European Weather Service: New AI weather model needs 1000 times less electricity
The weather service for dozens of European countries now also publishes forecasts generated by AI. This requires drastically less energy.
(Image: Wallenrock/Shutterstock.com)
The European Center for Medium-Range Weather Forecasts (ECMWF) is making available data from a new AI-based weather model that is significantly more accurate and requires around 1000 times less energy to produce a forecast. The research institute announced this on Tuesday, explaining that the model, called AIFS (Artificial Intelligence Forecasting System), is the first operational forecasting model that relies on machine learning for the widest range of parameters. Its availability will have a positive impact on how the national weather services of the 35 member states – including Germany, Austria, and Switzerland – produce their forecasts.
Soon even more AI weather forecasts
ECMWF Director General Florence Rabier calls the deployment of the AIFS a milestone that will also revolutionize weather research. Like the forecasts produced using traditional technology, the system will be fed with more than 800 observations collected daily from a wide variety of sensors. Initially, individual forecasts will be produced, but the aim is to create 50 slightly different forecasts at a time. These so-called ensemble forecasts are intended to represent the entire range of possible weather conditions. Later, hybrid models consisting of physics-based and AI-based models will be made available.
Videos by heise
The AI forecasts will be produced four times a day and will then be available as open data. Initially, however, the resolution is still lower than that of the physics-based models, according to ECMWF. The release comes less than three months after the presentation of an AI-based weather forecast by Google subsidiary Deepmind, which stated that the high-resolution GenCast model is better at predicting small-scale weather over a period of 15 days than ECMWF. Even more impressive was the fact that a single processor, rather than a supercomputer, was sufficient to create the model. This is now also possible with ECMWF, where the accuracy of the results is said to be around 20 percent higher.
(mho)
