Snow Model Verification Using Ensemble Prediction and Operational Benchmarks

Hydrologic model evaluations have traditionally focused on measuring how closely the model can simulate various characteristics of historical observations. Although advancing hydrologic forecasting is an often-stated goal of numerous modeling studies, testing in a forecasting mode is seldom undertaken, limiting information derived from these analyses. One can overcome this limitation through generation, and subsequent analysis, of ensemble hindcasts. In this study, long-range ensemble hindcasts are generated for the available period of record for a basin in southwestern Idaho for the purpose of evaluating the Snow– Atmosphere–Soil Transfer (SAST) model against the current operational benchmark, the National Weather Service’s (NWS) snow accumulation and ablation model SNOW17. Both snow models were coupled with the NWS operational rainfall runoff model and ensembles of seasonal discharge and weekly snow water equivalent (SWE) were evaluated. Ensemble predictions from both the SAST and SNOW17 models were better than climatology forecasts, for the period studied. In most cases, the accuracy of the SAST-generated predictions was similar to the SNOW17-generated predictions, except during periods of significant melting. Differences in model performance are partially attributed to initial condition errors. After updating the SWE state in the snow models with the observed SWE, the forecasts were improved during the first 2–4 weeks of the forecast window and the skills were essentially equal in both forecasting systems for the study watershed. Climate dominated the forecast uncertainty in the latter part of the forecast window while initial conditions controlled the forecast skill in the first 3–4 weeks of the forecast. The use of hindcasting in the snow model analysis revealed that, given the dominance of the initial conditions on forecast skill, streamflow predictions will be most improved through the use of state updating.