Sign in to continue learning
AI and Climate Change
15h40m
Hello everyone. My name is Lester Mackey, and today I'd like to tell you a story. It's a story of climate, of machine learning, and of the Wild West. The story begins with this guy, Judah Cohen. Judah is a climatologist and the Director of Seasonal Forecasting at Atmospheric and Environmental Research. And one day Judah came to me and said, Lester, I'm concerned. I'm concerned that the community is not making the best use of historical data in weather and climate forecasting. Instead, the landscape is dominated by these dynamical models, these purely physics-based models of how the oceans and atmosphere evolve. Now, I'm no expert in dynamical models, but here's a cartoon. First, we need to estimate today's weather conditions. We can measure the weather at certain places, and then we estimate it everywhere else. Okay, and now we know how physics works. We know how the atmosphere evolves. We know how oceans evolve. And so we can actually then simulate what future weather is going to look like, given today's weather conditions. This involves simulating what are called partial differential equations. You can never really do this exactly. What you need to do in practice is you break up space into small buckets, small grid cells, like I've shown you on this globe here. And you also discretize time into small increments. And then you can actually simulate that in practice using a supercomputer. All right, so this is what we do in practice today. But the accuracy of this approach is limited, and this has been known at least since the 1960s. It's limited by what's called chaos. This means that if there are any errors in your inputs, then those errors are going to be rapidly amplified, so that by the time I'm trying to predict two weeks from now, my predictions will be terrible. Now, the community has a couple of ways of trying to get around this. The first is called ensembling. This just means let's run our models, let's simulate the future using different initial conditions, some perturbations of what we observed today, and then get the predictions from each of those models and average them together. That's called an ensemble forecast. The other thing that's commonly done is called debiasing. I take the predictions of my model, I look at their predictions over the last 20 years, I subtract away the average prediction, and I add back the average of the ground truth, debiasing. And in some cases, this is the only way in which historical data is going to enter into the forecast. Okay, and that was concern number one, and Judith says, I also have a second concern, and that's sub-seasonal forecasts are especially poor. What does that mean? Well, I think we're all pretty familiar with weather forecasts. These are short-term predictions of temperature and precipitation. Typically, we trust these weather forecasts if they're about a day out, and we don't really trust them anymore if they're 10 days out, 14 days out. And there's a good reason for that. It's because they're very bad. But on the other extreme, we have what are called climate forecasts. These are predictions over months, seasons, years, decades, and typically we have more skill, more accuracy at predicting these climate forecast phenomena because they're averaging over more time. So it's a less noisy target, and we can do a better job of predicting it. But in between these two regimes, weather and climate is what's called sub-seasonal forecasting. This is predicting temperature and precipitation two to six weeks from now. And here, the predictions are especially bad because the weather models have broken down due to chaos. The climate models aren't good yet because you're not averaging over enough time. And also, the forecasts that you make two to six weeks from now depend on the weather today locally, and they also depend on weather conditions around the globe, making it a very challenging prediction problem. And it turns out that Judah was not the only person concerned about this. Around the time that he first approached me, the U.S. Bureau of Reclamation, which is a government agency, was also concerned about this problem. Now, the USBR is responsible for managing water in the western United States. They provide irrigation water to farmers. They're the second largest hydropower producer in the U.S. And they note that in the past decade, every state in the western U.S. has experienced drought in a way that affects both the local economy and the global economy. And they were so concerned about being able to anticipate this in advance that they decided to hold a contest. They called it the Sub-Seasonal Climate Forecast Rodeo. Rodeo because it's the western U.S. and that's where the cowboys are. And they said, if you can do a better job at forecasting temperature and precipitation at the sub-seasonal scale, two to six weeks from now, then we'll give you $800,000 in prizes. Okay, so we decided to enter this contest, but a strained aspect of this, what you might call a data science challenge, was that there was no data provided. Really, the USBR just said, do whatever you want. Just find some way of predicting temperature and precipitation better. So we created our own data set. We call it the Sub-Seasonal Climate USA data set. It consists of temperature data, precipitation data, pressure, some larger scale phenomena, like the Madden-Julian Oscillation, the state of El Nino, and it even has forecasts from some of these dynamical models. We put it on Azure, we update it daily on the cloud, and you can access it via the Sub-Seasonal Data Python package. Now, on top of this, we decided to build some hybrid learning plus physics forecasting models. And we found that these models double or triple the forecasting skill of the operational, just physics-based models that are used by the U.S. And also outperform state-of-the-art machine learning and deep learning methods targeted for this problem. And so what I'm showing you here on the top is the CFSV2, the Climate Forecasting System V2. This is the operational sub-seasonal forecasting system that the U.S. uses. And on the bottom, I'm showing you our new, what we call Adaptive Bias Correction method, which blends together forecasts from CFSV2 and machine learning corrections. And you can see that across the U.S., we're seeing improvements that double or triple its forecasting accuracy. This is for temperature, and we see the same sort of thing for precipitation. Again, double or tripling of the forecasting skill. So that's all the time we have today. But if you want to learn more about this work, I can point you to any of these papers here. We have a few publications and preprints on archive. One's called Improving Sub-Seasonal Forecasting in the Western U.S. with Machine Learning. Another is called Online Learning with Optimism and Delay. And finally, Learned Benchmarks for Sub-Seasonal Forecasting. Thank you.