基于深度学习的双循环井水动力调控工艺参数多目标优化设计  

Multi-objective optimization design of process parameters for hydrodynamic regulation of dual circulation wells based on deep learning

作　　者：马彦玲 方樟 周睿 刘治国 蒲生彦[2] 丁小凡 MA Yanling;FANG Zhang;ZHOU Rui;LIU Zhiguo;PU Shengyan;DING Xiaofan(College of New Energy and Environment,Jilin University,Changchun 130021,China;College of Ecology and Environment,Chengdu University of Technology,Chengdu 610059,China;Shenyang Academy of Environmental Sciences,Shenyang 110167,China)

机构地区：[1]吉林大学新能源与环境学院,长春130021 [2]成都理工大学生态环境学院,成都610059 [3]沈阳环境科学研究院,沈阳110000

出　　处：《环境工程学报》2024年第4期968-977,共10页Chinese Journal of Environmental Engineering

基　　金：国家重点研发计划资助项目(2020YFC1808300)。

摘　　要：循环井群优化设计是循环井技术污染场地修复应用时需考虑重要问题。本研究提出了一种基于深度学习方法的循环井群模拟-优化耦合技术。以双循环井作为井群运行的典型单元,以影响双循环井运行效果的关键参数作为输入变量,利用FloPy调用MODFLOW、MODPATH和MT3DMS模块构建双循环井驱动下的地下水流、粒子追踪及溶质运移数值模型,将表征运行效果指标循环效率(Pr)和污染物去除效率(η)作为模型输出参数,构建数据集;利用卷积神经网络模型建立替代模型;利用NSGA-Ⅱ算法建立以双循环井运行效果的多目标优化模型。在此基础上对陕西省西安市某示范场地双循环井进行优化设计,在兼顾Pr和η的前提下,以Pr为主要目标时的最优设计方案为相同抽注模式且均为上抽下注模式下,抽注水流量为185 m^(3)·d^(-1),井间距为20m,该方案下固定时间内Pr达到99%,η达到27%;以η为主要目标时的最优设计方案为双循环井抽注方式相反,抽注水流量为163m^(3)·d^(-1),井间距为13m。该方案下固定时间内Pr达到95%,η达到34%,并给出了2种方案的井结构参数。研究结果表明基于卷积神经网络方法建立的替代模型精确度高,预测双循环井Pr时的R2达到0.972,MAPE达到0.008,预测η时的R2达到0.986,MAPE达到0.009,且相较于传统数值模拟方法大大节省了数值计算时间。模拟-优化耦合技术可以应用于解决双循环井前期优设计问题,通过输入实际场地的水文地质参数,即获得双循环井最优的运行效果以及对应的优化设计方案,可为循环井技术的设计制造提供参考。The optimal design of groundwater circulation wells(GCWs)is crucial for the implementation of this technique to site remediation.This work proposed a deep-learning-based optimal technology by coupling circulation wells simulation with optimization.The key parameters affecting the operation of a dual-circulation wells,a typical unit of well cluster,were treated as input variables;The modules MODFLOW,MODPATH,and MT3DMS were invoked by FloPy to construct the groundwater flowing model,particle tracking model,and solute transporting model.The simulated performance indicators(circulation efficiency(Pr),and pollutant removal efficiency(η))are treated as output variables to build a training dataset.Convolutional neural network model was employed to establish an alternative model.The non-dominated sorting genetic algorithm(NSGA-II)was applied in building a multi-objective optimization model with the operation effect characterization indicators of the dual-circulation wells.The optimal design of dual-circulation wells at the demonstration site in Xi'an City,Shaanxi Province was carried out.Considering both Pr andη,the ideal design for a dual-circulation well,with Pr as the primary target,maintains the same pumping mode with the distraction screen in the upward and injection screen in the downward,with a pumping rate of 185 m^(3)·d^(-1) and a well spacing of 20 m.This scheme results in a 99%Pr rate and a 27%ηover a set duration;Conversely,whenηis the primary target,the ideal design scheme involves the reverse pumping mode,a 163 m^(3)·d^(-1) pumping rate,and a 13 m well spacing.This scheme attains a 95%Pr and a 34%ηover a set period.The parameters for the well structure of the both schemes are also provided in the research.The findings showed that the alternative model based on the convolutional neural network method has high computational precision,evidenced by an R2 of 0.972 and a MAPE of 0.008 in forecasting Pr in the dual-circulation wells,alongside an R2 of 0.986 and a MAPE of 0.009 in predictingη.This approach

关 键 词：地下水 双循环井 深度学习 卷积神经网络 多目标优化设计 NSGA-Ⅱ遗传算法 

分 类 号：X523[环境科学与工程—环境工程]