An Approach to Create Probabilistic Streamflow Forecasts from HRRRE Probabilistic Quantitative Precipitation Forecasts

Floods are the most frequent natural disasters that occur in the United States, causing the loss of around a hundred lives and billions of dollars annually. The biggest driver for modeling flooding events are the atmospheric inputs, specifically rainfall. Unfortunately, precipitation forecasts have generally been thought to contain too much error to be used as reliable inputs, and thus, flood forecasts are often not made until after precipitation has already fallen. The accuracy of atmospheric model forecasts continues to improve, both in timing and location, suggesting there may be ways to provide flood forecasts earlier by using the rainfall forecasts. One challenge, however, is that warm season precipitation which often leads to flooding is typically associated with relatively small-scale thunderstorms, and errors in timing and location are still usually large enough to have serious impacts on flood forecasts within small basins. One possible way to mitigate these errors is to use an ensemble of models since the ensemble provides a measure of uncertainty which is usually conveyed through probabilistic forecasts, and the mean, at least for many parameters, is typically more skillful than the forecast from any individual member.

The present study examines whether skillful probabilistic streamflow forecasts can be obtained using the High Resolution Rapid Refresh Ensemble (HRRRE) probabilities of precipitation exceeding specified threshold accumulations. The HRRRE consists of nine members providing forecasts over the continental United States on a 3 km horizontal grid. A key component of this work is the creation of expected rainfall amounts at every grid point for probability of exceedance values of 5, 10, 25, 50, 75, 90, and 95 percent. The rainfall amounts for each of the probability of exceedance values were calculated using bilinear interpolation from the probabilistic quantitative precipitation forecast (PQPF) generated by the HRRRE at 0.5 inch, 1 inch, 2 inches, and 3 inches, which is based on a neighborhood technique applied by the HRRRE developers. The grid point precipitation amounts associated with the probability of exceedance values are then input into a hydrologic model.

To be able to compare to the River Forecasting Center forecasts, the hydrologic model used in this study was the Sacramento Soil Moisture Accounting Model (SAC-SMA), and it was applied to 12 different river basins across the upper Midwest during the months of May, June, July, and August of 2017 and 2018. Preliminary results from the model runs have shown that the measured USGS discharge values most frequently match the HRRRE-based prediction, 39% of all cases, when the probability of exceedance was in the lowest quartile, 0-25%. The second most frequent match happens at the other end of the spectrum where 27% of the events had an exceedance probability in the 75-100% quartile. These findings are counterintuitive if one expects the measured discharges to most often match the predictions based on probability of exceedances of roughly 50%. There are at least four possible reasons to explain this discrepancy: (1) HRRRE is under dispersive, (2) the neighborhood approach used to create PQPF from the HRRRE has shortcomings, (3) the technique to determine rainfall amounts at given exceedance probabilities should be adjusted, or (4) errors are present in the SAC-SMA model. It is well-known that high resolution ensembles are often under dispersive, so at least some of the behavior is likely caused by that. Also, the current neighborhood approach to determine PQPF may be over sampling the neighboring grid points. Next, with the limited number of PQPF values, some external boundary values were added to run the bilinear interpolation of PQPF thereby possibly altering the probability of exceedance values and discharge forecasts. In addition, results are sensitive to the initial conditions of the SAC-SMA. Forecasters at the River Forecasting Centers regularly adjust the model to be in better agreement with observed values, a practice that we did not attempt to reproduce in this study. Future work may examine differences in PQPF values produced from HRRRE and the now operational HREF (High Resolution Ensemble Forecast), explore alterations in the techniques of determining probabilities of exceedance, and assess the variations in the SAC-SMA model with observed rainfall to better understand forecast errors associated wtih the SAC-SMA model.