基于机器学习的西南岩溶泉流量模拟研究  

Modelling karst spring flow in Southwest China based on machine learning

作　　者：马从文 张志才[1] 陈喜 程勤波[1] 彭韬[3,4] 张林[3,4] MA Congwen;ZHANG Zhicai;CHEN Xi;CHENG Qinbo;PENG Tao;ZHANG Lin(College of Hydrology and Water Resources,Hohai University,Nanjing,Jiangsu 210098,China;School of Earth System Science,Tianjin University,Tianjin 300072,China;Institute of Geochemistry,Chinese Academy of Sciences,Guiyang,Guizhou 550081,China;Puding Karst Ecosystem Observation and Research Station,Chinese Academy of Sciences,Puding,Guizhou 562100,China)

机构地区：[1]河海大学水文水资源学院,江苏南京210098 [2]天津大学地球系统科学学院,天津300072 [3]中国科学院地球化学研究所,贵州贵阳550081 [4]中国科学院普定喀斯特生态系统观测研究站,贵州普定562100

出　　处：《中国岩溶》2024年第1期48-56,共9页Carsologica Sinica

基　　金：自然科学基金重点项目(41571130071);面上项目(41971028,41571020)。

摘　　要：岩溶泉对西南岩溶区生态系统稳定和经济社会发展具有重要意义。受岩溶区独特水文地质结构与多重水流过程控制,岩溶泉流量具有复杂的动态变化特征,机器学习模型为其模拟和预测提供了有效手段。然而,岩溶泉域降雨−泉流量过程及其时空变异特征对机器学习模型结构与模拟精度的影响仍不明晰。本文选取西南典型岩溶泉,基于长短期记忆网络(LSTM)建立岩溶泉流量模拟模型,利用泉域实测逐小时降雨与泉流量序列进行模型训练与验证。在此基础上,分析了不同降雨−泉流量过程对岩溶泉流量模拟精度的影响,以及岩溶水文地质结构对降雨−泉流量响应滞时的控制作用。研究结果显示,山坡岩溶泉与流域出口岩溶泉训练期纳什效率系数(NSE)分别为0.942与0.951,验证期分别为0.831与0.834。对于山坡岩溶泉与流域出口岩溶泉,利用全年实测序列训练的模型预测雨季泉流量存在较大偏差,NSE分别为0.793与0.798,而利用雨季实测序列训练的模型预测雨季泉流量,精度显著提升,NSE分别为0.956与0.962,且此差异在暴雨频繁的5、6、7月尤为显著。受浅薄土壤与表层岩溶带分布影响,山坡岩溶泉LSTM模型时序步长显著小于流域出口岩溶泉。Karst springs are important for ecosystem and economic development in Southwest China.Controlled by the unique karst hydrogeological structure and multiple water flow processes,the karst spring flow has complex dynamic characteristics,thus posting a great challenge to simulate and predict the dynamic process of karst spring flow which can reflect the characteristics of rainfall–spring flow in the karst basin.As a data-driven model,the machine learning model omits the necessity of considering complex physical processes,showing its significant advantages in the simulation and prediction of nonlinear system variables.Therefore,it provides an effective approach for simulation and prediction of karst spring discharge.However,the influence of the flow processes and hydrogeological conditions on the structure and simulation accuracy of machine learning model is still unclear.Among the machine learning algorithms,LSTM,as the most popular algorithm in recent years,is widely used in the simulation and prediction of various long-time series data.LSTM adds a cell state similar to"conveyor belt"in the hidden layer,and the cell state is adjusted by forgetting gate,input gate and output gate.This structure can effectively solve the long-term transportation and memory problems of time series data,and is more suitable for runoff simulation and prediction than the traditional neural network algorithm.In this study,for the processes of rainfall–spring flow in karst areas,two typical karst springs(hillside karst spring and outlet karst spring)representing different geomorphic units in Southwest China are selected.Through hyper parameter optimization,a double hidden layer and double input LSTM model are adopted to build a machine learning model of typical karst spring flow.The measured meteorological and hydrological data from 0:00 on January 1,2017 to 24:00 on December 31,2019 are used.2017–2018 is the training period and 2019 is the validation period.The model has been trained and verified.Based on the simulation results,the

关 键 词：机器学习 LSTM 岩溶泉流量 响应滞时 岩溶降雨−泉流量过程 

分 类 号：P641.134[天文地球—地质矿产勘探]