Improved weather forecasting using artificial intelligence is set to take a huge leap forward with the launch of a new European system that significantly outperforms conventional forecasting methods.
While technology companies and weather offices around the world are already using AI to predict the weather, the European Centre for Medium-Range Weather Forecasts (ECMWF) has announced that its operational model sets a completely new standard, making global forecasts available to all users at any time.
Florence Rabier, ECMWF Director General, said it was a breakthrough that would change the science of weather and forecasting. “The launch of the AI Forecasting System allows us to generate the widest range of parameters ever using machine learning.”
Tested over the past 18 months, the experimental version has shown the system to be about 20% more accurate at making key forecasts than the best traditional methods , which use millions of global weather observations and process them in supercomputers using physical equations.
Faster response times to extreme weather events
Florian Pappenberger, Forecast Director at ECMWF, said the new system can predict a tropical cyclone’s track up to 12 hours in advance , giving it extra time to warn of extreme weather events.
photo: Carlos Barria / Reuters
Other AI-based medium-range forecasting systems in development include GenCast and GraphCast from Google DeepMind, Pangu-Weather from Huawei, FourCastNet from Nvidia, and FuXi, developed by the Shanghai Academy of Sciences and Fudan University, all trained on more than 40 years of meteorological data collected by ECMWF.
Comparing the accuracy of competing AI systems is difficult because their performance varies depending on the variables and time scales analyzed. The results published by ECMWF give some idea of the performance of the systems, but do not indicate which model is the best.
AI will predict wind force at 100 meters
Pappenberger stressed, however, that the European system is unique in that it predicts many more parameters than standard models that forecast temperature, precipitation and wind. For example, it also predicts solar radiation and wind speed at a height of 100 meters – typical for wind turbines – which is particularly useful for the renewable energy sector.
Pappenberger said the current limit for reliable short-term forecasts in Europe is six to seven days for precipitation and wind and up to 14-15 days for temperature.
“Machine learning models have a real chance of extending this period because they can capture information that we can’t yet represent well enough in physics-based models.”
While ECMWF forecasts are publicly available, the agency does not issue extreme weather warnings or tailored forecasts for the industry, leaving those tasks to national weather services and private companies.
Anemoi is Europe's answer to AI-based forecasts
ECMWF and a group of European national meteorological offices have created an open technical platform for AI forecasting systems called Anemoi, named after the Greek wind deities. The architecture is based on the same “graph neural network” that Google DeepMind’s forecasting models use.
ECMWF plans to further improve the system by increasing its spatial resolution and moving from the current version, which generates single forecasts, to so-called ensemble forecasting. This means generating 50 forecasts simultaneously with slightly different initial conditions, which will allow for a more precise definition of the range of possible weather scenarios.
In the future, as Kirstine Dale, director of AI at the UK Met Office, noted, a combination of both physics-based and data-driven simulations will be needed to achieve fast, reliable and accurate forecasts.
JS
