USGS2001
When a model is calibrated by nonlinear regression, calculated diagnostic statistics and measures of uncertainty provide a wealth of information about many aspects of the system. This report presents a method of ranking the likely importance of existing observation locations using measures of prediction uncertainty. It is suggested that continued monitoring is warranted at more important locations, and unwarranted or less warranted at less important locations. The report develops the methodology and then demonstrates it using the hydraulic-head observation locations of a three-layer model of the Death Valley regional flow system. The predictions of interest are subsurface transport from beneath Yucca Mountain and 14 Underground Test Areas. The advective component of transport is considered because it is the component most affected by the system dynamics represented by the scale model being used. The problem is addressed using the capabilities of the U.S. Geological Survey computer program MODFLOW-2000, with its ADVective-Travel Observation (ADV) Package, and an additional computer program developed for this work. The methods presented in this report are used in three ways. (1) The ratings for individual observations are obtained by manipulating the measures of prediction uncertainty, and do not involve recalibrating the model. In this analysis, observation locations are each omitted individually and the resulting increase in uncertainty in the predictions is calculated. The uncertainty is quantified as standard deviations on the simulated advective transport. The increase in uncertainty is quantified as the percent increase in the standard deviations caused by omitting the one observation location from the calculation of standard deviations. In general, observation locations associated with larger increases are rated as more important. (2) Ratings for largely geographically based groups are obtained using a straightforward extension of the method used for individual observation locations. This analysis is needed where observations are clustered to determine whether the area is important to the predictions of interest. (3) Finally, the method is used to evaluate omitting a set of 100 observation locations. The locations were selected because they had low individual ratings and were not one of the few locations at which hydraulic heads from deep in the system were measured. The major results of the three analyses, when applied to the three-layer DVRFS ground-water flow system, are described in the following paragraphs. The discussion is labeled using the numbers 1 to 3 to clearly relate it to the three ways the method is used, as listed above. (1) The individual observation location analysis indicates that three observation locations are most important. They are located in Emigrant Valley, Oasis Valley, and Beatty. Of importance is that these and other observations shown to be important by this analysis are far from the travel paths considered. This displays the importance of the regional setting within which the transport occurs, the importance of including some sites throughout the area in the monitoring network, and the importance of including sites in these areas in particular. The method considered in this report indicates that the 19 observation locations that reflect hydraulic heads deeper in the system (in model layers 1, 2, and 3) are not very important. This appears to be because the locations of these observations are in the vicinity of shallow observation locations that also generally are rated as low importance, and because the model layers are hydraulically well connected vertically. The value of deep observations to testing conceptual models, however, is stressed. As a result, the deep observations are rated higher than is consistent with the results of the analysis presented, and none of these observations are omitted in the scenario discussed under (3) below. (2) The geographic grouping of the observations found one major area of importance not identified by the individual observation analysis. Five of the 49 groups are categorized as most highly important. The most important groups were those that, when omitted, produced mean increases greater than 10 percent at any UGTA site or Yucca Mountain. Four of the five groups were dominated by one individual observation. However, one group, located in Ash Meadows, had no individual observations ranked of high importance but collectively, when omitted, increased uncertainty substantially. Other groups also located in Ash Meadows, including intermediate depth observations, consistently ranked as more important than all other groups. (3) To demonstrate the importance of omitting a set of low-rated observations, one scenario is considered in which the 100 individually lowest-rated shallow and intermediate-depth observation locations are omitted. The measure of overall prediction uncertainty increased by just 0.59 percent, indicating that the wells associated with these observations probably could prudently be measured less frequently.