Improving Medium-Range Streamflow Forecasting Across the U.S. Middle Atlantic Region

摘要

Short- to medium-range (forecast lead times from 0 to 14 days) streamflow forecasts are subject to uncertainties from various sources. A major source of uncertainty is due to the weather or meteorological forcing. In turn, the uncertainties from the meteorological forcing are propagated into the streamflow forecasts when using the meteorological forecasts (i.e., the outputs from a Numerical Weather Prediction (NWP) model) as forcing to hydrological models. Additionally, the hydrological models themselves are another important source of uncertainty, where uncertainty arises from model structure, parameters, initial and boundary conditions. To advance the science of hydrological modeling and forecasting, these uncertainties need to be quantified and modeled, using novel statistical techniques and robust verification strategies, with the goal of improving the skill and reliability of streamflow forecasts. This, ultimately, may allow generating in advance (i.e., with longer lead times) more informative forecasts, which could eventually translate into better emergency preparedness and response.;The main research goal of this dissertation is to develop, implement and verify a new regional hydrological ensemble prediction system (RHEPS), comprised by a numerical weather prediction (NWP) model, different hydrological models and different statistical bias-correction techniques. To implement and verify the new RHEPS, the U.S. middle Atlantic region (MAR) is selected as the study area. This is a region of high socioeconomic value with populated cities and, at the same time, vulnerable to floods and other natural disasters. To meet my research goal, the following objectives are carried out: Objective 1 (O1) - To choose a relevant NWP model or system by evaluating and verifying the outputs from different meteorological forecasting systems (i.e., the outputs or forecasts from their underlying NWP models); Objective 2 (O2) - To verify streamflow forecasts generated by forcing a distributed hydrological model with meteorological ensembles, and to develop and evaluate a statistical postprocessor to quantify the uncertainty and adjust biases in the streamflow forecasts; Objective 3 (O3) - To develop, implement and rigorously verify a multimodel approach for short- to medium-range streamflow forecasting. The overarching hypothesis of this dissertation is that the combination and configuration of the different system components in the streamflow forecasting system can have a significant influence on forecast uncertainty and that hydrological multimodeling is able to significantly enhance the quality of streamflow forecasts. The RHEPS is used to test this hypothesis.;To meet O1, precipitation ensemble forecasts from two different NWP models are verified. The two NWP models are the National Centers for Environmental Prediction (NCEP) 11-member Global Ensemble Forecast System Reforecast version 2 (GEFSRv2) and the 21-member Short Range Ensemble Forecast (SREF) system. The verification results for O1 reveal the quality of the meteorological forcing and serve to inform the decision of selecting a NWP model for O2. As part of O2, the meteorological outputs from the GEFSRv2 are used to force the NOAA's Hydrology Laboratory-Research Distributed Hydrological Model (HL-RDHM) and generate short- to medium-range (1-7 days) ensemble streamflow forecasts for different basins in the MAR. The streamflow forecasts are postprocessed (bias-corrected) using a time series model. The verification results from O2 show that the ensemble streamflow forecasts remain skillful for the entire forecast cycle of 7 days. Additionally, postprocessing increases forecast skills across lead times and spatial scales, particularly for the high flow conditions. Lastly, with O3, a multimodel hydrological framework is tested for medium-range ensemble streamflow forecasts. The results show that the multimodel consistently improves short- to medium-range streamflow forecasts across different basin sizes compared to the single model forecasts.