基于机器学习的冰冻圈典型流域水文过程模拟研究  

Modeling of hydrological processes in cryospheric watersheds based on machine learning

作　　者：宋轩宇 许民 康世昌 孙立平 SONG Xuanyu;XU Min;KANG Shichang;SUN Liping(State Key Laboratory of Cryospheric Science,Northwest Institute of Eco-Environment and Resources,Chinese Academy of Science,Lanzhou 730000,China;University of Chinese Academy of Sciences,Beijing 100049,China;School of Mathematics,Lanzhou Jiaotong University,Lanzhou 730070,China;School of Science,Northeastern University,Shenyang 110819,China)

机构地区：[1]中国科学院西北生态环境资源研究院冰冻圈科学国家重点实验室,甘肃兰州730000 [2]中国科学院大学,北京100049 [3]兰州交通大学数理学院,甘肃兰州730070 [4]东北大学理学院,辽宁沈阳110819

出　　处：《地学前缘》2023年第4期451-469,共19页Earth Science Frontiers

基　　金：国家自然科学基金项目(41971094,41871055);中国科学院青年创新促进会人才项目(2019414);中国科学院-澳大利亚联邦科学和工业研究组织(CAS-CSIRO)国际合作项目(131B62KYSB20190042)。

摘　　要：机器学习模型由于其优越的模拟预测性能而被广泛地应用于水文学研究,但其在高海拔地区的冰冻圈流域水文过程模拟研究方面尚有待深入。本研究基于Back Propagation神经网络(BP)、广义回归神经网络(GRNN)、径向基函数神经网络(RBF)、支持向量回归(SVR)、遗传优化BP神经网络(GA-BP)和双层长短期记忆神经网络模型(LSTM),对两个典型冰冻圈流域,即叶尔羌河流域和疏勒河流域的水文过程开展模拟研究,结合精度评价指标(NSE、RMSE和R)以及水文过程频率曲线对模型模拟效果进行综合分析。结果表明,双层LSTM模拟能力在叶尔羌河流域远优于其他模型,而疏勒河流域LSTM模拟效果与其他模型模拟结果相近,双层LSTM更适用于冰冻圈流域水文过程模拟。通过损失函数对模型参数化方案进行评价发现,LSTM模型在研究区模拟效果主要受优化器影响,叶尔羌流域学习衰减速率和初始学习率影响次之,而疏勒河流域初始学习率影响次之。对整个研究时段的径流突变检验分析结果表明,模型输入数据中降水和极端降水总量对研究区水文过程变化影响较大,气温次之。Machine learning models are widely used in hydrological research for their high predictive accuracy,however,their application in high-altitude cryospheric watersheds is seldom mentioned.In this study,machine learning models for two typical cryospheres,Yarkant and Shule river basins,were developed using BP neural network(BP),GRNN neural network(GRNN),RBF neural network(RBF),support vector regression(SVR),genetic optimization BP neural network(GA-BP)and double-layer long-term and short-term memory neural network(LSTM)algorithms,and model performance was evaluated using evaluation indexes NSE,RMSE and R and runoff frequency curves.The double-layer LSTM model performed much better than and similar to other models for the Yarkant and Shule River Basins,respectively;and overall the double-layer LSTM algorithm was more suitable for modeling hydrological processes in cryosphere basins.The loss function was used to evaluate the model parameterization scheme.It was found that the performance of the LSTM models was mainly affected by the optimizer,followed by the learning attenuation rate and initial learning rate for the Yarkant River Basin,and by the initial learning rate for the Shule River Basin.Model testing under abrupt changes in runoff suggested that climatic factors could have different hydrological impacts,meanwhile,precipitation and heavy precipitation R95p had the greatest impacts on the hydrological process in the study area,followed by temperature,during the entire study period.

关 键 词：气候变化 机器学习 冰冻圈流域 水文过程模拟 参数化 

分 类 号：P338[天文地球—水文科学]