Spatially consistent postprocessing of precipitation over complex topography

Over the last decade statistical postprocessing has become a standard tool to reduce biases and dispersion errors of probabilistic numerical weather prediction (NWP) ensemble forecasts. Most established postprocessing approaches train a statistical model using raw ensemble statistics on a typically small set of stations.  While raw ensemble statistics are available from high resolution NWP grid data, observations are missing at most grid points. Hence, the generation of spatial fields of forecast scenarios requires both some kind of interpolation and reshuffling of forecast quantiles based on a dependence template. The most widely used reshuffling approach, ensemble copula coupling (ECC), applies a reordering based on the raw ensemble rank order structure. ECC relies on the assumption that the spatial dependence structure of the raw ensemble is spatially consistent with the observed fields. This assumption may not always hold for hourly precipitation in particular over complex topography, since even high resolution models do not achieve a perfect representation of the real topography.

In this study, hourly CombiPrecip fields, which are a blend of precipitation observations from station and radar data, at a spatial resolution of 1 km over Switzerland serve as observations. Hourly precipitation raw ensemble forecast fields covering lead times up to 120 hours with a spatial resolution of 2 km are provided by COSMO-E. This enables us to postprocess hourly  COSMO-E ensemble precipitation forecasts over Switzerland at different spatial scales, from a single global ensemble model output statistics type model, over regional quantile regression  models up to grid point-wise local analog models. The mismatch in spatial resolution between COSMO-E and CombiPrecip as well as  the general issue of non-representative model topography over Switzerland’s complex topography may affect the spatial consistency of the (postprocessed) forecast fields. Starting with an analysis of systematic errors and spatial consistency of COSMO-E precipitation forecasts , we assess the potential for spatially multivariate postprocessing approaches, which are able to incorporate the spatial information from CombiPrecip and are yet simple and computationally efficient. To this end, we analyse the effects of using standard and new postprocessing model designs that vary in the (analog-based) selection of training data, spatial aggregation, postprocessing model parametrizations, and methods to obtain physically realistic forecast scenarios in space.