Regional data-driven weather modeling with a global stretched-grid

TL;DR

This study introduces a regional data-driven weather forecasting approach that uses a global stretched-grid and graph neural networks to deliver high-resolution forecasts ($2.5\mathrm{km}$) over the Nordics while maintaining a coarser global domain. The encoder–processor–decoder architecture employs a graph transformer to handle multi-resolution connectivity, with staged training that merges ERA5 reanalyses and MEPS high-resolution data. Evaluation against MET Norway’s MEPS and IFS shows improved instantaneous $2\mathrm{m}$ temperature RMSE and strong performance for 24 h temperature aggregates and precipitation, though extreme-event representation remains a challenge. The work demonstrates a scalable, low-cost path toward seamless regional forecasts suitable for public use and highlights future directions for hourly predictions and probabilistic outputs.

Abstract

A data-driven model (DDM) suitable for regional weather forecasting applications is presented. The model extends the Artificial Intelligence Forecasting System by introducing a stretched-grid architecture that dedicates higher resolution over a regional area of interest and maintains a lower resolution elsewhere on the globe. The model is based on graph neural networks, which naturally affords arbitrary multi-resolution grid configurations. The model is applied to short-range weather prediction for the Nordics, producing forecasts at 2.5 km spatial and 6 h temporal resolution. The model is pre-trained on 43 years of global ERA5 data at 31 km resolution and is further refined using 3.3 years of 2.5 km resolution operational analyses from the MetCoOp Ensemble Prediction System (MEPS). The performance of the model is evaluated using surface observations from measurement stations across Norway and is compared to short-range weather forecasts from MEPS. The DDM outperforms both the control run and the ensemble mean of MEPS for 2 m temperature. The model also produces competitive precipitation and wind speed forecasts, but is shown to underestimate extreme events.

Figures (10)

• Figure 1: (a) Map with annotated grid points centered around the Nordics. Global grid points are green, regional grid points are gray. (b) Input grid on the boundary between global and regional domain. (c) The encoder processes information into a mesh node from the 12 nearest grid points. (d) The processor mesh contains latent information, and has finer refinement over the regional than the global domain. (e) The decoder processes information back to a grid node from the 3 nearest mesh nodes. (f) On the mesh, information is processed between nodes at various refinements, here represented by the three largest ($M^0, M^1, M^2$) refinement layers.
• Figure 2: Model training follows a four-stage procedure. First, the ddm is pre-trained on 43 years of ERA5 data with a global resolution of 100 km (stage A) and 31 km (stage B). In stage C, we combine the 31 km global ifs dataset with the 2.5 km regional meps dataset with a training period of 3.3 years. Finally, the model is fine-tuned by auto-regressive rollout training over four prediction time steps (stage D).
• Figure 3: Forecasts of 10 m wind speed (colormap) and mean sea-level pressure (blue isobars) produced by MEPS and IFS (top row) and the stretched-grid ddm (bottom row). Forecasts were initialized at 06Z on January 27th, 2024.
• Figure 4: Forecasts of 10 m wind speed produced by the stretched-grid model, MEPS and IFS over the mountains in northern Norway. Shows lead time 24 h for a forecast initialized at 18Z on January 28th, 2024.
• Figure 5: Root mean squared error for 2 m temperature (a), 10 m wind speed (b), 6 h precipitation amount (c) and mean sea-level pressure (d). The lead time for 6 h precipitation represents the end of the time period.