Effect of (quasi-)optimum model parameter sets and model characteristics on future discharge projection of two basins from Europe and Asia

Uncertainty is an inherent, unavoidable feature in the modeling of natural processes. This is particularly a sensitive issue when dealing with forecasting, especially in the context of climate change impacts. Apart from the uncertainty introduced by different climate projections, additional sources of uncertainty have been acknowledged in the literature: driving models, input information, regionalization, parameter choice, downscaling techniques, among others. In this study, we focus as primary goal on the uncertainty introduced by various set of parameters in the twenty-first century projections of runoff in two large river basins: the Rhine in Europe and the Ganges in Asia. For this purpose, we apply a robust parameter estimation optimization algorithm to account for the uncertainty given by a quasi-optimum parameter set choice. A total of 1000 robust and well-performing parameter sets are found and applied for uncertainty analysis. Here we analyze how much discharge projections diverge, given sets of parameters with similar performance. Results suggest that parameter uncertainty is strongly related to model complexity in both basins. To contrast this uncertainty with other important sources, five general circulation models together with four climate change emission scenarios (Representative Concentration Pathways, RCP) are analyzed. Two hydrological models are used as well to test the impact of model conception on uncertainty. Results show that the contribution of parameter choice to uncertainty for the Ganges is rather stable in time and comparatively small for the periods 2006 to 2035, 2036 to 2065, and 2070 to 2099. In the case of Rhine, results are more heterogeneous and change over time, with increasing importance of GCM/RCPs toward the end of the century. Major differences are attributable to GCMs ranging from 60 to 80% followed by RCPs in the range 12–30%, whereas differences due to parameter sets range from 3 to 8%.