to ensure that my work in Academia um was really contending with those so this idea of you no matter
• 111:00 - 111:30 again no matter where you're doing your work if you're doing your work in Academia if you're doing your work in Industry I think trying to kind of work across and collaborate across those sectors can be really important in gaining the perspectives that are necessary to do this work well um and in terms of sort of top tips I would give um take classes in multiple disciplines so if you're a student you'll notice that you know a lot of the things we talked about here some of them are about electrical engineering some of them are about computer science some of them are about policy and regulation really engaging with these different areas can
• 111:30 - 112:00 be can be quite important to to actually you know working in this kind of space and in addition kind of internships and S conds or attending conferences hosted by different communities can be other ways to really understand the kind of interdisciplinary perspectives that are necessary to do work in this area engaging with local and Community organizations is also really um one way to to do work in this area so for example there are many cities that actually have you know um energy
• 112:00 - 112:30 advocacy groups or local electricity cooperatives and and kind of engaging with those groups can be one way to um to to work in this space um and in general again take advantage of the community around you reach out to people um kind of you know for example others were taking the summer school or more broadly and when you do that Outreach do your homework don't send a random note but instead kind of engage with what that person has done um and understand how you might actually work together in more
• 112:30 - 113:00 detail um I think for my journey into mln and Power Systems um my formative experiences around um electricity access reliability and lowend coverance offices really just came I would say more experiential I am from Cameron but I live in multiple places throughout subs and Africa so just being able to understand what uh were the common threads around um energy access
• 113:00 - 113:30 reliability and Power Systems in these different areas sort of sparked my interest in this space and um I was really able to engage with it then um as I did my uh Ms and and PhD um one thing that I think um helped me so of uh concretize or formalize um my interest and maybe work in this space was um doing internships that um were
• 113:30 - 114:00 true to both the problems I wanted to solve and the methods I wanted to use so for example being able to work on P Energy Systems style problems and also leveraging um computer science um algorithms and and principles I think helped me figure out which um places to to do and internship and eventually join as a as a as a researcher um I also spent um sometime um with working with
• 114:00 - 114:30 different stakeholders and maybe that'll be alongside some of my tips um I worked with uh Kenya power um um reg um Umi which are all utilities to understand so the challenges that they were facing the problems that they wanted to answer and how um we could use um ML and uh data to help them so my my number one tip is find something that you actually care about um I know that it can be hard to
• 114:30 - 115:00 try to answer the question in a vacuum um so to be able to figure out what you care about it it I'd say it starts from trying out different things so participate in in in competitions uh do some literature review learn more um about um the different aspects and domains thanks for Turing the the summer school on Power and energy systems but those will all go to like build your knowledge and intuition Bank on um what
• 115:00 - 115:30 you care about and what you should actually spend more time um doing for me I particularly found competitions very helpful because it helped me merge um domain knowledge with really the competitional skills to be able to apply ml in this aspect and then engage stakeholders uh actually um talking to utilities working with them help me figure out what problems they really care about what problems I thought they care about but that they don't so um it really helps honing on um what you
• 115:30 - 116:00 should be spending your time on if um you want to have impact and then finally enjoy the process and don't necessarily put pressure on yourself to have it figured out um up front but rather um think of it more as a journey which I'm still also um currently learning from uh in terms of resources we wanted to uh provide you some hopefully starting points on either uh data sets um resources simulators that you can um
• 116:00 - 116:30 hopefully um ignite your journey with uh the on our CCI Wiki we have an ex systems page that um lists out a bunch of data sets um a bit of background and really the different aspects that come into play when think about power energy systems so I would encourage you to go through this page and um learn more as maybe a starting uh place for the journey um in for some open data sets
• 116:30 - 117:00 there's the open energy data initiative from um the department of energy and while this data set is very uh us uh Focus or us Centric it's not necessarily uh a bad place to test out uh different approaches to build intuition to Horn skills so I would also encourage you to go um under this over maybe three paby of data on there so there's really quite a variety of things that you can um engage with there um Pria mentioned um I
• 117:00 - 117:30 data sets and um that's another space that you can go to for simulators um data sets um to test out some of your algorithms or to build some algorithms out and uh we're also uh working on updating the wish list but there is a base version there that you can see what the linkages are between the different data sets the type of ml methods that you can apply to those um problems data
• 117:30 - 118:00 sets and um what um climate relevant problem they'll be trying to um address so that's another resource you can utilize um if you're looking for Community which uh prer encouraged us to do that's the CCI Community platform and uh specifically um the power energy space which is a good place to find like-minded people interested in uh topics or um potential collaborators that you can work with and if you look
• 118:00 - 118:30 specifically at the power energy tar in our our newsletter you'll be kept up to date with uh some of the information that we have um on the changing on the changes that are occurring in the space um at this point maybe while uh people can take a minute to post what your next steps um would would be in uh either ML and climate change or ML and energy Journey we're interested in learning more about what that would be but we might also take this a moment to
• 118:30 - 119:00 answer a few questions before we are um out of and then alongside posting that in the chat we're also happy to um start um answering questions but we're definitely
• 119:00 - 119:30 very curious to hear what um what you might do with um with the information you learned today um given that we're running a bit uh out of time we'll do our best to answer um the questions posted in the Q&A uh by typing um so that we can that way we can answer
• 119:30 - 120:00 as many questions as possible also just a reminder that we've posted the form for the the Google form for feedback in the chat so please take a moment to fill that out when you have the chance e
• 120:00 - 120:30 Pria and Simone do you have maybe a bit
• 120:30 - 121:00 of time to answer some questions live uh to stay on and if you do okay fantastic so there are still a couple of open questions in the Q&A I think that would be nice to answer live so um we'll just go in order so one of the questions I was posed earlier on was what are some resources to design micr
• 121:00 - 121:30 GRS good question I actually don't have a great answer to this since I don't work in the like power grid design space I don't know Simone if you have any thoughts [Music] um not of the top of my head they are components to micro GD design um but not off the top of my head no I don't have a
• 121:30 - 122:00 good resource for it would you have any pointers for other people or places to look for resources for that um my starting point for for this would always be in literature um yeah and again would be to figure out what what aspect of micro micro grid design am I focused on um is it in
• 122:00 - 122:30 the is it in the interconnectedness is it in um the planning phase of uh uh Power Generation relative to like consumers yeah I'll try to hone down on like what aspect I'm interested in and then my starting point will always the literature um and I just Googling around I saw some initial like one link from the um uh US Department of energy called
• 122:30 - 123:00 the micro grid design toolkit so I have not used it before but I put a link in the in the um as a response to the question in case it's something worth looking into great um thank you there's also a question on um let's see uh for downscaling solar Radiance case study so um this one of the case studies are there papers follow
• 123:00 - 123:30 on papers that do temporal prediction um that's actually a good question I haven't looked into um temporal prediction using that study that we mentioned PR are you familiar um but what I might suggest is we do put the reference um which will share and once you put that you can check the CIT the citations and specifically look into that but can also quickly check to see if there were any
• 123:30 - 124:00 followups on temporal prediction great thanks um there's two questions from usra um first question is if we only have a meteorological data that how can we use them to carry out studies on forecasting for renewable energy apart from a simple weather forecasting management
• 124:00 - 124:30 system yeah so this is a situation where in principle if you don't have historical data on renewable energy production what you want to somehow try to do is match your meteorological forecast with knowledge of for example where specifically are their solar panels on the ground and what is is their capacity or where exactly are the wind turbines on the ground and what is their capacity um now of course this data is not always readily available
• 124:30 - 125:00 either um in principle you would ideally survey different kind of renewable energy providers and and and um owners of of power generation infrastructure to figure this out but some of the work that Simone talked about in terms of for example um detecting solar panels from satellite imagery and trying to understand or estimate their capacity from that that can be uh an intermediate approach where then you use your meteorological data alongside maybe a
• 125:00 - 125:30 satellite imagery derived estimate of what your on the ground solar capacity is and then you kind of match those up to to infer what you think the solar power production would be um there are some challenges with that right because for example solar power production is highly influenced by cloud cover and maybe you don't fully have that information but maybe you can infer that from historical satellite imagery as well so in this case it it does end up being kind of stitching together multiple sources of data that give you different pieces of information and
• 125:30 - 126:00 trying to figure out what what together those might imply um PR I actually bring up a good point around like um the work that if I may say leads to research outputs versus the work that actually leads to practical usage of um of um say a models like you mentioned the actual deployment might be a combination of multiple things to get um to a working
• 126:00 - 126:30 solution great thank you um another question by usra which is how can we manage the renewable energy production using IA which I take to to be Ai and forecasting what methods are what are the methods and strategies yeah so there's two ways to to think about when you already have renewable energy resources you know how to manage them so one of them is as we talked
• 126:30 - 127:00 about Ai and machine learning for forecasting so if you have more insight and foresight on how much renewable energy you have that allows you to do things like flexibly change how much electricity is being consumed at a given time using strategies like demand response um in a way that is matched with the amount of renewable energy Supply that's actually available the other option you have is to um if you for example have too much energy on the grid you can um for tail or basically
• 127:00 - 127:30 waste um solar wind power for example that's available on the grid but the goal is that we want to avoid that because then you know we're probably using fossil fuels to generate more power than is needed there and presumably you instead prefer to just generate less fossil fuel power rather than curtailing Renewable Power so um those are some of the the strategies that one can use great and there's just oh there's
• 127:30 - 128:00 three more um so this is from an anonymous attendee um you talked about minimization of costs as part of the optimizations could you comment on the risks of prolonging the use of fossil fuels within this approach instead of for instance looking at the minimization of emissions within the con within existing cost ranges you want to take this one Simone or should I um you can start so I was just reading it sounds good so um yeah so in an Ideal
• 128:00 - 128:30 World if we were really trying to like optimize power grids for emissions the for example AC optimal powerf flow problem instead of putting cost in the objective minimize cost we would put something like minimize emissions um in practice unfortunately this isn't how power grids are actually operated or how power markets are actually constructed in some sense the quickest way to get that to be reflected is if you have carbon pricing associated with power
• 128:30 - 129:00 generation then all of a sudden that gets internalized to cost and now your optimization problem is reflecting that but that doesn't fully necessarily solve the the problem either um one thing that is kind of fortunate is that a lot of renewable energy or a lot of lowc carbon Technologies and particularly Renewable Energy Technologies like solar and wind the What's called the marginal cost of producing power um is zero because the sun and the wind are free whereas buying
• 129:00 - 129:30 coal or buying natural gas produced power is not free so there are ways that luckily some aspects of cost do already align um with various aspects of of low carbon power but that's not always true and so that we have to do additional things to try to internalize that great there are some more questions coming in um I wanted to make sure that you're respectful of the speaker time
• 129:30 - 130:00 so yeah I just wanted to check how many more questions I think you'll have time to answer I could probably stay for another 10 minutes if that work so you have to drop off you should feel free obviously yeah I I I'll do five more and then I'll hop off if that's okay good great thanks so much so here's a question from Nan what strategies can be used to gather realistic data from de's Behavior behind meter for instance generation of electricity by solar panels before
• 130:00 - 130:30 consumption um so to start so there are a couple of things there one is um there are actually some programs like um pick con Street for example that are actually trying to outfit small like numbers of homes but certain homes with um very targeted
• 130:30 - 131:00 metering infrastructure in order to then assess you know what um there the de behavior is um and what the consumption behavior is and so forth and in some sense the hope is that by instrumenting you know multiple houses hopefully that are diverse and that are representative of different characteristics that you can get a sense of what's Happening Now in practice there's a lot of selection bias in terms of which households actually participate and where these things are installed but that is one
• 131:00 - 131:30 approach in principle getting targeted information on a smaller set of households in order to then extrapolate that more broadly um the other is that there are you know existing entities that that work very closely with consumers to do things like help um make their demand more flexible and these companies then get access to um more targeted consumer data they can't necessarily release it because of privacy issues but there can be ways if you're for example a researcher company to strike up bilateral agreements with
• 131:30 - 132:00 those entities if you you would like to maybe use that data do research with them commercialize with them or something else or um again there are some of these efforts to to try to generate synthetic data based on this realistic data so um that people can do studies even without those bilateral agreements I also think it depends on once sensitivity to um how accurate or realistic the data um needs to be um and
• 132:00 - 132:30 the balance between that and scale so if you are or for your specific application you're okay maybe with a smaller data set that is a true reflection for that um location then um I do 100% agree that um interacting with stakeholders may be treat data might be like viable Pathways now if you are trying to get a sense of what it might be at scale say for the US
• 132:30 - 133:00 you might have to um sacrifice a little bit around um um what I call the accuracy of the data depending on what you're trying to use it for for a method that can help you do that um at scale say very simply um looking at just panel size um um l the latitude um declination all of that to to estimate um output great thank you um there's just
• 133:00 - 133:30 two more questions here one from Harry where do you see opportunities for foundation models um pre I.E pre-trained large models that can be reusable in power system analysis besides weather prediction models that have already been done so I'll let you start because I think some of the work you do is actually more amendable to that than the work I do and I'm happy to talk about why I'm skeptical but maybe let's start with the more positive response um yeah I mean in principle foundational
• 133:30 - 134:00 models can help you um if well trained on the right data can help you uh uh Downstream with like fine ching and not having to retrain but specifically in like power and edgy systems we need to be careful um one around are do we already have in place the uh foundational models that are built on the right set of of data to answer the
• 134:00 - 134:30 questions that we have so um I'm not sure that there's a clear set of foundation models out there that can be readily applied in different aspects of P edgy systems maybe in more so in like the remote sensing um Community than in um other aspects for example I would say in operations and um it's optimal power flow I'm not sure that that that exists so I think one
• 134:30 - 135:00 would really need to be careful around what exactly we're trying to answer and where foundational models might play a role maybe in like questions that have to do with like imagery potentially language but um for other dimensional of Power Systems I won't be as um optimistic that it it's there just yet or even that it's the right thing um given the constraints that um we have for great
• 135:00 - 135:30 operations yeah so I'd say something similar in that if you really think about what a machine learning model or Foundation model is trying to do it's trying to in some sense figure out generally representative patterns that can be applied across different settings maybe with some additional fine-tuning and I think there are certain situations where that makes a lot of sense so when you're analyzing satellite imagery there's something common about kind of remote sensing data and image data that you know there there are some
• 135:30 - 136:00 fundamental features in there that potentially are usable across different sets of applications um and so I I can see places where Foundation models would be helpful for tasks like you know detecting solar panels from satellite imagery or detecting issues in powerg goodd infrastructure using satellite imagery because maybe there's something enough common among different satell imagery related tasks that there there's some shared features associated with that similarly with things like electricity demand uh prediction where electricity demand is driven by factors
• 136:00 - 136:30 like Behavior where it's hard to write rules down for but maybe there's some commonalities in those behaviors between different places this is a place where machine learning which is basically learning rules implicitly from data Maybe can be helpful in the idea that there maybe are common rules between these places but when you're talking about actually optimizing the the fundamental power grid somehow it seems that the things that are common between different Power grids are in some sense the underlying
• 136:30 - 137:00 physical equations and constraints that we already know how to write down and the things that then are are difficult are the Things That Vary on top of those systems so the fact that the P parameters are slightly different between different Power grids and that's something where you wouldn't expect generalization to work very well so I think in some sense because here we have the ability to write down explicit rules that give that common um grounding whereas in other settings where we can't
• 137:00 - 137:30 write down explicit rules that's maybe where you want a foundation model but here yeah I think we just know how to write down the rules that commonly bind different Power grids and we don't necessarily need to relearn them from data great thank you and uh uh one last question from Team AI thanks so much to the speakers also for staying online what please what are your thoughts around localizing customer
• 137:30 - 138:00 energy Data before requesting the data upon approval uh from each customer using Edge um and fog Computing I'm not really sure what fog Computing refers to but to boost their confidence in smart AI homes we are working on home appliance energy saving Technologies but we need to build customer confidence
• 138:00 - 138:30 yeah it's a good question I don't have a ton of expertise in this particular area um but I guess upon reading that I do think that there are uses of things like right like Federated learning or decentralized storage of data or things like that to protect consumer data from being um fully centralized fully exposed to various places um but I think building consumers confidence I think it's really
• 138:30 - 139:00 um it involves really demonstrating to them that you are actually right preserving privacy so just like sort of saying oh I'm storing it in a decentralized place well is that decentralized place secure what are the usage policies around that um if you're talking about sharing data up can you can provide any guarantees or certificates that you um when aggregating the data didn't expose individual consumer data so some of the considerations we talked about in the synthetic smart meter gen data generation and Faraday use case for
• 139:00 - 139:30 example apply where you can write down metrics but you have to be able to demonstrate them to to Consumers as well so I think this is a very reasonable and promising area but of course confidence needs to be earned um and so making sure that you're you're you're kind of doing that in through customer engagement and by by providing the right kinds of information and metrics is is yeah exactly this just took me back to the case study that you pointed out um I think the last case study PR where is one has to actually show um metrics that
• 139:30 - 140:00 demonstrate confidence and not necessarily contrive say contrive metrics there's a balance there between like contriving metrics that show confidence versus metrics that actually show confidence but I think one has to be quite powerful on what matters to Consumers and reflecting that in the analysis so um of of of the system great I think that does it for
• 140:00 - 140:30 all the questions um so thanks again um everyone for coming to today's lecture and also thanks to my support ta Ruth um so uh yeah with that I think I'll just thank again the speakers Priya and Simone for an excellent lecture thanks everyone thanks for the great participation and questions and everything thank you everyone and thanks
• 140:30 - 141:00 to our tears also