Study on runoff simulation on northern slope of Tianshan Mountains considering vertical zonality and glacier ablation

Abstract

[Objective] Runoff in glacial basins on the cold and arid northern slope of the Tianshan Mountains is jointly influenced by vertical zonality and the nonlinearity of glacier ablation. However, the significant vertical gradient and high glacier runoff contribution rate lead to prominent challenges in this region, namely poor snowmelt runoff simulation and insufficient characterization of the physical mechanism of baseflow. Taking the Manas River Basin as the study area, the performance adaptation patterns of physically-based and data-driven models under vertical differentiation are investigated. [Methods] Integrating the CMADS dataset, a parallel runoff simulation framework combining SWAT, LSTM, and CNN-LSTM was established. The Nash-Sutcliffe Efficiency(NSE), Root Mean Square Error(RMSE), and Percent Bias(PBIAS) were used to quantitatively evaluate the performance of monthly runoff simulation from 1979 to 2018, and the causes of model errors during the snowmelt period(April-July) and low-flow period(November-March) were analyzed. [Results] The result showed that: CNNLSTM exhibited superior overall simulation performance, with an overall NSE of 0. 829—representing a 3. 6% and 2. 2%improvement over LSTM(0. 800) and SWAT(0. 811), respectively. During the snowmelt period(April—July), CNN-LSTM still performed best(NSE= 0. 787), which was 6. 8% and 3. 0% higher than that of SWAT(0. 737) and LSTM(0. 764),respectively. In the low-flow period( November-March), CNN-LSTM and LSTM showed similar performance( both NSE=0. 859). Given the small runoff base in the low-flow period, although SWAT had a lower NSE(0. 782) than the two models, its explicit physical parameterization of groundwater processes still endowed the result with higher process robustness and interpretability, thus possessing unique analytical value. [Conclusion] Under the vertical zonality gradient, differences in model performance are jointly driven by hydrological process mechanisms and model structural advantages. CNN-LSTM learns the spatiotemporal patterns of meteorological factors via a data-driven approach and exhibits excellent adaptability to nonlinear processes. SWAT, with explicit parameterization based on physical mechanisms, exhibits unique process consistency and interpretability. The advantage boundaries and applicable scenarios of different models in runoff simulation in cold and arid regions are identified, providing a key scientific basis for the development of integrated physically-based and data-driven runoff simulation models.

Keywords

runoff simulation / CNN-LSTM model / SWAT model / CMADS / Manas River Basin / influencing factors