Research project P4S/251/BEHRA (Research action P4S)
Artificial intelligence (AI) is rapidly transforming weather prediction. Since 2022, several machine-learning weather prediction (MLWP) models have been developed that outperform traditional numerical weather prediction (NWP) models for a range of standard meteorological variables. These models are generally trained on reanalysis datasets, which are produced by combining observations with NWP models through data assimilation.
Most global MLWP models rely heavily on the ERA5 reanalysis produced by the European Centre for Medium-Range Weather Forecasts (ECMWF). ERA5 has a horizontal resolution of approximately 30 km. While this is sufficient for many large-scale weather features, it is too coarse to adequately represent small-scale and high-impact weather, particularly extreme precipitation. Deep convection, for example, requires convection-permitting resolutions of around 1 km or finer to be explicitly represented. High-resolution reanalysis datasets are therefore needed to make MLWP more competitive with NWP for forecasting high-impact weather.
The Royal Meteorological Institute of Belgium (RMI) is developing its own MLWP capability using reanalysis datasets prepared by international partners. In this project, we will complement these datasets by assimilating Belgian observations that are currently not included in the relevant international reanalyses. The project will develop a convection-permitting reanalysis system based on a Rapid Update Cycle (RUC) configuration that is currently operational at RMI and runs on the high-performance computing infrastructure of ECMWF. The system includes surface data assimilation and a three-dimensional variational data assimilation (3D-Var) scheme. The current 3D-Var system assimilates GNSS, Mode-S, AMDAR and radiosonde (TEMP) observations. For the reanalysis, radar observations will be added to the assimilation system.
Given the available computational resources, the reanalysis will initially focus on a recent period and will then progressively be extended further back in time. The primary objective is to produce a 1 km resolution reanalysis covering five years. Subject to available computational resources, a longer period may subsequently be processed. The optimal configuration of the reanalysis system will also be investigated. This may include, for example, a wind-farm parameterisation to account for the increasing influence of offshore and onshore wind farms on local and mesoscale weather, particularly over the North Sea.
The resulting high-resolution reanalysis will provide a valuable dataset for future MLWP development at RMI. In particular, it could be used to train a high-resolution version of AIFS, the MLWP model currently being developed by ECMWF and its partners. The AI model training itself is outside the scope of this project, which focuses specifically on the development and production of the high-resolution reanalysis. The resulting datasets will also be made available to support further national and international research projects.
