Limiting factors in model performance for the multi-objective calibration of a pan-European, semi-distributed hydrological model for discharge and sediments

Abstract

An inevitable challenge faced during calibration of large-scale hydrological models is the task of reducing equifinality of various model structures and parameter sets. HYPE (Hydrological Predictions for the Environment) is a semi-distributed, continuous simulation hydrological model which provides alternative routines for many simulated processes and utilizes both domain-wide parameters and parameters tied to physiography. The fourth-generation HYPE model for the pan-European domain, E-HYPE4, was calibrated for discharge and sediments within a framework which allowed for the evaluation of factors limiting model performance. Calibration was performed using a multi-phase approach in which an initial multi-objective calibration for discharge and sediments was followed by an exhaustive sediment calibration to evaluate combinations of erosion and sedimentation/resuspension routines. During each calibration phase, ensembles of parameter sets and model routines were simultaneously evaluated against discharge, evapotranspiration, and sediment observations. In total, 20,000 candidate parameter sets were evaluated during the discharge calibration, and a further 20,000 candidate model setups were assessed during the sediment calibration. Model performance was best with a highly regionalized model, and the largest drop in achievable model performance occurred when transitioning from an ensemble of candidates to a single model setup for the full model domain. However, much of the gains in performance with a highly regionalized model could be achieved with a much less regionalized model. Inclusion of sediments in the discharge calibration process reduced equifinality, and evaluation of the erosion routines indicated that a simple index-based routine performed equally well as a more complex, process-based routine . Finally, analysis of model performance by subbasin attributes revealed the dominant factors — such as landuse, glaciers, abstractions/regulations, groundwater, and lakes/wetlands — affecting model biases for different geographical regions.