High-resolution climate simulations for Central Europe: An assessment of dynamical and statistical downscaling techniques

To bridge the resolution gap between the outputs of global climate models (GCMs) and finer-scale data needed for studies of the climate change impacts, two approaches are widely used: dynamical downscaling, based on application of regional climate models (RCMs) embedded into the domain of the GCM simulation, and statistical downscaling (SDS), using empirical transfer functions between the large-scale data generated by the GCM and local measurements. In our contribution, we compare the performance of different variants of both techniques for the region of Central Europe. The dynamical downscaling is represented by the outputs of two regional models run in the 10 km horizontal grid, ALADIN-CLIMATE/CZ (co-developed by the Czech Hydrometeorological Institute and Meteo-France) and RegCM3 (developed by the Abdus Salam Centre for Theoretical Physics). The applied statistical methods were based on multiple linear regression, as well as on several of its nonlinear alternatives, including techniques employing artificial neural networks. Validation of the downscaling outputs was carried out using measured data, gathered from weather stations in the Czech Republic, Slovakia, Austria and Hungary for the end of the 20th century; series of daily values of maximum and minimum temperature, precipitation and relative humidity were analyzed. None of the regional models or statistical downscaling techniques could be identified as the universally best one. For instance, while most statistical methods misrepresented the shape of the statistical distribution of the target variables (especially in the more challenging cases such as estimation of daily precipitation), RCM-generated data often suffered from severe biases. It is also shown that further enhancement of the simulated fields of climate variables can be achieved through a combination of dynamical downscaling and statistical postprocessing. This can not only be used to reduce biases and other systematic flaws in the generated time series, but also to further localize the RCM outputs beyond the resolution of their original grid. The resulting data then provide a suitable input for subsequent studies of the local climate and its change in the target region.