Context-Aware Blending for Seamless Probabilistic Precipitation Forecasting

Nowcasting (NWC) and numerical weather prediction (NWP) systems provide high-resolution forecasts, but their performance depends on lead time as well as on the underlying observations, model configurations, and physical parameterisations across spatial and temporal scales. The ensemble nowcasting system STEPS, operated by DWD, delivers highly accurate precipitation forecasts for the first few hours, whereas the regional ICON ensemble variants ICON-RUC-EPS and ICON-D2-EPS outperform STEPS at longer lead times. These systems differ in their physical parameterisations, computational cost, and forecast ranges (up to +48h for ICON-D2-EPS and +14h for ICON-RUC-EPS). Combining STEPS with these NWP models offers the potential to produce a seamless probabilistic forecast up to 48h that outperforms each individual system.

Within the SINFONY 3.0 project at DWD, we develop a machine learning approach to generate a calibrated, seamlessly blended probabilistic forecast of hourly precipitation. Radar observations from DWD’s network are used for training. The method builds on recent advances in machine learning-based post-processing: Grönquist et al. [1] use deep U-Nets for bias and spread estimation, Rempel et al. [2] focus on the blending of ensemble nowcasting and ensemble NWP  and Primo et al. [3] generate calibrated probabilistic distributions using a neural network and include additional contextual data features such as seasonal and orographic parameters to improve predictions.

We introduce a U-Net-based neural network architecture with context-dependent modulation that adaptively weights the contributing forecast systems. This allows the relative importance of each input source to be adjusted dynamically based on lead time, seasonality and orography. The inclusion of contextual information therefore supports calibration, temporal consistency, and overall forecast skill.

[1] Peter Grönquist, Chengyuan Yao, Tal Ben-Nun, Nikoli Dryden, Peter Dueben, Shigang Li, and Torsten Hoefler. Deep learning for post-processing ensemble weather forecasts. Philos. Trans. A Math. Phys. Eng. Sci., 379(2194):20200092, April 2021.
[2] Martin Rempel, Peter Schaumann, Reinhold Hess, Volker Schmidt, and Ulrich Blahak. Adaptive blending of probabilistic precipitation forecasts with emphasis on calibration and temporal forecast consistency. Artificial Intelligence for the Earth Systems, 1(4), October 2022.
[3] Cristina Primo, Benedikt Schulz, Sebastian Lerch, and Reinhold Hess. Comparison of model output statistics and neural networks to postprocess wind gusts. In: A. Ott, W. Reichel, J. Warwicker, editors. Applications of Mathematics in Sciences, Engineering, and Economics Cham: Springer Nature Switzerland; 153–180; 2026.