Weather Forecasting Using Artificial Intelligence

Mark Altaweel

Updated:

Weather forecasting is something we have become accustomed to; when it is wrong, we might complain, but often at least we have some confidence our forecasts can get predictions correctly at least for the immediate future, such as a day or two in advance. With artificial intelligence becoming more pervasive, scientist are now trying to improve weather forecasting using the benefits of AI. One such promising application is GraphCast that is now building on recent advances in deep learning technology.

Weather forecast accuracy

Most weather forecasts’ accuracy drops off considerably after 5 days; GraphCast promises relatively accurate forecasts for 10 days with predictions given for any global location quickly (less than one minute).[1] 

Medium-range weather predictions with AI

Forecasting in ranges of ten days is considered to be a medium-range forecast; the industry standard, High Resolution Forecast (HRES) produced by the European Centre for Medium-Range Weather Forecasts (ECMWF), produces forecasts to this range.

GraphCast is a Google DeepMind effort that uses deep learning; it promises to not only be more accurate than the industry standard forecasts but it has been pitched as a better tool to forecast extreme weather events. In fact, such events are becoming more common and motivation for creating GraphCast is, in part, driven by the increasing frequency of such events.

In a recent study, GraphCast was able to predict hundreds of weather variables for the entire planet up to 10 days in advance at 0.25° resolution. In comparison to traditional models such as HRES, predictions were more accurate in 90% of cases tested. Predictions were particularly better at events such as tropical cyclones, atmospheric rivers, and extreme temperatures, including areas more severely affected by these events. Thus, it could potentially be better at saving lives and property loss by providing more accurate forecasts of cyclones or damaging events well in advance of these events.

How GraphCast predicts weather

GraphCast’s neural network basis is a Graph Neural Network (GNN) that takes graph input. It takes the two most recent weather states of the Earth, which are the current time and 6 hours earlier, and then provides forecasting for 6 hours ahead for  0.25° latitude-longitude grid. This is roughly a 28 x 28 km grid, at least near the equator, with the grid covering the entire globe.

Predicted weather is then used to forecast additional weather states up to 10 days in advance. The grid representing the Earth’s surface applies a million grid points. For each grid point, the model predicts surface variables, including temperature, wind speed and direction, and mean sea-level pressure.

It also predicts 6 atmospheric variables at 37 levels of altitude, including humidity, wind speed and direction, and temperature. A decoder is used to map the processed output back to the latitude-longitude grid to provide specific location forecasting. GraphCast was trained using ERA5 data from 1979 to 2017. Additionally, a lower resolution (1 degree resolution, 13 pressure levels) model with 13 pressure levels can be used.[2]

Open source weather prediction

GraphCast is offered as an open-source project which is important for the advancement of the model and its transparency to the scientific community.[3] The idea is that more advanced deep learning methods should help improve medium- and long-term weather forecasting in the years to come as extreme events and climate change begin to drastically alter the Earth’s weather.

Given the need to improve forecasts, ECMWF is now testing GraphCast to see if it has the potential to become the main weather forecasting tool for medium-scale forecasts used in Europe.[4] 

Current weather models depend on Numerical Weather Prediction (NWP), which are algorithms that replicate physical models of how different aspects of weather works. Such equations are very time consuming, such as fluid dynamic models, and many are required in any large weather model. These algorithms are then run in unison but even then error in these models accumulate such that most models begin to produce high error between 5-10 days of forecasting.

Google DeepMind has also been creating other weather forecasts, including Nowcasting, a 90-minute weather forecaster, and MetNet-3, which is a a regional weather forecasting model for 24 hour weather forecasts. 

Artificial intelligence is gaining ground in weather forecasting

What is clear is AI-based weather forecasting will slowly replace some of our traditional weather models we have become increasingly reliant on. Before, scientists tried to improve physical-based algorithms that were used to forecast weather incrementally. Now, scientists can train deep learning models that use weather data to improve their forecasting ability, which should mean our accumulated weather data should help improve models rather than having to improve our physical algorithms. The goal will be to create better weather AI-based models that can predict short-, medium-, and long-term forecasts.

Having better forecasts will improve not only day-to-day knowledge of our weather but better prepare us for our changing climate

References

[1]    For more on GraphCast, see:  https://deepmind.google/discover/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/.

[2]    The article that introduces the model can be found here:  Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton-Rosen, Z., Hu, W., Merose, A., Hoyer, S., Holland, G., Vinyals, O., Stott, J., Pritzel, A., Mohamed, S., & Battaglia, P. (2023). Learning skillful medium-range global weather forecasting. Science382(6677), 1416–1421. https://doi.org/10.1126/science.adi2336.

[3]    GraphCast’s code can be found here:  https://github.com/google-deepmind/graphcast.

[4]    For more on ECMWF’s tests, see:  https://charts.ecmwf.int/products/graphcast_medium-mslp-wind850?base_time=202403170000&projection=opencharts_europe&valid_time=202403170000

Related

Photo of author
About the author
Mark Altaweel
Mark Altaweel is a Reader in Near Eastern Archaeology at the Institute of Archaeology, University College London, having held previous appointments and joint appointments at the University of Chicago, University of Alaska, and Argonne National Laboratory. Mark has an undergraduate degree in Anthropology and Masters and PhD degrees from the University of Chicago’s Department of Near Eastern Languages and Civilizations.

Free weekly newsletter

Fill out your e-mail address to receive our newsletter!
Email: