摘要:This study evaluates the accuracy of short-range (1-h to 18-hr lead-time) forecasts from the High-Resolution Rapid Refresh (HRRR) model for five extreme storms in the United States: (1) the September 21–23, 2016, frontal storms in Iowa, (2) the April 28-May 1, 2017, frontal storms in the Southern Midwestern US, (3) the August 25–31, 2017, Hurricane Harvey storms in Texas, (4) the September 13–17, 2018, Hurricane Florence storms in the Carolinas, and (5) the September 4–6, 2019, Hurricane Dorian storms in the Carolinas. The evaluation was carried out by comparison with gauge-corrected Multi-Radar/Multi-Sensor (MRMS-GC) products. In terms of temporal variability, there was a good agreement between the forecasted and observed precipitation on an hourly basis. Thus, the HRRR products provide relatively reliable forecasts. However, the forecasts were mostly biased: they tend to overestimate rainfall for both hurricanes, underestimate rainfall for tropical storms in Iowa, and produce almost unbiased estimates for the frontal storms in Southern Midwestern US. In terms of spatial pattern, the forecasts are able to capture the spatial pattern of hurricanes, however, they produce too many, localized, high-rain intensities for the frontal storms than what the observations show. With regard to the effect of lead times, the 1-h lead forecasts have often lower accuracy than the other lead-time forecasts, while there was no much systematic difference in accuracy among the 2-h to 18-h lead-time forecasts. The bias estimates in the forecast are also examined at different spatial scales, ranging from 2 km × 2 km all the way to 128 km × 128 km. The results show that the bias estimated at smaller spatial scales vary within a large range, mostly within the range of −100% to 100%, indicating that the bias estimates obtained at large scale (hundreds of km grids) are not applicable to bias estimates at small scales, and vice versa. Local-bias correction approaches are therefore preferable over global bias-correction approaches.