Regional Data-Driven Weather Forecasting: Preliminary Results of a Probabilistic model using HARMONIE-AROME 2.5 km Data for the Iberian Peninsula.

The rapid emergence of Data-Driven Weather Prediction (DDWP) models has revolutionized global forecasting, with ECMWF’s Artificial Intelligence Forecasting System (AIFS) [1] showing the ability to capture complex dynamics and demonstrating competitive skill against traditional Numerical Weather Prediction (NWP) models. While models such as the AIFS are highly effective, most are designed for global applications and trained on datasets like ERA5. Despite its extensive temporal coverage, the spatial resolution of ERA5 is often insufficient to resolve the fine-scale atmospheric processes critical for limited-area modeling where kilometer scale resolution is essential, e. g., to properly capture extreme events. Their adaptation to convection-permitting scales remains a significant frontier, especially in regions with complex topography, such as the Iberian Peninsula. This study addresses these limitations by presenting an evaluation of a high-resolution regional AI forecasting model for the Iberian Peninsula, that bridges the gap between global and regional systems and builds on the ANEMOI open-source framework.

Our approach utilizes the Anemoi framework to train specialized models for the Iberian Peninsula, leveraging the valuable global context of ERA5 for limited-area modeling. Central to our methodology is a tiered training strategy designed to harmonize the extensive temporal record of ERA5 with the high spatial resolution of AEMET’s HARMONIE-AROME [2] dataset (2.5 km spatial resolution). Following initial hyperparameter tuning on ERA5 O96 and pretraining on O320 grids, we perform fine-tuning with HARMONIE to better capture localized weather phenomena that global models typically fail to resolve. To ensure spatial continuity and provide external boundary conditions, we employ a stretched grid as done in Bris [3]. This technique integrates high-resolution data over the target domain with lower-resolution information from the surrounding global region, creating a robust prediction system that mitigates training limitations arising from the comparatively shorter temporal extent of regional datasets.

Initial deterministic experiments in this setup revealed that while the model captures general atmospheric patterns, it struggles to represent the magnitude and frequency of localized extreme events. To address this, our research has transitioned toward the application of probabilisticforecasting architectures available in Anemoi framework. Specifically, this study evaluates the performance of Generative Diffusion Models in comparison to the Continuous Ranked Probability Score (CRPS) optimized probabilistic approach developed at AEMET.

Ultimately, this work serves as a practical demonstration of how the Anemoi ecosystem can be implemented to establish high-resolution, computationally efficient probabilistic regional forecasting workflows within a National Meteorological Service context, such as AEMET. Beyond this implementation, our primary scientific objective is to assess the validity of Generative Diffusion Models as a robust approach to accurately represent uncertainty and capture the localized extreme events over the complex terrain of the Iberian Peninsula, which remain a challenge for both traditional NWP and current data-driven systems.