Improving runoff simulation in cold alpine regions based on VIC-glacier by combining LSTM error correction technology

Xinjiang’s cold alpine region, known as the “solid desert reservoir”, harbors over 40 % of China’s modern glaciers. Its abundant glacial resources significantly impact the ecological and economic development of the entire Northwestern region. Quantifying the proportion of glacier meltwater and forecasting future runoff trends have become critical research focuses in cold alpine areas. This study developed a coupled Variable Infiltration Capacity-Glacier (VIC-glacier) model for the upper Hotan River Basin, a typical cold alpine region, and optimized it using the Shuffled Complex Evolution (SCE-UA) automatic calibration method. The simulation results were further improved by integrating Long Short-Term Memory (LSTM) and Autoregression (AR) error correction techniques, with their performances compared and validated. The results show that the VIC-glacier model with LSTM error correction demonstrates significantly enhanced accuracy across different forecast periods. The Nash-Sutcliffe Efficiency (NSE) of short-term forecasts reaches 0.9 in both training and testing periods, with the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) reduced compared to the uncorrected model. Compared with AR correction, the LSTM model consistently outperforms in multiple foresight periods. This model mitigates uncertainties in water resources caused by glacier changes and enables more accurate future runoff predictions, providing a scientific basis for integrated water resource management in Xinjiang’s cold alpine regions.