Generation of WRF-Hydro Probabilistic Streamflow Forecasts by Shifting Ensemble QPF Based on a Climatology of Forecast Rainfall Displacement Errors

Errors associated with precipitation location in Quantitative Precipitation Forecasts (QPF) present challenges when using for hydrologic prediction. Streamflow predictions are watershed-based, and erroneous shifts in the location of predicted precipitation can cause an over-prediction in some basins and under-prediction in others. Unfortunately, QPF skill is lowest during the warm season when the heaviest rain events occur in much of the United States, often causing flooding. A key reason for the low skill is the small-scale intense nature of warm season rainfall and the displacement errors that are typical in the positioning of convective systems in the models used to provide QPF. Our prior work had shown that a random shifting of ensemble member QPF to try to account for displacement errors improved the containing ratio for high streamflow events simulated using the Sacramento Soil Moisture Accounting hydrologic model over that obtained using QPF from the raw HRRRE members, but Ranked Probability Scores did not improve. That result suggests a better-informed method of shifting the QPF, perhaps based on a climatology of displacement errors, may be needed to significantly improve streamflow forecasts.

To further explore methods to account for displacement errors and therefore reduce errors for streamflow prediction, we examine two convection-allowing model ensembles, the High Resolution Rapid Refresh Ensemble (HRRRE) and the High Resolution Ensemble Forecast (HREF), for 30 cases where precipitation warranted a flash flood watch or warning within the North Central River Forecasting Center area of responsibility. The center of mass for the initiation hour and the 18-hour accumulation of each model’s QPF are compared to the center of mass of the observed Quantitative Precipitation Estimates (QPE) for the same time intervals. Correlations between the displacement at initiation and at the 18-hour accumulation are examined as a possible guide for precipitation forecast pre-processing prior to running a streamflow forecast. Analysis of displacement errors between QPF and QPE revealed a somewhat systematic displacement in total precipitation to the west for HRRRE, and no preferred directional shift with HREF. Analysis for relationships between initiation hour displacement and displacement of the total accumulated rainfall yielded mixed results, with crude methods of subtracting initiation hour displacement errors from accumulation displacement errors generally making displacement errors larger. However, dividing the dataset based on the general intercardinal direction of the initiation hour displacement, calculating individual weights for latitude and longitude for each, and then subtracting the initiation hour displacement resulted in some improvement of the displacement error of the full period precipitation.

Both the original random shifting approach using half degree and one-degree latitude and longitude shifts, along with approaches using different weighting based on our climatological study are being tested using each member of HRRRE, with the resulting QPF fed into the WRF-Hydro model. All the necessary fields for the NoahMP LSM are included in the shifting such as, shortwave and longwave radiation, surface pressure, temperature, wind, and specific humidity, as well as the primary driver, the QPF. 104 USGS gauge points located within the upper two thirds of Iowa and the Des Moines river watershed in Minnesota are incorporated into the WRF-Hydro to allow for a large sample size when investigating model output. By taking model displacement errors into account using weighting of the hydrologic model outputs, this should diminish errors in hydrologic forecasts associated with precipitation uncertainty, leading to applications in further research and operations.