Predictability of precipitation-related hazards: insights from ensemble archives with advanced statistical methods
Loading...
Date issued
Authors
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Reuse License
Abstract
Accurate precipitation forecasts are essential for hydrology and weather-related risk management, yet they remain challenging even in convection-permitting numerical weather prediction. Limiting factors are the incomplete knowledge of initial conditions as well as simplifications in the representation of physical processes. While strategies to deal with initial condition uncertainty are well-established, isolating and representing model uncertainty remains challenging. This thesis examines practical predictability of hazardous precipitation in convection-permitting forecasting systems by extracting flow-conditioned systematic model behaviour and hazard-relevant information from long operational ensemble records. It combines two complementary perspectives, regime-dependent precipitation error structures over complex terrain and probabilistic hail guidance inferred from precipitation patterns when hail is not forecast explicitly.
In the first study, winter orographic precipitation errors over the Harz exhibit a robust low-dimensional structure that recurs under similar large-scale conditions. A principal component analysis (PCA) of precipitation differences between radar data (RADKLIM) and high-resolution numerical weather prediction data (COSMO-DE-EPS) identifies three dominant spatial modes that together explain roughly two thirds of the winter error variance. The modes represent ridge-centred amplification, synoptic placement mismatches with different spatial extent, as well as a mismatch between upstream and downstream precipitation amounts. ERA5 composites and the moist Froude number link these modes to distinct flow and stability regimes, while ICON case studies and semi-idealised experiments relate them to physically distinct precipitation responses under different terrain-flow regimes and show that microphysical perturbations mainly affect intensity rather than the dominant spatial error structure. The results indicate that isolating the underlying issues with the model representation remains challenging even when controlling for the large-scale flow configuration and when excluding cases where initial condition uncertainty (as represented in the ensemble system) controls forecast error.
In the second study, high-resolution precipitation patterns contain a reproducible signal that supports skilful probabilistic discrimination between hail-producing and non-hail convective situations, with best test accuracies of 85.91\% (CNN) and 85.71\% (RF) using RADKLIM precipitation field snapshots alone. Transferring this precipitation-based signal to COSMO-DE-EPS requires bias correction, and quantile mapping enables a physically plausible application to model precipitation fields. Verification against hail tracks shows strong sensitivity to displacement-tolerant neighbourhood settings in space and time. Under such settings, the COSMO--RF product reproduces the main hail-active regions and captures interannual variability. Analysis of dynamic, thermodynamic and aerosol (dust) conditions during mis-forecasts provides complementary context, but does not yield a single clear discriminator between hits, misses, and false alarms in the present setup.
Overall, the thesis shows that useful predictability can be extracted in two complementary ways. Over complex terrain, predictability emerges through the potential for flow-conditioned and physically interpretable structure in systematic precipitation error that may be applied to correct model precipitation fields. For hail, predictability emerges through a transferable hazard signal in precipitation patterns that can be converted into calibrated probabilities once systematic intensity biases are corrected, even though event-scale localisation remains limited.