RainPro-8: An Efficient Deep Learning Model to Estimate Rainfall Probabilities Over 8 Hours

We present a deep learning model for high-resolution probabilistic precipitation forecasting over an 8-hour horizon in Europe, overcoming the limitations of radar-only deep learning models with short forecast lead times. Featuring a compact architecture, it enables more efficient training and faster inference than existing models. Extensive experiments demonstrate that our model surpasses current operational NWP systems, extrapolation-based methods, and deep-learning nowcasting models, setting a new standard for high-resolution precipitation forecasting in Europe, ensuring a balance between accuracy, interpretability, and computational efficiency. Code is available at URL. Recent advances in artificial intelligence have generated significant interest in deep learning for weather forecasting (Rasp et al., 2024; An et al., 2024). Although deep learning has achieved remarkable success in both nowcasting (Gao et al., 2024b; Gong et al., 2024) and medium-range forecasting (Lam et al., 2023; Price et al., 2024), significant challenges remain. Deep learning models for now-casting are often limited to very short lead times (up to two hours). In contrast, medium-range models, which predict broader atmospheric dynamics for up to 10 days, typically operate at coarser resolutions and are influenced by precipitation-specific biases in the training datasets (Lavers et al., 2022). As a result, they struggle to capture small-scale precipitation features, like local showers, often leading to the exclusion of precipitation forecasts in medium-range models (Lam et al., 2023). This work addresses the challenge of forecasting precipitation for up to 8 hours at high spatiotemporal resolutions, bridging the gap between nowcasting and medium-range forecasting. Forecasting over an 8-hour horizon is critical for timely predictions that help mitigate risks like flooding and optimize resource management in agriculture, energy, or transportation.