Benchmarking model performance in complex human-water systems.

Human activities must now be considered as an integral part of the water cycle. Consequently, the integration of human-water interactions into hydrological modelling is essential for the large-scale simulation of flow. However, whilst the last decade has seen substantial advancements in the guidance available for modelers on how best to benchmark and evaluate flow simulations in natural catchments, there is little discussion surrounding how these practices may differ in more complex, human-impacted catchments. Here we discuss some of the key issues in benchmarking model performance in human-impacted catchments and demonstrate these using a large-sample of reservoir-impacted catchments across Great Britain. We find that evaluation metrics designed for natural systems do not always translate to those impacted by human activity, where reservoir-impacted flow timeseries can have a substantially different distribution. In light of the new parameters and model assumptions associated with representing human activities within the natural water cycle, we suggest that the integration of uncertainty quantification and sensitivity analysis (UQ and SA) for robust model evaluation is particularly important. We discuss the need for clear accessible workflows for the application of UQ and SA in the evaluation of complex and large-scale water resource system modelling.