The application of machine learning and deep learning methods to hydrological modelling has advanced significantly in recent years, offering alternatives to traditional conceptual and physically based approaches. Within the numerous algorithms, long short-term memory (LSTM) networks have proven themselves particularly useful for the task of streamflow modelling. This thesis provides a collection of publications that investigate the capabilities, limitations and interpretability of LSTM for the purpose of streamflow modelling and climate change impact assessment within the lowland Ems catchment in Northwest Germany. Within a comparative performance evaluation, LSTM and its predecessor, the recurrent neural network, demonstrate superior accuracy compared to the conceptual HBV model across various statistical performance metrics. However, a decline in performance was observed during low-flow conditions in certain sub-catchments. The evaluation of the flow duration curve revealed that the ML models more effectively capture the water balance, while HBV better represents streamflow dynamics. To enhance the interpretability of LSTM, six explainable artificial intelligence techniques were applied. These methods consistently identified seasonal patterns in the temporal relevance of hydroclimatic input data. In combination with an observed correlation between the internal LSTM states and catchment-scale soil moisture dynamics, the findings suggest that LSTM models are capable of implicitly learning the relevant hydrological processes. Following, the capabilities of LSTM to model climate change impact scenarios, particularly when they extend beyond historically observed climate conditions, are addressed. An ensemble of climate change projections is provided as hydroclimatic input to evaluate the performance of LSTMs and conceptual models. While all models reveal heterogeneous alterations in streamflow under future climate conditions, significant differences emerge based on the model type. Results provide evidence that LSTMs, in combination with the temperature-based Haude formula for estimating potential evaporation, work inadequately under altered climatic regimes, raising concerns about their applicability in long-term projections. The study also indicates the potential need to incorporate physical constraints into LSTM architectures to ensure model robustness and hydrological plausibility beyond the historical training range. Collectively, this thesis contributes important insights into the applicability and interpretability of LSTM models in streamflow modelling. Despite the presence of a physically realistic representation of soil moisture dynamics of the Ems catchment, no robust change signals for streamflow under climate change can be derived. Those results underscore the potential of LSTM model approaches for accurate streamflow simulation, however, they require us to always critically question LSTM results, particularly when they are applied outside the training range.