Incorporating accumulated temperature and lagged precipitation effects in LSTM modeling improves the streamflow simulation in glacier dominated basins

This study focuses on the Dongkemadi (DKMD) Glacier Basin, located in the source region of the Yangtze River Basin. We employed a Long Short-Term Memory (LSTM) neural network model optimized using Bayesian methods for streamflow simulation in a glacier-dominated basin and evaluated the model performance with different feature combinations. Additionally, we applied SHAP (SHapley Additive exPlanations) analysis to assess the contributions of the selected features. The Bayesian-optimized LSTM model could effectively simulate the complex runoff processes. Incorporating 22-day PDD and 23-day lagged precipitation improve prediction accuracy, increasing the testing set NSE from 0.826 to 0.879. SHAP analysis indicates that temperature and its cumulative effect (PDD) are identified as the primary meteorological drivers for glacier runoff, while lagged precipitation, despite improving overall accuracy, shows a lower contribution compared to temperature/PDD. • We developed a Bayesian LSTM model for streamflow simulation in a glacier basin. • Incorporating 22-day PDD and 23-day lagged precipitation improved the model accuracy. • SHAP analysis highlights temperature and PDD as key glacier melt drivers. • We propose a framework to identify optimal lag durations across glacier basins.