Learning from the National Water Model: Regional Improvements in Streamflow Prediction through Experimental Parameter and Physics Updates to the WRF-Hydro Community Model

The National Weather Service's Office of Water Prediction now provides real-time, spatially and temporally continuous hydrologic prediction over the contiguous U.S. (CONUS) through the National Water Model (NWM). The NWM provides real-time estimates and forecasts out to 30 days of streamflow across 2.7 million stream reaches as well as distributed snowpack, soil moisture, and evapotranspiration at 1-km resolution. It is built on the WRF-Hydro community modeling system, which couples state-of-the-art land surface models with a suite of additional process-based hydrologic modules, adding capacity to simulate lateral redistribution of surface and subsurface water, groundwater dynamics, and channel routing. This hydrologic model implementation over a region as diverse as CONUS has facilitated the rapid evaluation of commonly used hydrologic process representations and parameter assumptions, as well as identification of where these assumptions fail.

In this study, we ask the question: to improve streamflow simulation in the NWM, where does standard model calibration suffice and where do we need process-level changes? We focus initially on processes impacting soil water movement. We use a multi-year retrospective CONUS-scale WRF-Hydro model run to identify hydrologically poorly performing regions. Based on error analysis and regional characteristics, we identify regions where model errors are likely due to shortcomings in surface and subsurface drainage dynamics (vs. forcings, snowpack dynamics, etc.). We then isolate representative basins in each of these regions and evaluate the impacts of standard soil parameter calibration, non-standard calibration against remote sensing products, and additional physical model complexity. Specifically, we test three levels of model improvements: (1) improved soil moisture drainage behavior through standard soil parameter calibration against streamflow, (2) incorporation of spatially variable surface detention storage calibrated against remotely-sensed inundation products, and (3) more accurate deep groundwater and soil water exchanges through the use of a 2-D coupled, process-based groundwater model. We quantify how these model improvements impact streamflow prediction in the various problem regions, and make recommendations for extrapolating these watershed-based findings to the full CONUS implementation of the NWM as well as other large-scale hydrologic modeling applications.