Tailoring Hydrologic Modeling for Improved Water Resources Decision Support: A Mixed Ensemble Approach

Abstract

Water resources decisions are often made in the context of compromise among stakeholder groups with very different interests, and the size and characteristics of aquifers is such that the impacts of present-day use may take decades to manifest and equally long to mitigate. Initially small uncertainties (e.g. the water level in a stream after 20 years of groundwater pumping from a nearby housing development) can become magnified over time, proportionately increasing the environmental and monetary costs of a miscalculated decision. Hydrologic models help predict consequences but are limited by sparse data and uncertainty. This combination of amplified responses over time, long time-scale impacts, and multiple concerned stakeholders suggests a need for multiple models, and it is worthwhile to pay special attention to less probable, still plausible models that predict consequential outcomes. In this thesis, these models are called models of concern (MOCs). We propose a method of combining two ensembles: one is composed of the best-fitting calibrated models, and another entirely of MOCs, to consider the importance of consequential outcomes. The importance of each model’s prediction to the stakeholder is defined through a utility function. Predictions may have low or high utility according to the associated consequences, and a stakeholder sets a utility threshold to express their level of risk tolerance. The mixed ensemble is formed through an iterative process that allows the stakeholder to consider, and reconsider, their willingness to accept risk according to the likelihood of a consequential outcome. This process represents stakeholder concerns more fully than a model, or even a single ensemble, which only considers goodness-of-fit. Preliminary results suggest that the mixed ensemble of best-fitting and MOC models increases the identification of consequential outcomes without overstating the potential for negative impacts. Future work is needed to test this method against more complex systems to integrate it with current methods of automated parameter estimation. This work considers the need for, and implications of, a new approach to ensemble modeling to better represent stakeholder concerns.