Selection and Use of Model Evaluation Metrics from Multi-decadal Hydrological Simulations to Support Operational Hydrological Forecasting

Continental scale hydrology presents forecasters with a number of challenges due to the diversity of dominant hydrological processes and limited observations across the full range of scales. The number of unique combinations of hydrological characteristics combined with the relative rarity of high impact events like major flooding and drought further complicate efforts to evaluate model performance at continental scale. Observations of volumetric streamflow from USGS gauge locations are the primary source of truth used to evaluate operational hydrological forecasts and simulations. Streamflow at a gauge is the result of processes influenced by a set of temporally and spatially varying upstream characteristics that may be unique to each gauged location. Streamflow observations also exhibit a large degree of temporal autocorrelation, but little spatial correlation outside of gauges within the immediate vicinity on the same stream. All these factors contribute to a dwindling source of independent and identically distributed observations that forecasters can use to meaningfully assess the uncertainty of operational forecasts.

This study presents an evaluation of the National Water Model (NWM) version 2.1 40-year Retrospective Simulation used by the Water Prediction and Operations Division of the Office of Water Prediction to inform the use of NWM forecasts when preparing continental scale flood guidance products. This evaluation incorporates resampling methods intended to reduce temporal autocorrelation, a concise set of model evaluation statistics, and methods used to quickly convey statistical sampling uncertainty both regionally and at individual gauges. The selected statistics reflect a measured approach meant to evaluate model output driven by observed precipitation, while allowing forecasters to efficiently judge model confidence in a particular region.