离心泵水力优化问题代理模型性能的参数化  

Parametric on performance of surrogate models in hydraulic optimization of centrifugal pumps

作　　者：甘星城 裴吉[1] 袁寿其[1] 王文杰[1] 颜爱忠 赵芸 GAN Xingcheng;PEI Ji;YUAN Shouqi;WANG Wenjie;YAN Aizhong;ZHAO Yun(National Research Center of Pumps,Jiangsu University,Zhenjiang,Jiangsu 212013,China;Sinoso Science and Technology Inc.,Nanjing,Jiangsu 211100,China)

机构地区：[1]江苏大学国家水泵及系统工程技术研究中心,江苏镇江212013 [2]中苏科技股份有限公司,江苏南京211100

出　　处：《排灌机械工程学报》2024年第12期1211-1220,共10页Journal of Drainage and Irrigation Machinery Engineering

基　　金：国家自然科学基金资助项目(51879121);江苏省自然科学基金资助项目(BK20190851);中国博士后科学基金资助项目(2024M751178)。

摘　　要：为了解决离心泵水力优化问题中代理模型的选择与样本数量的确定缺乏系统理论依据的问题,基于3400组不同的管道泵设计样本,针对离心泵优化设计中常用的3种代理模型(人工神经网络、响应面模型和克里金模型)进行了参数化分析.通过对比不同样本数量下3种模型的性能表现与预测精度,发现响应面模型在处理离心泵多参数优化问题时的拟合精度较差;而克里金模型和人工神经网络在单部件优化中表现出色.在多部件联合优化或复杂目标函数下,仅多隐藏层级联前馈神经网络具备一定的可靠性,但仍难以完全满足优化需求.此外,研究表明,当样本数较少(为参数总量的1.0~1.2倍)时,单隐层前馈神经网络的性能最优;而当样本数充足(大于参数总量的2.0倍)时,多隐藏层前馈神经网络与级联前馈神经网络表现出更高的可靠性.In order to solve the problem of the lack of system theoretical basis for the determination of surrogate model in the optimization of hydraulic performance of centrifugal pump,a parametric analysis was conducted based on 3400 pipeline pump design samples,three commonly used surrogate models in centrifugal pump optimization were evaluated:artificial neural networks,response surface models,and Kriging models.The results,comparing model performance and prediction accuracy across various sample sizes,show that the response surface model struggle with accuracy in multi-parameter optimization scenarios,while Kriging models and artificial neural networks excel in single-component optimizations.For more complex objectives or multi-component optimizations,only deep feedforward neural networks with multiple hidden layers demonstrate some reliability,but still fall short of meeting all optimization requirements.Additionally,the study reveals that single-layer feedforward neural networks perform best with smaller sample sizes(1.0 to 1.2 times the total number of parameters),while deep feedforward and cascade feedforward networks prove more reliable when the sample size exceeds twice the total number of parameters.

关 键 词：离心泵 优化设计 代理模型 机器学习 参数化分析 

分 类 号：TH311[机械工程—机械制造及自动化] S277.9[农业科学—农业水土工程]