Abstract (eng)
Numerical weather prediction models have limited skill in the first few forecast hours. Therefore simple nowcasting methods, like the linear extrapolation of current radar information, are often used to predict the very near future more precisely. However, these extrapolation methods tend to rapidly decrease in skill after a few hours. Many efforts have been undertaken to combine recent observations or extrapolations of the latest states with numerical weather prediction (NWP) model computations. In order to achieve such a blending of current information and NWP forecasts, a data assimilation method is used in this thesis. The Local Ensemble Transform Kalman Filter (LETKF) is exploited to re-weight pre-existing ensemble forecasts according to recent observations, thus creating an adapted forecast, or rather, a nowcast. This LETKF-based nowcasting method is performed on the probability of hourly precipitation rates exceeding a threshold. The use of probabilities rather than precipitation intensities avoids unphysical jumps in the nowcasts. For a period of six days in July 2020, this LETKF-based nowcasting is performed on GeoSphere Austria’s 17-member C-LAEF ensemble. INCA analyses, which consist of a combination of recent radar observations and surface precipitation measurements, are used as observations. The nowcasts are, on average, able to outperform the underlying ensemble forecast for lead times of one to three hours, depending on the setting. For lead times of four hours and longer, however, the method in the presented setting can not outperform the ensemble forecast on average. Propagation of precipitation patterns outside of the localisation scale is found to be the main reason for decreasing nowcasting performance with increasing lead time. The method is tested for three different precipitation rates and performs better for lower intensities. Sensitivity studies are carried out to assess the influence of both the spatial resolution of the probability field and observation-error covariances. While the exact setting of observation errors does not affect the outcome significantly, results are better for coarser spatial resolutions.