The rainfall outputs from the latest convection-scale Weather Research and Forecasting (WRF) model are shown to provide an effective means of extending prediction lead times in flood forecasting. In this study, the performance of the WRF model in simulating a regional sub-daily extreme rainfall event centred over Beijing, China is evaluated at high temporal (sub-daily) and spatial (convective-resolving) scales using different domain configurations and spin-up times. Seven objective verification metrics that are calculated against the gridded ground observations and the ERA-Interim reanalysis are analysed jointly using subjective verification methods to identify the likely best WRF configurations. The rainfall simulations are found to be highly sensitive to the choice of domain size and spin-up time at the convective scale. A model run covering northern China with a 1 : 5 : 5 horizontal downscaling ratio (1.62 km), 57 vertical layers (less than 0.5 km), and a 60 h spin-up time exhibits the best performance in terms of the accuracy of rainfall intensity and the spatial correlation coefficient (R′). A comparison of the optimal run and the initial run performed using the most common settings reveals clear improvements in the verification metrics. Specifically, R′ increases from 0.226 to 0.67, the relative error of the maximum precipitation at a point rises from -56 to -11.7 %, and the root mean squared error decreases by 33.65 %. In summary, re-evaluation of the domain configuration options and spin-up times used in WRF is crucial for improving the accuracy and reliability of rainfall outputs used in applications related to regional sub-daily heavy rainfall (SDHR).