On the last day of June, researchers from Forschungszentrum Jülich (FZJ), the European Weather Centre for Medium-Range Weather Forecasts (ECMWF), Deutscher Wetterdienst (DWD), and the University of Bonn met at the ECMWF Bonn site in Germany for the third in-person progress meeting of the RAINA project.
Key achievements presented during the meeting included:
- Linking Global Weather Forecasts to Local Precision
Encouraging progress has been made in enhancing precipitation, wind gust, and temperature forecasts over Germany using a fully machine learning-based forecasting system. The approach integrates GraphCast global forecasts with spatial downscaling via DeepMind’s CorrDiff model, to produce high-resolution, locally refined predictions. These forecasts demonstrate reduced biases, improved realism, particularly for precipitation fields, and stable performance across forecast lead times up to 24 hours. - RiverMamba Reimagined: From Floods to Rainfall
The RiverMamba model, developed by the partners from the University of Bonn, has shown strong performance in predicting river discharge, thereby outperforming traditional models. Building on this success, RiverMamba will be adapted to rainfall forecasts by training it on satellite-based observations from the IMERG datasets. - Learning to Rain: First experiments with the WeatherGenerator
Initial experiments with the WeatherGenerator prototype have focussed on improving the representation of precipitation in the model. For this, the WeatherGenerator was trained on the ERA5 reanalysis and the observation-based IMERG dataset. Early findings indicate that targeted model adaptions, e.g. the way how the precipitation is input to the model, can significantly enhance predictive performance.
The meeting highlighted continued progress toward delivering advanced forecasting tools to support disaster preparedness and climate resilience.
