Intercomparing LSTM and RNN to a Conceptual Hydrological Model for a Low-Land River with a Focus on the Flow Duration Curve

Machine learning (ML) algorithms slowly establish acceptance for the purpose of streamflow modelling within the hydrological community. Yet, generally valid statements about the modelling behavior of the ML models remain vague due to the uniqueness of catchment areas. We compared two ML models, RNN and LSTM, to the conceptual hydrological model Hydrologiska Byråns Vattenbalansavdelning (HBV) within the low-land Ems catchment in Germany. Furthermore, we implemented a simple routing routine in the ML models and used simulated upstream streamflow as forcing data to test whether the individual model errors accumulate. The ML models have a superior model performance compared to the HBV model for a wide range of statistical performance indices. Yet, the ML models show a performance decline for low-flows in two of the sub-catchments. Signature indices sampling the flow duration curve reveal that the ML models in our study provide a good representation of the water balance, whereas the HBV model instead has its strength in the reproduction of streamflow dynamics. Regarding the applied routing routine in the ML models, there are no strong indications of an increasing error rising upstream to downstream throughout the sub-catchments.