Modernization of the seamless point-based forecasting system MOSMIX using AI

MOSMIX is a point-based, global forecasting system providing predictions up to 14 days ahead for a wide range of meteorological variables. It features a seamless transition from observations to forecasts by downscaling and combining output from the global NWP models IFS and ICON over the previous three days. Among forecasters at Deutscher Wetterdienst (DWD), MOSMIX has earned a strong reputation for its high quality, particularly for variables such as wind and temperature, and is therefore widely used in operational forecasting and warning processes.

Despite being operational for decades and having reached a mature state, MOSMIX is based on increasingly outdated technology. The system relies primarily on multiple linear regression, with nonlinearity represented through a large number of situation-dependent equations. Specifically, separate equations are defined for each model run, location, variable, lead time, and season. While these millions of equations ensure high forecast quality, they also necessitate additional measures to maintain consistency across variables and lead times. As a result, the overall complexity of MOSMIX has become a bottleneck for further development.

Moreover, MOSMIX produces high-quality forecasts only at locations where observations are available. Although interpolation is possible for sites near observation stations, the system cannot generate fully gridded forecast fields.

To address these limitations, DWD is currently developing a new system intended to replace MOSMIX. This presents a significant challenge, as the new system is expected to match or exceed the performance of the current operational system while retaining its key features. At the same time, it should provide gridded forecasts, at least over the European domain.

In this presentation, we introduce a prototype of a hybrid grid- and point-based system that aims to achieve MOSMIX-level performance at station locations while simultaneously producing gridded forecasts. The approach extends the station embedding methodology of Rasp and Lerch (2018) and Schulz and Lerch (2022), enabling the neural network to be applied consistently on a spatial grid. Model performance is evaluated using out-of-sample verification, in which the stations used for validation are excluded from the training process. We also address the issue of consistency across variables.

Preliminary results for wind gusts indicate that, while it remains challenging to outperform MOSMIX in terms of standard metrics such as RMSE, the new approach can achieve comparable performance.

Literature:

Rasp, S., and S. Lerch, 2018: Neural Networks for Postprocessing Ensemble Weather Forecasts. Mon. Wea. Rev., 146, 3885–3900, https://doi.org/10.1175/MWR-D-18-0187.1.

Schulz, B., and S. Lerch, 2022: Machine Learning Methods for Postprocessing Ensemble Forecasts of Wind Gusts: A Systematic Comparison. Mon. Wea. Rev., 150, 235–257, https://doi.org/10.1175/MWR-D-21-0150.1.