多年冻土区冻结层上水动态及模拟——以青藏高原风火山流域为例  

Dynamics and Simulation of Suprapermafrost Water in Permafrost Regions:A Case Study of the Fenghuoshan Watershed on the Qinghai-Tibetan Plateau

作　　者：邹禧 王根绪 吴碧琼 宋春林 郭林茂 李金龙 ZOU Xi;WANG Genxu;WU Biqiong;SONG Chunlin;GUO Linmao;LI Jinlong(College of Water Resource&Hydropower,Sichuan University,Chengdu 610065;State Key Laboratory of Hydraulics and Mountain River Engineering,Sichuan University,Chengdu 610065;China Yangtze Power Co.,Ltd.,Yichang 443002,Hubei,China)

机构地区：[1]四川大学水利水电学院,成都610065 [2]四川大学山区河流保护与治理全国重点实验室,成都610065 [3]中国长江电力股份有限公司,湖北宜昌443002

出　　处：《山地学报》2025年第1期28-45,共18页Mountain Research

基　　金：国家自然科学基金联合基金(U2240226);国家自然科学基金(423B2102)。

摘　　要：冻结层上水是寒区冻土水文循环的关键层,揭示其动态演变规律,对认知冻土区地下水运移机制及精准预测具有重要科学意义。然而,由于多年冻土区原位监测数据的匮乏,以及非线性适应型水文过程模型构建的缺失,冻结层上水动态时空预测精度难以满足科学研究和工程实践需求。本研究以青藏高原风火山小流域(海拔4063~5398 m)为典型研究区,基于2021-2023年原位观测气象数据(精度±0.1℃/±0.1 mm)、逐日土壤水热(精度±1℃/±0.03 m^(3)·m^(-3))及冻结层上水位(精度±0.14 cm)原位监测数据,揭示坡面尺度冻结层上水动态的水热时空协同机制;集成气温、降水、土壤温湿度和初始水位等多要素,构建及评估基于长短期记忆神经网络(LSTM)的冻土水文预测模型的适应性。研究发现:(1)冻结层上水动态具有显著季节分异特征,其水位波动(年变幅0~1.53 m)与活动层土壤温湿度呈现一致性,基于Boltzmann函数的平均拟合优度为0.90。(2)所构建的基于LSTM方法的冻结层上水位预测模型(学习率0.002)在坡面多梯度验证中表现出卓越性能,平均纳什效率系数(NSE)为0.83,证实模型具备复杂冻土水文过程的解析能力。(3)模型敏感性分析表明,在仅采用气温、降雨和初始水位作为输入的基准情景下,模型仍能维持0.72的NSE值,土壤温湿度参数的引入使模拟精度平均提升8.9%,这为数据稀缺区提供了可靠的简化预测方案。本研究提出的多要素耦合建模方法,克服了传统水文模型的物理机制限制,可以为多年冻土区水资源管理与生态水文系统稳定性提供可靠的技术支撑。Suprapermafrost water,a crucial layer in the hydrological cycle of permafrost regions,plays a pivotal role in deciphering groundwater dynamics and enabling reliable predictions in permafrost environments.However,limited by the lack of in-situ monitoring data of suprapermafrost water in the Qinghai-Tibetan Plateau and the absence of robust mathematical frameworks tailored to permafrost hydrogeology,the accuracy of spatio-temporal predictions for suprapermafrost water dynamics is insufficient to meet the needs of scientific research and engineering practice.In this study,it selected the Fenghuoshan watershed(altitude:4063-5398 m),a representative permafrost region on the Qinghai-Tibetan Plateau,as research target.It utilized in-situ meteorological data(accuracy:±0.1℃/±0.1 mm),daily soil water-heat(accuracy:±1℃/±0.03 m^(3)·m^(-3)),and suprapermafrost water table(resolution:±0.14 cm)monitoring data in the period of 2021-2023 to intepret the spatio-temporal interactions and spatial heterogeneity of the three variables at the hillslope scale.Multiple factors such as air temperature,precipitation,soil temperature-humidity,and initial water table were integrated into a permafrost hydrological prediction model based on the Long Short-Term Memory(LSTM)neural network,with its adaptability evaluated.0x09(1)Supra-permafrost water tables exhibited pronounced seasonal variability,synchronized with intra-annual fluctuations of soil temperature/moisture in active-layer.A Boltzmann function-based model effectively captured phase transition dynamics during freeze-thaw cycles,achieving a mean goodness-of-fit of 0.90.(2)A novel Long Short-Term Memory(LSTM)neural network model,specifically optimized for permafrost hydrology,demonstrated exceptional predictive capability across diverse slope positions.The model attained an average Nash-Sutcliffe efficiency coefficient(NSE)of 0.82.(3)Sensitivity analysis revealed that the model maintained an NSE of 0.72 using only temperature,precipitation,and initial water tables as inputs.In

关 键 词：多年冻土 冻结层上水 土壤温湿度 地下水位模拟 LSTM 

分 类 号：P642.1[天文地球—工程地质学]