Automatic identification of systematic model failures in ensemble precipitation forecasts

Quantiative precipitation forecasting remains a major challenge even for kilometre-scale ensemble forecasating systems. However, operational high-resolution ensemble systems provide a large data-set from which - if combined with observational data - insight into systematic issues in the model physics can be gained. Here, we explore statistical methods to automatically identify systematic error patterns and their relation to the larger-scale conditions at example problem of precipitation at the Harz mountain range in northern Germany. For the analysis COSMO-D2-EPS forecasts for the years 2011-2018 are combined radar-derived and station-calibrated surface precipitation estimates provided by the German Weather Service (DWD). For the identification of common precipitation error patterns, empirical orthogonal function (EOF) analysis has been employed. For the winter season the leading order principal components show error features located on the elevated topography in the Harz region. Analysis of large-scale conditions (derived from ERA5) for each principal component shows systematic differences in upstream wind direction and speed, temperature, and specific humidity. In the summer seasons patterns are less localised, but some regional structure is maintained especially for the first principal component. Also the differentiation in large-scale conditions between EOFs is less. The challenges in summer are presumably related to a large contribution of convective precipitation. Overall, the leading 5 principal components explain 70,4% (48,2%) of the variance in winter (summer). To gain a better understanding of the relationship of error models to larger-scale conditions, as well as the physical mechanisms of model errors, simulations of precipitation at representative dates for principal components 1 and 2 were performed using the ICON-D2 model.