The goal for this study was to examine the performance of quantitative precipitation forecasting (QPF) obtained from a high resolution convection-allowing model and two coarser resolution operational weather prediction models to better understand any QPF improvements in the convection-allowing runs. The ARW-WRF model was run over the period from March through November 2013 with 4 km grid spacing to better understand the limits of predictability of short-term (12 h) QPF that might be used in hydrology models. WRF runs were performed using NAM and GFS output as the first guess fields in the 3-dimensional variational data assimilation system. Radar data were assimilated in WRF runs. Several verification methods were used to compare the QPF from the high-resolution runs with coarser operational GFS and NAM QPF. Three traditional grid-to-grid verification methods, as well as two spatial techniques, neighborhood and object-based, were used to verify QPF for 1h, 3h, 6h and 12h precipitation accumulation intervals and two grid configurations.

In general, skill increased more as accumulation interval increased than for spatial scale increasing. At the same neighborhood scale, the grid spacing on which the verifications were done had less impact on the high resolution WRF model than the coarser models. NAM had the worst performance not only for model skill but also for spatial features due to the existence of large dry bias and location errors. Even for some severe floods with large rain coverage, NAM still underpredicted the magnitude of the total rain volume. Moreover, the finer resolution of NAM did not offer any advantages in predicting small-scale storms compared to the coarser GFS model. Both neighborhood and traditional techniques suggested that WRF had much higher skill for larger precipitation thresholds. In addition, WRF had the smallest displacement errors and was able to most correctly forecast the intensity magnitude of simple precipitation objects. The total rain volume obtained from WRF for severe floods is the closest to the observations though WRF still underestimated it. All of the models had the best performance from midnight to early morning, because the least wet bias, location and coverage errors were present then. The lowest skill happened from late morning through afternoon. The major challenge for skill improvement during this period was large displacement errors. The displacement errors started to grow in late morning and reached peak values around late afternoon.

QPF for river basins had higher biases than for the full domain. WRF overpredicted the total rain volume for all river basins and diurnal time periods examined. The skill for the largest basin shared similar characteristics with the full domain, but the smaller basins had much larger discrepancies.