Analysis of uncertainty in hydrometeorological ensemble forecasts

Ensemble hydrometeorological forecasting has great potential for improving flood predictions and use in water management systems, however, the amount of data used and created with an ensemble forecast requires a careful and intentional approach to understand how useful and skillful the forecast is. The NOAA National Water Model (NWM) was run using downscaled NOAA Global Ensemble Forecast System (GEFS) meteorological forcings for the 2016-2017 wet season (October-March) in California to create an 11-member hydrologic forecast ensemble. To evaluate the performance of these ensemble forecasts, we chose to study streamflow sites within Sonoma County, California, a rain-dominated region which includes the Russian, Navarro, and Napa Valley River basins. The uncertainty in the downscaled GEFS precipitation and the NWM forced with the downscaled GEFS meteorological inputs was compared to deterministic North American Land Assimilation System (NLDAS) precipitation and a NLDAS-forced NWM run. For the purpose of our analysis, we verified the 5-day forecast lead time. The NLDAS and NLDAS forced-NWM was compared to precipitation and streamflow observations and found to be adequate for comparison purposes. The analysis included seasonal statistics and event-based performance characterization as a way to assess the skill in a statistically rigorous way, as well as provide insight into individual events that caused hydrologic impacts. Ensemble performance was evaluated using a “percent coverage” framework, defined as the percentage of time the observation falls within the middle 80% of the ensemble forecast. The percent coverage for the precipitation was consistently better than that of the streamflow. Rank histograms were created to characterize the ensemble distribution and its relationship to both NLDAS and the NLDAS-driven NWM. The rank histograms showed that the ensemble spread of the precipitation was fairly consistent and reliable, but showed a clear trend of overestimation (positive bias) in the streamflow forecasts. Correlations were made for several metrics to relate precipitation to streamflow, which showed that the changes in spread were well correlated, but the biases of the ensemble were not. The results indicate overall that the NWM may benefit from further evaluation and correction of biases, as well as a better understanding of how errors and uncertainties from meteorological forcings affect hydrologic model skill and ensemble spread