Understanding uncertainty contributions throughout the hydrologic modeling system in stochastic flood frequency analysis

Stochastic flood frequency (FF) estimation using hydrologic models is typically approached using Monte Carlo methods where the precipitation inputs, model parameters, and model initial conditions are varied for many simulations and return periods are then estimated empirically. However, the fractional uncertainty contributions of these components are not well understood across the full range of return periods used by agencies such as Reclamation (e.g. up to 1e6 or larger return periods). A more complete understanding of the component uncertainty contributions should help guide future R&D and subsequent investments, e.g. if event precipitation inputs are more important for the return period of interest, investment can be guided to improved precipitation event estimation techniques for a watershed to reduce uncertainty.

We explore these key components of the modeling chain by using a multi-model framework, the Framework for Understanding Structural Errors (FUSE), and varying model parameters using values consistent with historical meteorological uncertainty. We do this through the use of an ensemble precipitation and temperature dataset for the CONUS for developing an ensemble of internally consistent hydrologic model states and parameters for four FUSE model instances mimicking well-known hydrologic models such as PRMS, HEC-HMS, VIC, and SAC-SMA. We explore the differences in FF estimates for the different model structures given the exact same generation methodology, which is key to properly understanding model structural uncertainty. Stochastic event simulations are then performed for multiple model structures, across initial conditions and input parameters, with variable precipitation input distributions taken from Reclamation estimates.