An evaluation of the impact of model structure on hydrological modelling uncertainty for streamflow simulation

Abstract

Operational flood management and warning requires the delivery of timely and accurate forecasts. The use of distributed and physically based forecasting models can provide improved streamflow forecasts. However, for operational modelling there is a trade-off between the complexity of the model descriptions necessary to represent the catchment processes, the accuracy and representativeness of the input data available for forecasting and the accuracy required to achieve reliable, operational flood management and warning. Four sources of uncertainty occur in deterministic flow modelling; random or systematic errors in the model inputs or boundary condition data, random or systematic errors in the recorded output data, uncertainty due to sub-optimal parameter values and errors due to incomplete or biased model structure. While many studies have addressed the issues of sub-optimal parameter estimation, parameter uncertainty and model calibration very few have examined the impact of model structure error and complexity on model performance and modelling uncertainty. In this study a general hydrological framework is described that allows the selection of different model structures within the same modelling tool. Using this tool a systematic investigation is carried out to determine the performance of different model structures for the DMIP study Blue River catchment using a split sample evaluation procedure. This investigation addresses two questions. First, different model structures are expected to perform differently, but is there a trade-off between model complexity and predictive ability? Secondly, how does the magnitude of model structure uncertainty compare to the other sources of uncertainty? The relative performance of different acceptable model structures is evaluated as a representation of structural uncertainty and compared to estimates of the uncertainty arising from measurement uncertainty, parametric uncertainty and the rainfall input. The results show first that model performance is strongly dependent on model structure. Distributed routing and to a lesser extent distributed rainfall were found to be the dominant processes controlling simulation accuracy in the Blue River basin. Secondly that the sensitivity to variations in acceptable model structure are of the same magnitude as uncertainties arising from the other evaluated sources. This suggests that for practical hydrological predictions there are important benefits in exploring different model structures as part of the overall modelling approach. Furthermore the model structural uncertainty should be considered in assessing model uncertainties. Finally our results show that combinations of several model structures can be a means of improving hydrological simulations.