Evaluating Gridded Precipitation Analysis Forcings for the National Water Model

Initial conditions for the National Weather Service's (NWS) National Water Model (NWM) forecasts are established via NWM analysis cycles, which run once per hour. Input forcing data for the NWM are provided by external operational sources, including gridded analyses and numerical weather prediction (NWP) model forecasts. Given that the NWM is a hydrologic model, it is particularly important that the precipitation forcings used for its analysis cycles be unbiased and minimally erroneous. Of additional concern is the behavior of the data used in model calibration, since systematic differences in the behavior of calibration forcings and analysis forcings may propagate into the model through its calibration parameters.

To inform model development and operational configuration, the Office of Water Prediction (OWP) has evaluated a variety of precipitation analyses and short-term forecasts that contribute to NWM analysis cycles. These include Multisensor Precipitation Estimates (MPE) produced at NWS River Forecast Centers, radar-only and gauge-corrected precipitation estimates from the Multi-Radar Multi-Sensor (MRMS) system, and short-term precipitation forecasts produced by the Rapid Refresh (RAP) and High Resolution Rapid Refresh (HRRR) systems. In addition to NWM analysis precipitation inputs, this study examines NWM calibration data sources, including precipitation grids from the North American Land Data Assimilation System (NLDAS), used in calibration through NWM version 2.0, and from the OWP-generated Analysis of Record for Calibration (AORC), which will be used to calibrate future NWM versions.

This study examines the skill and bias of each of the above sources of gridded precipitation against observations collected by a variety of ground networks, including the Automated Surface Observing System (ASOS), the Hydrometeorological Automated Data System (HADS), the NWS Cooperative Observer Program (COOP), and several others.

This evaluation primarily covers the domain of the coterminous United States (CONUS) and the four-year period from October 2014 to October 2018. A variety of skill metrics are produced across both temporal and spatial dimensions, including Mean Absolute Relative Error, Log Multiplicative Bias, Probability of Detection, False Alarm Ratio, and correlation.