One way to analyse model structure uncertainty is the application of multi-model ensembles. The differences among the predicted model realizations in the ensemble provide an idea of the uncertainty from the hydrological models. In previous studies by Vansteenkiste et al. (2011) and Vansteenkiste et al. (2012) a hydrological impact analysis for a medium scaled catchment in Belgium, the Grote Nete, was performed using three lumped conceptual hydrological models (NAM, VHM and PDM) and using two different distributed model codes (WETSPA and MIKE-SHE). The different hydrological models were calibrated using exactly the same dataset. Subsequently, all calibrated models were used to assess the changes in extreme flows due to climate change.

In this study, the ability of using flexible model structures to create and evaluate multi-model ensembles in both a lumped as well as distributed way is investigated. Furthermore, the effect of equifinality is addressed and data uncertainties are taken into account in the evaluation of the multi-model ensemble.

For the lumped case, the VHM model (Willems, 2011) was used as a starting point and the application of automated calibration, based on specific performance criteria for low and high flow, was tested to verify the usefulness in calibrating the ensemble of model structures. The outcome of the automatic calibration was found comparable to the manual calibration of Vansteenkiste et al. (2011), in terms of statistical performance as well as for modelling specific summer and winter events. However, different parameterizations are obtained when optimizing towards different performance criteria (Nash-Sutcliffe efficiency and separate criteria focusing on high and low flows), indicating a lack of identifiability of the parameters.

The calibration scheme was then used to calibrate 24 different versions of the VHM model (slightly different in model structure). Comparison of the ensemble of model structures showed some advantages of adding complexity in the routing of overland flow or baseflow. However, caution is needed since, both in terms of performance measure and model structure, the parameter variations are striking, revealing the lack of parameter identifiability due to overparameterization and the compensation action between parameters introduced by the optimisation. No model structure really outperforms the others and analysis would probably benefit from more distinctive model structure hypotheses.

Furthermore, an implementation for developing a distributed model ensemble based on a flexible model framework is presented and applied in a similar model ensemble development and analysis. A grid-cell based discretization provides the ability to calculate water balances in each grid-cell with geographical heterogeneity using different model conceptualisations (infiltration, interception, soil moisture,...). A local drainage network is used to connect the individual cells and different routing conceptualisations can be combined and tested.

Moreover, due to the clear effect of the adopted performance measure on the resulting parameterization and its dependence on observed data, the uncertainty of the observed hydrograph can play a significant role in the evaluation of model structures. An evaluation approach taking into account the uncertainty of the discharge measurements based on the rating curve derivation is proposed and applied on the ensemble of optimized model structures. The approach can be used to evaluate models structures in relation to the observation error and to identify behavioural structure realizations. From this set of behavioural structures a predictive uncertainty of the ensemble of models can be derived. The results of the model residuals relative to the measured error indicate a similar systematic deviation of the VHM model ensemble and the NAM and PDM models used in Vansteenkiste et al. (2011).