Towards Reliable Prediction of Streamflow in Semi-Arid Watersheds using a Mass-Conserving LSTM

Reliable streamflow prediction is essential for taking preventive action to manage floods and droughts. Long Short-Term Memory (LSTM) networks have gained significant attention in hydrology due to their strong ability to learn nonlinear rainfall–runoff relationships and capture temporal dependencies within hydrological time series. However, it often struggles in semi-arid watersheds due to weak rainfall–runoff correlation and higher rainfall variability, and evapotranspiration dominating the water balance. In addition, the model is frequently criticized for its black-box nature, highlighting the need for approaches that combine data-driven learning with hydrological realism. To address these limitations, the recently developed Mass-Conserving LSTM (MC-LSTM) incorporates mass-balance constraints directly into its architecture, ensuring physically consistent predictions. Despite this advancement, its suitability for semi-arid environments remains underexplored, raising the scientific question: Does incorporating mass balance improve model performance in semi-arid watersheds? To investigate this, a comparative analysis of LSTM and MC-LSTM was conducted across ten semi-arid watersheds in India. The results demonstrate that MC-LSTM consistently outperformed the traditional LSTM, substantially reducing overall bias and high-flow bias. Moreover, LSTM performance deteriorated with increasing potential evapotranspiration (PET), indicating its limited ability to capture PET-driven processes in semi-arid catchments. In contrast, the MC-LSTM, by enforcing mass balance, was better able to represent these dynamics to some extent. Overall, these findings highlight the importance of integrating mass-balance principles into deep learning models to enhance the reliability of streamflow prediction in semi-arid watersheds.