Ensemble Streamflow Forecasts Using Spatially Shifted QPF

Heavy rainfall events, which can cause flash flooding, are quite common in the U.S. Upper Midwest. Quantitative precipitation forecasts (QPF) are used in combination with hydrologic models to help provide an early warning of flash flooding. However, a single QPF lacks representation of the precipitation forecast uncertainty. Ensemble QPF represent the uncertainty in precipitation, however errors in precipitation amount and location can present challenges when applied for hydrologic prediction. In this work, we propose an approach to address the uncertainty in QPF associated with spatial displacement errors by systematically shifting an ensemble QPF to create additional ensemble members. We test ensemble QPF from the advanced National Oceanic and Atmospheric Administration (NOAA) High Resolution Rapid Refresh Ensemble (HRRRE) prediction system, which is currently in the experimental stage. The HRRRE consists of nine members with varying initial conditions and has 3-km horizontal grid spacing. We create an additional 72 ensemble members by shifting each of the nine HRRRE members in the cardinal (N, S, E, and W) and intermediate (NE, NW, SE, SW) directions. Shifting distances of 0.5^º and 1.0^º are used based on our past work of typical displacements for warm season precipitation events. The 81 ensemble QPF members are then input into the calibrated National Weather Service (NWS) Hydrology Laboratory-Research Distributed Hydrologic Model (HL-RDHM) to create an 81-member ensemble streamflow forecast. We are testing this method for select watersheds within the Upper Mississippi River Basin and comparing the performance of the 81-member ensemble to the original HRRRE 9-member ensemble (raw).

Results show that the systematic shifting of QPF increases the overall forecast spread compared to using the nine raw QPF outputs. This results in an improved containing ratio, a measure of how often the observation falls within the bounds of the ensemble. The 1^º shift had the highest score. Analysis of a simple two category forecast of above or below minor flood stage, shows that while the false alarm rate goes up with the spatially-shifted ensemble, the probability of detection goes up as well. Overall, we find an improvement in predicting the possibility of a minor flood event with the spatially-shifted ensemble. Although the spatial shifting appears to improve the categorical forecast for minor flooding, we found that the probabilistic forecasts derived from the ensemble do not score better than the raw ensemble for probabilistic metrics such as the ranked probability score. This is because the 81-member ensemble is skewed towards non-events as the precipitation footprint tends to be small for these events, and our basins are also small, so the forecasted precipitation often gets shifted outside the watershed bounds. To address this, we are experimenting with different weighting techniques to refine the probability of the ensemble forecast.