Development of large-ensemble hydrology simulations in support of water resources planning and management

Abstract:

Water resource management decisions are made against a backdrop of uncertainty. To capture this model uncertainty, a move is currently afoot in the hydrologic community towards multi-model ensembles. However, the current generation of operational and experimental simulation platforms can be characterized as “small ensemble” systems. Typically, fewer than five hydrologic models are used to characterize the substantial uncertainty in simulation of hydrologic processes and the multi-model systems have a poor probabilistic portrayal of risk.

We have developed a hydrologic modeling approach, SUMMA (the Structure for Unifying Multiple Modeling Alternatives), that is built on a common set of governing equations and a common numerical solver. In developing SUMMA we recognize that most hydrologic models are based on the same set of governing equations that describe the temporal evolution of thermodynamic and hydrologic model states. As a result, different hydrologic models can be thought of as individual instantiations of a master modeling template. The difference among models relates to decisions on (1) the methods used to represent spatial variability (both throughout the model domain and within model elements); (2) the methods used to parameterize individual physical processes; (3) the methods used to estimate model parameters; and (4) the methods used to solve the model equations. Given this perspective, SUMMA provides a unified approach to hydrologic modeling that integrates different modeling methods into a consistent structure with the ability to instantiate alternative hydrologic models.

This presentation will illustrate the capabilities of SUMMA to quantify model uncertainty. First, we review results from an analysis of model shortcomings and provide a typology of the different sources of uncertainty in hydrologic modeling. Next, we summarize the extent to which SUMMA can mimic the behavior of existing models, through controlled test cases for a subset of model domains. This work on model mimicry guides the careful and deliberate selection of ensemble members, so that the ensemble has a comprehensive coverage of the model hypothesis space. Finally, we evaluate the statistical properties of large ensemble systems developed using SUMMA, to improve the probabilistic characterization of risk.