The accurate simulation and prediction of runoff in alpine glaciated watersheds is of increasing importance for the comprehensive management and utilization of water resources. In this study, long short-term memory (LSTM), a state-of-the-art artificial neural network algorithm, is applied to simulate the daily discharge of two data-sparse glaciated watersheds in the Tianshan Mountains in Central Asia. Two other classic machine learning methods, namely extreme gradient boosting (XGBoost) and support vector regression (SVR), along with a distributed hydrological model (Soil and Water Assessment Tool (SWAT) and an extended SWAT model (SWAT_Glacier) are also employed for comparison. This paper aims to provide an efficient and reliable method for simulating discharge in glaciated alpine regions that have insufficient observed meteorological data. The two typical basins in this study are the main tributaries (the Kumaric and Toxkan rivers) of the Aksu River in the south Tianshan Mountains, which are dominated by snow and glacier meltwater and precipitation. Our comparative analysis indicates that simulations from the LSTM shows the best agreement with the observations. The performance metrics Nash-Sutcliffe efficiency coefficient (NS) and correlation coefficient (R²) of LSTM are higher than 0.90 in both the training and testing periods in the Kumaric River Basin, and NS and R² are also higher than 0.70 in the Toxkan River Basin. Compared to classic machine learning algorithms, LSTM shows significant advantages over most evaluating indices. XGBoost also has high NS value in the training period, but is prone to overfitting the discharge. Compared with the widely used hydrological models, LSTM has advantages in predicting accuracy, despite having fewer data inputs. Moreover, LSTM only requires meteorological data rather than physical characteristics of underlying data. As an extension of SWAT, the SWAT_Glacier model shows good adaptability in discharge simulation, outperforming the original SWAT model, but at the cost of increasing the complexity of the model. Compared with the oftentimes complex semi-distributed physical hydrological models, the LSTM method not only eliminates the tedious calibration process of hydrological parameters, but also significantly reduces the calculation time and costs. Overall, LSTM shows immense promise in dealing with scarce meteorological data in glaciated catchments.

高山冰河流域的径流的精确模拟和预测对于水资源的综合管理和利用具有越来越重要的意义。在这项研究中，采用了最先进的人工神经网络算法-长短期记忆（LSTM），来模拟中亚天山山区两个数据稀疏的冰川流域的日流量。比较中还使用了其他两种经典的机器学习方法，即极端梯度增强（XGBoost）和支持向量回归（SVR），以及分布式水文模型（土壤和水评估工具（SWAT）和扩展SWAT模型（SWAT_Glacier））本文旨在提供一种有效且可靠的方法，用于对观测气象数据不足的冰川高寒地区的流量进行模拟。本研究中的两个典型盆地是南天山山脉阿克苏河的主要支流（库马里克河和托克斯坎河），主要支流是雪，冰川融水和降水。我们的比较分析表明，LSTM的模拟与观察值显示出最佳的一致性。性能指标Nash-Sutcliffe效率系数（NS）和相关系数（在库玛力河流域的训练和测试期间，LSTM的R ²）均高于0.90，而NS和R ²在Toxkan流域也高于0.70。与经典的机器学习算法相比，LSTM在大多数评估指标上均显示出显着优势。XGBoost在训练期间也具有较高的NS值，但易于过放电。与广泛使用的水文模型相比，尽管数据输入较少，但LSTM在预测精度方面具有优势。此外，LSTM仅需要气象数据，而不需要基础数据的物理特征。作为SWAT的扩展，SWAT_Glacier模型在放电模拟中显示出良好的适应性，优于原始的SWAT模型，但代价是增加了模型的复杂性。与通常复杂的半分布式物理水文模型相比，LSTM方法不仅省去了繁琐的水文参数校准过程，而且大大减少了计算时间和成本。总体而言，LSTM在处理冰川积水的稀缺气象数据方面显示出巨大的希望。