Assessing the Hydrologic Suitability of MC-LSTM for Reliable Streamflow Prediction in Human-Influenced Watersheds

Anthropogenic activities such as rapid urbanization, highly regulated dams and reservoirs, and widespread land-use changes significantly modify natural streamflow dynamics, making streamflow prediction increasingly challenging for traditional hydrological models. Deep learning approaches, particularly the long short-term memory (LSTM) network, have gained popularity due to their ability to capture long-term dependencies in hydrological time series. However, the purely data-driven nature of LSTM limits its reliability in human-influenced watersheds. The Mass-Conserving LSTM (MC-LSTM) addresses this limitation by incorporating mass-balance constraints directly into its architecture, enabling physically consistent predictions. Despite this advancement, a systematic comparison between LSTM and MC-LSTM in human-influenced hydrological systems remains limited. In this study, we evaluate the predictive advantage and hydrologic suitability of MC-LSTM across 51 human-influenced watersheds in India. The watersheds are categorized into low- and high-human-influenced categories using a composite disturbance index (CDI), derived from the number of dams, reservoir storage, cropland fraction, built-up fraction, and population density.  This setup allows us to address a key question: Does incorporating mass balance constraints into LSTM improve streamflow reliability in highly regulated watersheds? The results show that MC-LSTM substantially outperforms traditional LSTM in highly human-influenced watersheds, yielding a significantly higher median NSE (MC-LSTM: 0.69; LSTM: 0.62). MC-LSTM also demonstrates several additional benefits, including improved high-flow prediction, reduced sensitivity to training data size, and slightly enhanced performance in semi-arid watersheds. In contrast, traditional LSTM tends to underestimate high-flow, depends on larger training datasets, and performs poorly in semi-arid and highly regulated basins. These findings underscore the importance of incorporating mass balance into DL-based hydrological models to enhance reliability in real-world applications.