Pan-European severe weather warnings by using deep learning and extrapolation techniques

We present initial results from the INLINE project funded by the Union Civil Protection Mechanism (UCPM). The objective is to develop probabilistic short-term warning products of severe weather with focus on localized phenomena, such as heavy rainfall, lightning and wind gusts in sub-hourly and kilometer-scale. The products are implemented in two different configurations: pan-European and one adapted to Finland by using locally available data sources.

The forecast models developed in INLINE are based on deep learning and extrapolation techniques. Our primary tool is the Simpler yet Better Video Prediction (SimVP): a deep learning model originally developed for prediction of multi-channel digital image sequences (Tan et al. 2025). The SimVP predictions suffer from progressive loss of small-scale features and extreme values with increasing lead time. Thus, we propose a modification to address this shortcoming. As an alternative we consider the extrapolation models implemented in pySTEPS (https://pysteps.github.io) and compare their forecast skill and computational performance with SimVP. We conclude that SimVP has up to 20% better forecast skill for lead times in the 2-4 hour range. This is mainly because of its ability to predict growth and decay of weather phenomena and spatiotemporal correlations between multiple variables. For very short time ranges (under 30 minutes), extrapolation methods still remain as a viable alternative to deep learning. In addition, we apply the perturbation generators implemented in pySTEPS to produce realistic forecast ensembles in a computationally efficient manner.

Radar-derived rain rate from the EUMETNET OPERA radar network is used as the main forecast variable. Additional variables include convective available potential energy (CAPE), convective inhibition (CIN) and wind u- and v-components. For training the models, we use ERA5 reanalyses, and use the corresponding IFS forecasts from ECMWF for real-time prediction. In the configuration localized to Finland, lightning flash density gridded to 10 km spatial resolution and 10-minute time window is used as an additional variable.

Gridded forecasts are translated into multi-hazard warnings by using combined thresholds that are defined by user-specified logical operations (e.g. rain rate over 10 mm/h and wind speed over 25 m/s). We present initial evaluation of the warnings by case studies from major storm events during 2023 and 2024. This is done using the European Severe Weather Database (ESWD) reports and emergency calls available from the Finnish PRONTO database as independent verification observations. Our results highlight the challenges in pan-European prediction of weather hazards, such as regional variability of data quality and availability of verification observations.