Comparison Of Uncertainty Analysis For Community Based Watershed Models PDF Download

This thesis provides a critique and evaluation of the Generalized Likelihood Uncertainty Estimation (GLUE) methodology, and provides an appraisal of sensitivity analysis methods for watershed models with calibrated parameters. The first part of this thesis explores the strengths and weaknesses of the GLUE methodology with commonly adopted subjective likelihood measures using a simple linear watershed model. Recent research documents that the widely accepted GLUE procedure for describing forecasting precision and the impact of parameter uncertainty in rainfall-runoff watershed models fails to achieve the intended purpose when used with an informal likelihood measure (Christensen, 2004; Montanari, 2005; Mantovan and Todini, 2006; Stedinger et al., 2008). In particular, GLUE generally fails to produce intervals that capture the precision of estimated parameters, and the distribution of differences between predictions and future observations. This thesis illustrates these problems with GLUE using a simple linear rainfall-runoff model so that model calibration is a linear regression problem for which exact expressions for prediction precision and parameter uncertainty are well known and understood. The results show that the choice of a likelihood function is critical. A likelihood function needs to provide a reasonable distribution for the model errors for the statistical inference and resulting uncertainty and prediction intervals to be valid. The second part of this thesis discusses simple uncertainty and sensitivity analysis for watershed models when parameter estimates result form a joint calibration to observed data. Traditional measures of sensitivity in watershed modeling are based upon a framework wherein parameters are specified externally to a model, so one can independently investigate the impact of uncertainty in each parameter on model output. However, when parameter estimates result from a joint calibration to observed data, the resulting parameter estimators are interdependent and different sensitivity analysis procedures should be employed. For example, over some range, evaporation rates may be adjusted to correct for changes in a runoff coefficient, and vice versa. As a result, descriptions of the precision of such parameters may be very large individually, even though their joint response is well defined by the calibration data. These issues are illustrated with the simple abc watershed model. When fitting the abc watershed model to data, in some cases our analysis explicitly accounts for rainfall measurement errors so as to adequately represent the likelihood function for the data given the major source of errors causing lack of fit. The calibration results show that the daily precipitation from one gauge employed provides an imperfect description of basin precipitation, and precipitation errors results in correlation among flow errors and degraded the goodness of fit.