基于物理信息神经网络的长距离顶管施工顶力预测  

Jacking force prediction of long-distance pipe jacking construction based on physical information neural network

作　　者：李博 刘宇翔 陈建国 杨耀红 张哲 LI Bo;LIU Yuxiang;CHEN Jianguo;YANG Yaohong;ZHANG Zhe(Henan Water Conservancy Engineering First Group Co.,Ltd.,Zhengzhou 450000,China;School of Water Conservancy,North China University of Water Resources and Electric Power,Zhengzhou 450046,China;Henan Provincial Key Laboratory of Hydrosphere and Watershed Water Security,Zhengzhou 450046,China)

机构地区：[1]河南省水利第一工程局集团有限公司,河南郑州450000 [2]华北水利水电大学水利学院,河南郑州450046 [3]河南省水圈与流域水安全重点实验室,河南郑州450046

出　　处：《人民长江》2025年第1期147-155,共9页Yangtze River

基　　金：河南省学科创新引智基地项目“智慧水利”(GXJD004);河南省水利厅科技攻关项目“长距离大直径复杂地层顶管施工关键技术研究及应用”(GG202358)。

摘　　要：长距离顶管施工过程中,准确预测顶力是有效控制施工安全质量及进度的关键问题。基于知识数据融合的机器学习建模方法,将顶力计算物理模型与多层感知机相融合,构建了物理-数据双驱动的物理信息神经网络模型(PINN),用物理机制约束神经网络的训练机制,并引入改进的麻雀搜索算法(ISSA)对模型超参数取值进行优化,建立了ISSA-PINN顶管施工顶力预测模型;以河南省郑开同城东部供水工程顶管施工为例,选取524组工程实测数据验证了模型的有效性。计算结果表明:ISSA-PINN模型具有较高的预测精度,相较于单纯数据驱动模型,在测试集和新数据集中的预测性能分别提升了0.07和0.17,说明物理模型的融入对降低机器模型的过拟合风险和提高泛化能力有积极影响;相比于SSA和粒子群算法,ISSA算法寻优速度更快、适应度更好。研究结果可为顶管工程施工顶力控制提供参考。Accurately predicting the jacking force is a crucial for effectively controlling the safety,quality,and schedule of the long-distance pipe jacking project.In this paper,based on the machine learning modeling method of knowledge data fusion,the jacking force calculation physical model was combined with the multi-later perceptron to construct a physics-informed neural network model(PINN)driven by physics and data,the physical mechanism was used to constrain the training mechanism of the neural network,and the improved sparrow search algorithm was introduced to optimize the hyperparameter value of the model.Therefore,the ISSA-PINN jacking force prediction model was established.Taking the water supply project for Zhengzhou-Kaifeng City demonstration area as an example,524 sets of actual measurement data were selected to validate the model's effectiveness.The results showed that,compared to data-driven models,the ISSA-PINN model achieves high prediction accuracy with 0.07 and 0.17 improvement on test datasets and new datasets respectively.This demonstrated that incorporating physical models have positive effects to reducing over fitting and enhancing generalization of MLP model.The ISSA algorithm demonstrated faster optimization speed and better fitness compared to the standard Sparrow Search Algorithm(SSA)and Particle Swarm Optimization(PSO).The research results can provide reference for jacking force control in pipe jacking engineering construction.

关 键 词：顶管施工 顶力预测 物理信息神经网络(PINN) 改进麻雀搜索算法(ISSA) 

分 类 号：TV672.2[水利工程—水利水电工程] TP181[自动化与计算机技术—控制理论与控制工程]