Climate change impact assessment on a German lowland river using long short-term memory and conceptual hydrological models

Lowland Ems catchment in North-west Germany Alterations of streamflow caused by climate change can impact the riverine ecosystems, socio-economic development and the effectiveness of flood protection measures. Therefore, hydrological climate impact studies are crucial for foresighted water management. In addition to conceptual and physical hydrological models, machine learning based models are frequently used in streamflow modelling. However, the applicability of machine learning models for climate impact studies remains relatively unexplored. In this study, we use a standard ensemble of climate change projections of the German Weather Service for the Ems catchment. We compare the climate change signals of two conceptual hydrological models to those of two different long short-term memory model approaches. Based on long-term simulations, climate change impact on different streamflow indices is assessed. The results indicate heterogenous alterations of streamflow depending on the applied model and its model-specific input data. Further analysis reveals that the application of the Haude formula for calculation of potential evaporation is questionable for climate change impact assessment despite its successful application in the study region in recent climate. For a successful application of LSTM in climate change impact studies, physical restrictions need to be defined to simulate a plausible catchment behavior also for climate conditions beyond the range of training data. • Comparison of conceptual and deep learning hydrological models. • Climate change impact assessment of streamflow is model dependent. • Simulated actual evaporation varies between the modelling approaches. • Physical restrictions are required in deep learning models for climate change studies.