Norway’s complex terrain, rugged coastline, and high-latitude environment make accurate 0–6 h weather prediction particularly challenging, while societal needs for reliable short-term forecasts, ranging from flash-flood warnings to renewable-energy management, continue to grow.
At the same time, Norway benefits from an exceptionally dense observation network that includes official stations, Nordic radar composites, Meteosat satellite imagery, and thousands of quality-controlled citizen weather stations.
This project leverages these rich data streams to develop a next-generation nowcasting system based on Graph Neural Networks (GNNs) within ECMWF’s open-source Anemoi framework.
The model fuses multi-modal inputs on their native grids using dedicated encoders and a graph-based representation that naturally handles irregular observation geometry and complex topography. Key innovations include a station-dropout strategy that enables robust use of heterogeneous and intermittently available citizen observations, and the integration of spatially aware loss functions to better capture fine-scale precipitation structures.
Using EuroHPC resources, we will train several multi-modal GNN variants on multi-terabyte datasets covering radar, satellite, station, and NWP fields, and perform extensive verification using pointwise, spatial, and event-based metrics.
The resulting prototype aims to deliver high-resolution (1 km, 5–10 min) nowcasts for precipitation, temperature, wind, humidity, and solar irradiance. Expected outcomes include substantial improvements in short-term forecast skill, enhanced situational awareness for public safety, and better operational integration of renewable energy.
By combining dense observations, advanced AI methods, and large-scale HPC, this project provides a scalable blueprint for national-level nowcasting and contributes to Europe’s broader transition toward AI-augmented weather prediction.
Ophelia Miralles, The Norwegian Meteorological Institute, Norway