Research papersA metric for attributing variability in modelled streamflows

- We present a metric for attributing variability in flow ensembles to its sources.
- The method, Quantile Flow Deviation (QFD), is applied to multi-model studies.
- The method gives guidance on the relative importance of different model choices.

Significant gaps in our present understanding of hydrological systems lead to enhanced uncertainty in key modelling decisions. This study proposes a method, namely “Quantile Flow Deviation (QFD)”, for the attribution of forecast variability to different sources across different streamflow regimes. By using a quantile based metric, we can assess the change in uncertainty across individual percentiles, thereby allowing uncertainty to be expressed as a function of magnitude and time. As a result, one can address selective sources of uncertainty depending on whether low or high flows (say) are of interest. By way of a case study, we demonstrate the usefulness of the approach for estimating the relative importance of model parameter identification, objective functions and model structures as sources of stream flow forecast uncertainty. We use FUSE (Framework for Understanding Structural Errors) to implement our methods, allowing selection of multiple different model structures. Cross-catchment comparison is done for two different catchments: Leaf River in Mississippi, USA and Bass River of Victoria, Australia. Two different approaches to parameter estimation are presented that demonstrate the statistic- one based on GLUE, the other one based on optimization. The results presented in this study suggest that the determination of the model structure with the design catchment should be given priority but that objective function selection with parameter identifiability can lead to significant variability in results. By examining the QFD across multiple flow quantiles, the ability of certain models and optimization routines to constrain variability for different flow conditions is demonstrated.