A non-stationary climate-informed weather generator for assessing future flood risks

Abstract

We present a novel non-stationary regional weather generator (nsRWG) based on an auto-regressive process and marginal distributions conditioned on climate variables. We use large-scale circulation patterns as a latent variable and regional daily mean temperature as a covariate for marginal precipitation distributions to account for dynamic and thermodynamic changes in the atmosphere, respectively. Circulation patterns are classified using ERA5 reanalysis mean sea level pressure fields. We set up the nsRWG for the central European region using data from the E-OBS dataset, covering major river basins in Germany and riparian countries. The nsRWG is meticulously evaluated, showing good results in reproducing at-site and spatial characteristics of precipitation and temperature. Using time series of circulation patterns and the regional daily mean temperature derived from general circulation models (GCMs), we inform the nsRWG about the projected future climate. In this approach, we utilize GCM output variables, such as pressure and temperature, which are typically more accurately simulated by GCMs than precipitation. In an exemplary application, the nsRWG statistically downscales precipitation from nine selected models from the Coupled Model Intercomparison Project Phase 6 (CMIP6), generating long synthetic but spatially and temporally consistent weather series. The results suggest an increase in extreme precipitation over the German basins, aligning with previous regional analyses. The nsRWG offers a key benefit for hydrological impact studies by providing long-term (thousands of years) consistent synthetic weather data indispensable for the robust estimation of probability changes in hydrologic extremes such as floods.