基于神经网络的二阶波-流和结构物相互作用  

Study on second-order wave-current interactions with a structure based on the neutral network

作　　者：王赤忠[1,2] 郑宇谦 葛晗 朱嵘华[1,2] WANG Chizhong;ZHENG Yuqian;GE Han;ZHU Ronghua(Ocean College,Zhejiang University,Zhoushan 316021,Zhejiang China;Yangjiang Offshore Wind Power Laboratory,Yangjiang 529500,Guangdong China)

机构地区：[1]浙江大学海洋学院,浙江舟山316021 [2]阳江海上风电实验室,广东阳江529500

出　　处：《华中科技大学学报（自然科学版）》2024年第4期16-21,共6页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(52171325)。

摘　　要：基于时域二阶势流理论,对位于均匀流中的水平半圆柱体在水面上的二阶波绕射问题进行了计算,得到了自由表面波时间历程、圆柱所受到的水动力时间历程及相应的一、二阶波和水动力幅值,并以此作为神经网络的训练样本,采用LM-BP神经网络来预测任意参数(波浪频率或波数、流速或傅汝德数)组合下波浪和水动力幅值或峰值及其时间历程,可以快速地获得精度较高的计算结果.研究结果表明:对于样本数较小的一阶、二阶波浪和水动力幅值预测,可以采用单个隐含层及较少的神经元数或节点数即可获得较高精度的预测结果;而对于样本数很大的波浪和水动力时间历程预测,需要至少2个隐含层及较多的节点数才能获得较满意的结果.Wave diffraction by a horizontal semi-circle in a uniform current was evaluated based on the potential flow theory through the time-domain second-order approach.The time histories of first-and second-order wave and hydrodynamic force and their amplitudes or peaks of their superpositions were obtained and taken as the training set.The back propagation neutral network with Levenberg-Marquardt algorithm(LM-BP)was employed to make fast and accurate prediction of the wave and force including their histories and amplitudes or peaks at any wave frequency and current speed.Research results show that higher accurate results on wave and force amplitudes or peaks with small sample can be obtained using a single hidden layer with fewer neurons,while two hidden layers at least with more neurons are required for predicting the histories of first-and second-order waves and forces with very large sample.

关 键 词：时域二阶理论 波-流-体相互作用 BP(逆向传播)神经网络 LEVENBERG-MARQUARDT算法 机器学习 

分 类 号：O35[理学—流体力学]