Options
Model complexity and diagnostic-tool based analyses of integrated and physically based models
Auteur(s)
Ghasemizade, Mehdi
Date de parution
2016
Mots-clés
Résumé
The proper management of water resources nowadays is a critical issue. In that sense, accurate measurement of water balance components is a prerequisite for the proper management of water resources since one cannot manage what one cannot measure. Due to the difficulty in direct measurements of some of the water balance components such as deep percolation, simulation models are applied. Recent increases in computational power have motivated the application of more complex models of coupled environmental processes. These models, however, require outnumbered parameters, which lead to the problem of over-parameterization, meaning that many different parameter sets can lead to identical fits to the observed data. Therefore, this study explores the application of integrated and physically-based model HydroGeoSphere (HGS) in the framework of a weighing lysimeter in north-east of Switzerland to pursue: I) comparing the performance of different levels of complexity (in terms of the number of parameters) for simulating daily water balance components (actual evapotranspiration, water content, and lysimeter discharge) where three model concepts were introduced; II) addressing the output uncertainty of each concept at different time scales; III) application of a global and temporal sensitivity analysis as a diagnostic tool to address how individual parameters of the model as well as their interactions can affect the output uncertainty; VI) using a time-varying identifiability analysis method to investigate when the maximum amount of information about model parameters can be derived, considering the available data. The results of the study indicated that the most complex concept outperformed the other simpler concepts in reproducing the daily water balance components based on the performance metrics of R<sup>2</sup> and RMSE. However, the ideal required level of complexity, when considered in terms of output uncertainty, was shown to be dependent on the time scales of the simulated outputs. Exploring the results of the sensitivity analysis revealed that the individual effects of model parameters as well as their interaction effects on model outputs are required to be analyzed simultaneously to allow for the reduction in output uncertainty. The identifiability analysis indicated that identifiability is a necessary but not sufficient condition for a parameter to allow for reduction in the model output uncertainty. Overall our research indicated that, based on the available data at the lysimeter scale, complex and integrated models, such as HGS, are attractive solutions to reproduce complex features of the system but they have the severe difficulties of parametrization, leading to their reduced predictive capabilities.
Notes
Thèse de doctorat : Université de Neuchâtel, 2016
Identifiants
Type de publication
doctoral thesis
Dossier(s) à télécharger
