PDF下载

[1]刘兴坡,王志强,李璟.基于BP神经网络的SWMM参数全局灵敏度分析方法[J].中国给水排水,2021,37(9):122-129.
　LIU Xing-po,WANG Zhi-qiang,LI Jing.Global Sensitivity Analysis Method for Parameters of Storm Water Management Model Based on BP Neural Network[J].China Water & Wastewater,2021,37(9):122-129.

点击复制

基于BP神经网络的SWMM参数全局灵敏度分析方法

中国给水排水[ISSN:1000-4062/CN:12-1073/TU] 卷: 第37卷期数: 2021年第9期页码: 122-129 栏目: 出版日期: 2021-05-01

Title:: Global Sensitivity Analysis Method for Parameters of Storm Water Management Model Based on BP Neural Network

作者:: 刘兴坡^1,2，王志强^1,2，李璟^1,2; (1.上海海事大学海洋科学与工程学院，上海201306；2.上海海事大学海洋环境与生态模拟研究中心，上海201306)

Author(s):: LIU Xing-po^1,2,WANG Zhi-qiang^1,2,LI Jing^1,2; (1. College of Ocean Science and Engineering, Shanghai Maritime University, Shanghai 201306, China; 2. Center for Marine Environmental and Ecological Modelling, Shanghai Maritime University, Shanghai 201306, China)

关键词:: SWMM模型参数; 全局灵敏度分析; BP神经网络; 修正Morris法; LH-OAT法; 线性回归法; Nash-Sutcliffe效率系数; 集对分析联系度

Keywords:: SWMM parameter; global sensitivity analysis; BP neural network; modified Morris method; LH-OAT (Latin hypercube one factor-at-a-time); linear regression method; Nash-Sutcliffe efficiency coefficient; set pair relational degree

摘要:: 为了准确高效地分析SWMM模型参数的灵敏度，识别模型关键校准参数，提出了基于BP神经网络的参数全局灵敏度分析方法。以汇水区面积、宽度、不透水率等12个参数作为模型输入，以Nash-Sutcliffe效率系数（NSE）代表模型输出，设计了3套验证方案及其比较参考准则，对BP神经网络法与修正Morris法、线性回归法、LH-OAT法、多项式回归法等灵敏度分析方法进行了比较分析。结果表明，应用BP神经网络的连接权值作为全局灵敏度分析依据时，BP神经网络法的灵敏度分析精度优于线性回归法、LH-OAT法和多项式回归法。因此，BP神经网络法是SWMM模型参数全局灵敏度分析的一种可行方法。SWMM模型参数灵敏度分析精度与分析模型的非线性拟合程度存在密切关系，优良的非线性拟合能力是BP神经网络方法优于其他灵敏度分析方法的重要原因。通过比较不同灵敏度分析方法的效果，提出了以Morris扰动灵敏度分析结果作为真值基准、以SWMM模型几何参数作为灵敏度指示参数、以线性回归法和拟合优度作为比较基准等方法，可为灵敏度分析方法效果评价提供借鉴。基于BP神经网络的SWMM模型参数灵敏度分析方法为解决复杂模型参数“参数灵敏度分析-参数寻优”多重循环校准开拓了新路径。

Abstract:: In order to accurately and efficiently analyze the sensitivity of SWMM parameters and identify the key calibration parameters of the model, a global sensitivity analysis method based on BP neural network was proposed. Twelve parameters (such as catchment area, width, impermeable area ratio, etc.) were inputted into the model, and Nash-Sutcliffe efficiency coefficient (NSE) was used to represent the model output. Three sets of verification schemes and their comparison reference criteria were designed, and sensitivity analysis methods (such as BP neural network, modified Morris method, linear regression method, LH-OAT method, polynomial regression method, etc.) were compared and analyzed. When connection weight of BP neural network was used as the basis of global sensitivity analysis, the sensitivity analysis accuracy of BP neural network method was better than that of linear regression method, LH-OAT method and polynomial regression method. Therefore, BP neural network method was a feasible method for global sensitivity analysis of SWMM parameters. The parameter sensitivity analysis accuracy of SWMM was closely related to the non-linear fitting degree of the analysis model. Excellent non-linear fitting ability was an important reason why BP neural network was superior to other sensitivity analysis methods. By comparing the effects of different sensitivity analysis methods, it was proposed that Morris perturbation sensitivity analysis results were used as the true value benchmark, SWMM geometric parameters were used as sensitivity indicator parameters, and linear regression method and goodness of fit were used as comparison benchmark, which could provide reference for the effect evaluation of sensitivity analysis methods. The parameter sensitivity analysis method of SWMM based on BP neural network develops a new way to solve the multi-cycle calibration of parameter sensitivity analysis and optimization of complex models.

更新日期/Last Update: 2021-05-01