基于自适应采样多保真度神经网络的土石坝邓肯—张E-B模型参数反演  

Parameter Inversion of Duncan Zhang E-B Model for Earth Rockfill Dams Based on Adaptive Sampling Multi-Fidelity Neural Network

作　　者：李江波 罗五洲 夏旦 吕小龙 凌骐 LI Jiang-bo;LUO Wu-zhou;XIA Dan;LV Xiao-long;LING Qi(Hunan Wuling Power Technology Co.,LTD.,Changsha 410000,China;Wuling Power Co.,LTD.,Changsha 410000,China;College of Mechanics and Engineering Science,Hohai University,Nanjing 211100,China;NARI Group Corporation,Nanjing 211100,China)

机构地区：[1]湖南五凌电力科技有限公司,湖南长沙410000 [2]五凌电力有限公司,湖南长沙410000 [3]河海大学力学与工程科学学院,江苏南京211100 [4]南瑞集团有限公司,江苏南京211100

出　　处：《水电能源科学》2025年第4期143-146,167,共5页Water Resources and Power

基　　金：国家重点研发计划(2018YFC0406703);国家自然科学基金重点项目(11932006)。

摘　　要：为缓解土石坝参数反演过程中需多次调用有限元计算造成计算量过大的问题,提出了基于多保真度神经网络(MFNN)结合自适应采样的参数反演方法。即基于MFNN对不同保真度数据信息的融合能力,充分利用低保真度数据计算量小、易获得的优势,提高代理建模效率;结合自适应采样方法能针对性地提高代理模型局部精度的能力,可进一步降低计算成本。对土石坝邓肯—张E-B模型参数反演的结果表明,所提方法仅需少量迭代步便达到收敛,且根据反演值计算出的沉降位移响应与实测值吻合良好,精度满足工程要求。To alleviate the problem of excessive computational complexity caused by multiple calls to finite element calculations during the parameter inversion process of earth rockfill dams,this paper proposes a parameter inversion method based on multi-fidelity neural network(MFNN)combined with adaptive sampling.By utilizing the fusion ability of MFNN for different fidelity data information,the advantages of low fidelity data with low computational complexity and easy access can be fully utilized to improve the efficiency of surrogate modeling.The adaptive sampling methods can improve the local accuracy of surrogate models and further reduce computational costs.The results of parameter inversion for the Duncan Zhang E-B model of earth rockfill dams indicate that the proposed method only requires a small number of iteration steps to achieve convergence,and the settlement displacement response calculated based on the inversion values is in good agreement with the measured values,with accuracy meeting engineering requirements.

关 键 词：土石坝 反演分析 邓肯-张E-B 代理模型 神经网络 

分 类 号：TV641[水利工程—水利水电工程] TV698.1