Accounting for spatial displacement errors in HRRRE QPF to create short-term ensemble streamflow forecasts
Most of the heavy precipitation that falls in the Upper Mississippi Valley occurs during the period of May to September. Most of that warm season precipitation falls as a result of mesoscale convective systems. Because convective rainfall is usually localized and intense, it is critical to account for any model error resulting in spatial displacements. This is especially important when producing hydrologic forecasts as those displacements cause rainfall to be moved across watershed boundaries. Previous research attempted to shift quantitative precipitation forecasts (QPFs) as input to a hydrologic model to expand streamflow ensembles to better account for spatial displacements of QPF. Climatological spatial displacement errors have been computed for multiple convection-allowing ensembles, including the High-Resolution Rapid Refresh Ensemble (HRRRE), paving the way for the present work to apply those findings to hydrologic ensemble forecasts. A skewed-normal random number generator was used to select locations within the range of the climatological spatial displacement errors identified in the previous work to which the individual members of the HRRRE precipitation forecasts were shifted. Every HRRRE member was shifted 6 times, creating a full ensemble of 63 model members. The full ensemble of QPF and the original HRRRE members were used as forcing for the WRF-Hydro v5.1.1 running in a National Water Model 2.0 configuration. Evaluation of the magnitude of peak discharge, timing of peak discharge, and differing ensemble performance among small and large watersheds were carried out using 50 stream gauges, over 29 storm events. Verification of the magnitude of peak discharge was done by using containment ratio, ranked probability score, and reliability as metrics of probabilistic forecast skill. Deterministic evaluation was conducted using critical success index, equitable threat score, and Heidke skill score, at ten probability of exceedance thresholds. In addition to using equal weights among model members, two weighting schemes were added to modify the influence of ensemble members based on the observed spatial displacement of HRRRE QPF at convective initiation (CI). Previous research had calculated centroids, or center of mass of the precipitation systems, for the HRRRE QPFs and observed rainfall for 30 storm events in the Upper Mississippi region. The centroids of the modelled rainfall systems were compared to the centroids of observed precipitation areas where there was at least 1mm of rainfall at CI. Within the framework of this study, the centroids of HRRRE QPFs at CI, taken from the prior research, were compared to the centroids of the shifts produced by the random number generator. Weights would then be assigned as the inverse of the distance between those centroids. One of the CI displacement weighting schemes used a climatology-based correction to adjust the location of the displaced HRRRE QPF centroid. That correction adjusted the displacement of HRRRE QPF observed at CI to match the displacement of the centroid of accumulated precipitation more closely for the full forecast event. The second of the two weighting schemes disregarded the correction and did not adjust the magnitude of the spatial displacement of the HRRRE QPF when finding the inverse distance weights. The ensemble weighting methods allowed for the creation of two ensembles with 9 members and were designed to test an ensemble of similar size to the HRRRE, but with added information about the predicted location of QPF at CI. One of the two new ensembles was made by selecting the members that had the smallest distances between the centroid of the shifted QPF and the centroid of the CI displacements. The other new ensemble was similar; however, one member was chosen out of each parent HRRRE member group. All members in the new 9 member ensembles were given equal weight, and effectively zero weight was assigned to the members that were not selected for the smaller ensemble.