Machine Learning Methods for the Postprocessing and Seamless Blending of Ensemble Forecasts

Numerical weather prediction (NWP) models can deliver high-resolution forecasts, but achieving optimal performance requires the fine-tuning of model parameters and adjusting the physical models for different spatial, temporal, and seasonal scopes. Models differ in forecast accuracy for specific lead times and seasonal environments, offering a variety of fragmented forecasts exhibiting temporal discontinuities. This happens with the two regional ICON variants for Germany, ICON-D2 and ICON-RUC, which are operational at DWD and differ in their physical models and forecast range. Although both operate on a 2 km scale, ICON-RUC focuses on convection, is more computationally expensive, and it provides shorter lead times (+14 hours for ICON-RUC vs. +48 hours for ICON-D2). Therefore, blending is necessary for a seamless forecast for up to 48 hours. Moreover, NWP systems are biased and postprocessing is required for calibration and error correction.
As part of the SINFONY 3.0 project at DWD, our goal is to use machine learning methods to develop a postprocessing framework that delivers an improved combined probabilistic forecast that is calibrated and seamlessly blends the output data from both NWP models, with a focus on hourly precipitation. DWD’s radar network measurements is used as the ground truth for training.
We build on previous work that utilizes ML to improve probabilistic forecasts: Grönquist et al. [1] used deep U-Nets for bias and spread estimation. Rempel et al. [2] focus on the blending of ensemble nowcasting and ensemble NWP, and also show that their network can improve calibration. Primo et al. [3] generate calibrated probabilistic distributions using a neural network and include additional contextual data features such as lead time, seasonal, and orographic parameters to improve predictions. This work presents a neural network that blends both NWP systems and provides a calibrated probability.

[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. arXiv [stat.AP], January 2024.