Evaluating the performance of a numerical weather prediction model for precipitation and temperature in Luxembourg and the Greater Region: insights from WRF and WRFDA 3D-VAR

We evaluate the Weather Research and Forecasting (WRF) model, with and without WRFDA 3D-VAR data assimilation, for precipitation and temperature forecasts in Luxembourg and the Greater Region during the July 2021 flood event. Conventional observations, Global Navigation Satellite System (GNSS) zenith total delay (ZTD) and their combination were assimilated 6-hourly over a month (20 June–20 July 2021), each cycle re-anchored to GFS so that all experiments share a common background. The climatological CV3 and an NMC-derived CV5 background error covariance were both tested, and forecasts verified against 264 independent surface stations in four countries. The covariance dominates the outcome. With CV3, assimilation raises rainfall frequency without improving discrimination and the equitable threat score falls from 0.383 to 0.337 at 0.1 mm per 6 h; with CV5 the same observations improve it, by +0.057 at 1 mm per 6 h. The gain is confined to low and moderate intensities, regionally uneven, and rests on an in-sample covariance. Adding ZTD to the conventional network gains +0.037, whereas adding conventional observations to ZTD gains nothing significant. For 2 m temperature, after height adjustment, no configuration is best. For this domain the background error covariance therefore matters more than the choice of observation type.