COMPARATIVE CALIBRATION OF A COMPLEX HYDROLOGICMODEL BY STOCHASTIC METHODS GLUE AND PEST

摘要

Hydrologic and environmental models are useful for prediction and understanding of processes. Model calibrationoften faces the problem of equifinality, by which empirical observations validate one set of model parameter values almostto the same degree as another very different set. This uncertainty in model parameter values implies uncertainty in modelpredictions. A stochastic approach to calibration is therefore appropriate to establish the likely range or probabilitydistribution of model parameters and predictions. In the literature, two commonly used methods are the GeneralizedLikelihood Uncertainty Estimation (GLUE) approach and the Parameter Estimation (PEST) software (a nonlinear parameterestimation and optimization package). This study makes a side‐by‐side comparison of GLUE and PEST applied to two models,the first a simple two‐parameter sinusoidal temperature model and the second a complex multi‐parameter multi‐purposehydrologic model, the Soil and Water Assessment Tool (SWAT) model, applied to the Salt Creek watershed in Illinois. TheSWAT model is calibrated for stream flow, corn and soybean yields, and nitrate load. The GLUE and PEST calibrations forthe simple model are straightforward. The same is not true, however, of the complex SWAT model, for which both GLUE andPEST are found to require some level of prior information to be effective. In this study, that information is obtained from adeterministic calibration of the system using a genetic algorithm (GA). The results indicate that generally there is a greaterflexibility in problem specification in GLUE than in PEST. This is desirable, although it also means that there is a greaterlevel of subjectivity. This flexibility, together with GLUE's independence from any assumption of model structure, makesGLUE suitable for calibrating large complex models where computational resources are available. On the other hand, forproblems where the presence of local optima is not significant, PEST is an attractive option, as it is able to identify the optimalset of adjustable model parameters at just a fraction of the computational cost of GLUE.