Evaluation of different calibration strategies for large scale continuous hydrological modelling

For the analysis of climate impact on flood flows and flood frequency in macroscale river basins, hydrological models can be forced by several sets of hourly long-term climate time series. Considering the large number of model units, the small time step and the required recalibrations for different model forcing an efficient calibration strategy and optimisation algorithm are essential. This study investigates the impact of different calibration strategies and different optimisation algorithms on the performance and robustness of a semi-distributed model. The different calibration strategies were (a) Lumped, (b) 1-Factor, (c) Distributed and (d) Regionalisation. The latter uses catchment characteristics and estimates parameter values via transfer functions. These methods were applied in combination with three different optimisation algorithms: PEST, DDS, and SCE. In addition to the standard temporal evaluation of the calibration strategies, a spatial evaluation was applied. This was done by transferring the parameters from calibrated catchments to uncalibrated ones and validating the model performance of these uncalibrated catchments. The study was carried out for five sub-catchments of the Aller-Leine River Basin in Northern Germany. The best result for temporal evaluation was achieved by using the combination of the DDS optimisation with the Distributed strategy. The Regionalisation method obtained the weakest performance for temporal evaluation. However, for spatial evaluation the Regionalisation indicated more robust models, closely followed by the Lumped method. The 1-Factor and the Distributed strategy showed clear disadvantages regarding spatial parameter transferability. For the parameter estimation based on catchment descriptors as required for ungauged basins, the Regionalisation strategy seems to be a promising tool particularly in climate impact analysis and for hydrological modelling in general.