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

Abstract. This study evaluates the Weather Research and Forecasting (WRF) model, with and without WRFDA 3D-VAR data assimilation (DA), for precipitation and temperature forecasts in Luxembourg and the Greater Region during the July 2021 flood event. Conventional observations (CONV), Global Navigation Satellite System (GNSS) Zenith Total Delay (ZTD), and their combination (CONV + ZTD) were assimilated using a 6-hour rapid-update cycle over a full month (20 June–20 July 2021), and verified against independent surface stations withheld from the assimilation, together with radar and GPM IMERG data. For precipitation, DA improved categorical skill: CONV yielded the largest gains in bias (+37.4 %) and probability of detection (+18.3 %), at the cost of a higher false alarm ratio, while CONV + ZTD gave more moderate but balanced improvements. Absolute-error metrics (RMSE, MAE, SMAPE) changed little and mostly not significantly, indicating that DA improves event detection rather than magnitude. For temperature, all configurations reduced bias, ZTD being the most effective (+97.4 %). Station-level Wilcoxon signed-rank tests confirm that the bias and detection gains are statistically significant (p < 0.001). The results demonstrate the complementary roles of conventional and GNSS-based observations for regional numerical weather prediction.