You may have had the experience of unboxing furniture and come across instructions that go something like this: "Assemble the bookshelf according to the provided diagram." Yes, I know what a bookshelf looks like. Probably wouldn't be reading the assembly instructions if I didn't need a little more help with the process. Or maybe you've opened a cookbook with an author who thinks you're already somewhat of a chef. "Deglaze the pan." What?
(Laughter)
OK, off I go on a separate search to understand whatever that means.
Instructions that tell you what to do and not how to do it are pretty useless. And yet, even when we're talking about something as important as climate change, we hear them all the time. “Transition to renewable energy.” “Electrify everything else.” “Deploy solutions that are equitable and fair.” Yes, let's do all of that. But how?
Answering how is where we understand which solutions are actually feasible, whether that be with today's infrastructure, our evolving regulatory environment or any of the other number of dependencies and constraints that we have to consider. How we solve climate change also depends on our very definition of the problem. It's a scientific challenge, a sociopolitical issue, an economic problem and so much more. And how we solve it will depend on how we frame it. There is no single answer.
I'm a scientist, so I approach climate change as a scientific challenge. I'm also a techno-optimist and artificial-intelligence product manager, so I also approach it as a technological one.
When it comes to a sustainable future, artificial intelligence can help us do three critical things. First, it can help us understand climate change and its effects on Earth's ecosystems. Second, it can help us optimize current systems and infrastructure, because we can't just start over from scratch today. And third, it can help us accelerate the breakthrough science we need, such as fusion as a carbon-free energy source.
Today, I'd like to talk about that second one, optimizing current systems, and specifically, how we can use AI to harness a superpower we already have in this fight: wind energy.
Renewables are unquestionably a key to a sustainable future, but the problem is they're unpredictable. Sometimes, the sun shines and the wind blows, and sometimes, it just doesn't. Now, for an electricity systems operator, who needs supply to meet demand in real time, 24-7, this is hugely problematic. Renewables can't be 100 percent reliably scheduled.
Now, unfortunately, fossil-fuel plants are the opposite. You can burn a specific amount of coal at a set time to deliver exactly the amount of electricity you want in a predictable time window. So ... if you're a power systems manager whose job is to literally keep the lights on, which source are you more confident depending on?
But here's one of the places where AI can come in. It is a powerful tool for forecasting. AI systems can ingest vast amounts of historical data and help us predict future events. So, while we can't eliminate the variability of wind, we can use AI to more accurately predict its availability. That was my team’s “what” to do. Use AI to accelerate the transition to renewables, like wind energy. The tough part was the “how” to do it.
First, we researched the challenge. We read papers, we spoke to domain experts, we found out everything we could about the problem. Our team, which is a mix of research scientists, engineers, a product manager, a program manager and an impact analyst, decided that a neural net trained on historical weather data and turbine power-production information would likely help us accomplish our goal.
Next, we needed to find two core elements: data to train the system and a partner who was willing to deploy it. Both of these can be major obstacles when it comes to deploying AI in real-world scenarios.
Let's start with data. There are massive gaps in climate-critical data -- not just in electricity, but in agriculture, transportation, industry and many other sectors. Some of our data, we could purchase or download for free -- weather forecasts, for instance. But some of the data we needed was proprietary, and this would be, like, turbine power-production information and other operational data from the wind farms. Now, we needed that proprietary data so that we could train our models to learn the relationship between historical weather and historical power production, so it could then then make predictions about future power availability based on what data said about future weather. Now it's probably worth mentioning here that we were looking at a few years of data on hourly resolution, not historical data at a timescale that would have massive climactic differences from present day.
In addition to data, we needed to find a partner with domain expertise and the willingness and scale to test new systems. You know, surprisingly, this can be a major hurdle when it comes to deploying AI in the real world. Believe it or not, it's not every wind-farm manager that wants to let a bunch of AI researchers test on their multimillion- or multibillion-dollar systems. But the thing is, in order to prove that AI works, we have to have deployment opportunities in the real world. Luckily for us, Google was a ready and willing partner. OK, yes, DeepMind is a part of Google, but it's not a given that they would let us test on their systems. Yet they let us test on 700 megawatts of their wind-power capacity, which is equivalent to a large wind farm in the United States. This made them an excellent proxy for external wind-farm operators. They also lent us an expert team to advise on metrics and benchmarks and to share the data that we needed. This is another critical component of AI for the real-world deployments. Working with a domain-expert team that can tell you what they need, how they need it to work, which constraints keep the system safe, what quantifiable metrics to use to measure AI performance and how much better that AI performance needs to be than their previous systems to make the cost of switching over even worth it. And that's just to name a few.
So at this point, we have our idea, we have our data, we have our deployment partner. Now, to test and deploy our system. Improving the accuracy of electricity-supply forecast is incredibly important. If predictions are higher than actual generation, renewable electricity managers may not have enough supply to meet demand. This, in turn, drives the purchase of carbon-intensive fossil fuels to cover that gap, because they're largely what makes up backup generation. Now, the good news. Our AI system performed 20 percent better than Google's existing systems. Even better news is that Google decided to scale this technology. And scaling is so important. We will run out of time in the climate countdown if we aren't deploying solutions that are widely applicable. This particular solution is being developed into a software product that French company Engie is among the first to pilot.
But, you know, it doesn't even take a major research organization to do this kind of work. Where we focused on AI for supply-side forecasting, a small UK-based nonprofit called Open Climate Fix is focusing on AI for demand-side forecasting. They found a willing partner in the UK National Grid, and are currently deploying forecasts that are two times more accurate than the UK grid's previously used systems. Now, all of this is to say is that AI can help us with the transition to renewable energy, but scientists and technologists, we're not going to be able to do that alone. We need to be working with partners and experts who can teach us the “how.”
So for those of you interested in this space, if you're a domain expert, please share the problems you face and the challenges that you have so that our sector can ensure that AI pursuits will have impact in the real world and not be purely academic. Even better, if you want to incentivize ML researchers to work on your problems, I'll let you in on a little secret: build a competition, and they will come.
(Laughter)
It's true. Just don't forget the datasets and metrics. If you are a data holder, where it’s safe and responsible to do so, please share data related to those challenges. If you're not sure whether the data you have is even climate-critical, you can check out Climate Change AI's website, where they have published a wish list of climate-critical datasets. Access to these datasets would unblock crucial research and innovation in AI for climate. If you're a deployment partner, please, let us know who you are, especially if you're willing to test innovative systems. And for everyone who's interested in this space, please know you do not have to be technical to work in tech. AI for climate action requires a variety of skill sets and a diversity of backgrounds that, yes, includes research scientists and engineers, but it also includes ethicists and policy experts, communication teams, product managers, program managers and so many more folks.
Now for the warning label. AI is not a silver bullet. It will not solve all problems driving climate change. It isn't even the right tool for many of the challenges that we face. AI is also not a technology without tensions. It needs to be deployed safely and responsibly. Not to mention, until our grids are run on clean energy, AI itself will carry a carbon footprint, as will any energy-intensive technology we use. But AI can be a transformational tool in our fight against climate change -- it's just on all of us to wield it effectively. The “why” we need to is absolutely harrowing. The “what” we can do is really exciting. But it’s the “how” we can do it that will illuminate feasibility and help us drive impact.
So, in your next climate action conversations, when someone presents you with an exciting "what," please help to advance the conversation to the impactful "how."
Thank you.
(Cheers and applause)
