Learning from differing errors between machine learning and a conceptual hydrological model - the case of convective storms and flash floods

In recent years, major floods triggered by convective events during the summer months repeatedly occurred in central western Europe, where flood regimes had previously been characterised by slowly developing inundations in winter. This imposes a great challenge on conceptual hydrological models, as other runoff mechanisms seem to dominate during convective storms than the storage-driven runoff production, which dominates flood formation in the wet season. Hence, most hydrological models that are used for operational flood forecasting struggle when applied to convective events and show deficiencies in capturing peak flows and timing flood volumes. It is thereby unclear to which extent the uncertainty in precipitation input and discharge observations, the influence of the catchment state, the model structure itself, or a combination of several or all of these factors, compromise successful predictions. To shed light on this question, we will compare how different model structures perform during high flows across a range of catchment physiographic settings. We will analyse the performance of a conceptual hydrological water balance model and identify its deficiencies by comparing it to a data driven model. This comparison will reveal whether uncertainties in the input and output data or model structural deficiencies are the major source of error. This will allow us to identify systematic errors, compare them and improve the model structure.