“Pseudo-deterministic” precipitation products, a clustering-based approach to combine and reduce ensemble information

Ensemble prediction systems have evolved as a standard in modern weather prediction. At the same time, with an increasing number of members and higher spatiotemporal resolution, the amount of data produced is growing fast. However, not all operational users are able to process the increasing amount of data, which calls for methods to reduce ensemble information.

The SINFONY project at Deutscher Wetterdienst (DWD) focuses on the probabilistic forecasting and Nowcasting-NWP combination of precipitation, with the aim of better representing convective heavy rainfall events. Ensemble reduction of SINFONY output data is based on two concepts. Firstly, it aims to compact the information of an ensemble into a new, combined member (the “pseudo-deterministic” member). The second approach is to retain most of the information in a member-reduced ensemble that also contains the newly created pseudo-deterministic member. This postprocessing of an ensemble can be helpful to users that have limited computational power at their disposal but still want to benefit from the probabilistic information of the ensemble. Another application of the pseudo-deterministic product could be its visualisation as a map with locally the most probable precipitation scenario.

The pseudo-deterministic member is defined as the locally most probable forecast of the ensemble identified by a k-means-based clustering of the members in limited areas and blending the resulting best local forecasts back together. This leads to a mixed forecast that is supposed to have higher statistical skill averaged over the model domain (here, Germany) compared to the weather model’s deterministic member. In addition, the clustering can also be used for the identification of precipitation scenarios that differ significantly from the locally most probable precipitation. As a result, a reduced ensemble can be derived that contains a large part of the forecast distribution. In this poster, we present and evaluate the developed algorithm for the ensemble reduction and explore how skill depends on the number of reduced ensemble members in convective weather situations.