A method for evaluating the importance of system state observations to model predictions, with application to the Death Valley regional groundwater flow system

Abstract

[1] We develop a new observation-prediction (OPR) statistic for evaluating the importance of system state observations to model predictions. The OPR statistic measures the change in prediction uncertainty produced when an observation is added to or removed from an existing monitoring network, and it can be used to guide refinement and enhancement of the network. Prediction uncertainty is approximated using a first-order second-moment method. We apply the OPR statistic to a model of the Death Valley regional groundwater flow system (DVRFS) to evaluate the importance of existing and potential hydraulic head observations to predicted advective transport paths in the saturated zone underlying Yucca Mountain and underground testing areas on the Nevada Test Site. Important existing observations tend to be far from the predicted paths, and many unimportant observations are in areas of high observation density. These results can be used to select locations at which increased observation accuracy would be beneficial and locations that could be removed from the network. Important potential observations are mostly in areas of high hydraulic gradient far from the paths. Results for both existing and potential observations are related to the flow system dynamics and coarse parameter zonation in the DVRFS model. If system properties in different locations are as similar as the zonation assumes, then the OPR results illustrate a data collection opportunity whereby observations in distant, high-gradient areas can provide information about properties in flatter-gradient areas near the paths. If this similarity is suspect, then the analysis produces a different type of data collection opportunity involving testing of model assumptions critical to the OPR results.

Sensitivity analysis, calibration, and testing of a distributed hydrological model using error‐based weighting and one objective function

“…The idea of using sensitivity analysis to decide what parameters to estimate using regression and the essential related idea that unestimated parameters (also called the parameter null space and fixed unessential factors) can be important to measures of prediction uncertainty appear in the literature from many fields. For example, see Sobol' [1993], Anderman et al [1996], Hill [1998], Tiedeman et al [2004], Moore and Doherty [2005], Hill and Tiedeman [2007] and Sobol' et al [2007]. This work focuses on using sensitivity analysis to design a productive regression and identify important observations.…”

Comparative analysis of formulations for conservative transport in porous media through sensitivity-based parameter calibration

“…Sensitivity of model response to these parameters typically varies in space and time. An important step in a parameter estimation procedure is to identify locations in the system where the model is most sensitive to its parameters [e.g., Tiedeman et al ., ]. This, in turn, constitutes the basis for model‐based experiment design and interpretation [e.g., Fajraoui et al ., , and references therein].…”

Global sensitivity analysis in hydrological modeling: Review of concepts, methods, theoretical framework, and applications

“…However, we also argue that SA is a useful perspective for conceptualizing and understanding hydrological models for several reasons. As indicated by Rakovec et al (2014), SA can be used to: (a) detect when increasing model complexity can no longer be supported by observations and whether it is likely to affect model predictions (e.g., Saltelli et al, 1999;van Werkhoven et al, 2008a;Doherty and Welter, 2010;Rosolem et al, 2012;Gupta et al, 2012;Foglia et al, 2013); (b) reduce the time required for model calibration by focusing estimation efforts on parameters that are important for calibration metrics and predictions (e.g., Anderman et al, 1996;Hamm et al, 2006;Zambrano-Bigiarini and Rojas, 2013); (c) determine the priorities for theoretical and site-specific model development (e.g., Hill and Tiedeman, 2007;Saltelli et al, 2008;Kavetski and Clark, 2010); and (d) identify the advantageous placement and timing of new measurements (e.g., Tiedeman et al, 2003Tiedeman et al, , 2004.…”