Flexible framework for diagnosing alternative model structures through sensitivity and uncertainty analysis

Abstract

To explore the significance of alternative model structures and their inadequacies, hydrological modeling frameworks that allow quick implementation and comparison of alternative structures were already developed, mostly in a lumped mode. These systems allow testing the suitability of different model components and combining them in a modular fashion. Components can be modified or added if none of the available components fulfils the problem-specific requirements. It is, however, important to highlight that there is no general superior model structure for all spatial resolutions used. So, combining this flexible handling of structural components with varying the spatial resolution is necessary to adapt the model building process to specific conditions of the system, the available data and the objectives of the study. The next important step is to define effective strategies to diagnose and compare competitive model structures. Only then one can propose model structure improvements. There are several approaches described in the literature that help model diagnosis, like sensitivity analysis, parameter optimisation and uncertainty analysis, but these are mostly used to evaluate one specific structure or compare only a limited number of different structures and are typically not used in conjunction, but rather individually. By changing either specific processes or spatial resolution, while fixing the remainder of the model structure, rigorous testing of the model structure is possible, by addressing the effect of individual model components or spatial resolution. We present a tool to diagnose alternative model structures and address the effect of individual model components. The tool allows improving the evaluation and selection process of appropriate model structures out of the possible combinations coming from these flexible model structures to ensure the model represents the dominant processes of the system with the required rigour. The presented strategy uses both uncertainty and sensitivity analysis in a Monte-Carlo based framework. Regional sensitivity analysis allows identifying and comparing critical parameters among the different structures for different objective functions. Uncertainty analysis quantifies output uncertainty and parameter identifiability for different likelihood functions among the different structures. Comparing the posterior distribution of the parameters with the initial sampled distribution defines how these are conditioned by the model evaluation process. Since working with flexible structures, analysis can be done on both common and non-common components and associated parameters of the different model structures in a lumped or distributed mode. Structural components can be changed one at a time or a predefined set of model structures can be compared, using the combination of the above mentioned techniques. Selection criteria are assessed and linked to specific objectives (looking to specific flow regimes, specific objective functions or adapted metrics like flow duration curves). Moreover, it expresses the significant effect of the selected objective function(s) and the importance of using multiple evaluation criteria supporting the research question instead of only trying to reproduce the observed hydrograph.