基于PSO-集成学习的混凝土收缩预测模型构建与分析  

Modeling and Interpretation of a PSO-Ensemble Learning Algorithm for Concrete Shrinkage Prediction

作　　者：屈兵 李学峰 赖秀英 翁向阳 林文聪 QU Bing;LI Xuefeng;LAI Xiuying;WENG Xiangyang;LIN Wencong(School of Civil Engineering,Putian University,Putian Fujian 351131,China;Fujian Lijian Inspection&Testing Group Co.,Ltd.,Putian Fujian 351131,China;Fujian Nanyu Engineering Construction Co.,Ltd.,Sanming Fujian 365001,China)

机构地区：[1]莆田学院土木工程学院,福建莆田351131 [2]福建荔建检验检测集团有限公司,福建莆田351131 [3]福建省南禹工程建设有限公司,福建三明365001

出　　处：《莆田学院学报》2025年第2期71-78,共8页Journal of putian University

基　　金：福建省自然科学基金资助项目(2024J01876);莆田市科技计划项目(2022GZ2001ptxy13)。

摘　　要：为充分考虑混凝土收缩行为中的多因素耦合效应,选取水灰比、环境湿度等14项特征参数,构建了基于集成学习AdaBoost、GBDT、XGBoost及CatBoost算法的混凝土收缩预测模型,采用粒子群优化(PSO)算法优化各模型的超参数。结果表明:PSO-CatBoost模型的预测精度最优,在测试集上的R^(2)、R_(MSE)和M_(AE)分别为0.9524、0.1099和0.0638;SHAP可解释性分析表明,对收缩预测影响最大的五个因素分别为测量时间、环境湿度、体表比、水灰比和干燥龄期,环境湿度与体表比的降低会加剧收缩变形,水灰比在约0.4~0.6范围内增加时,混凝土收缩变形增加。To comprehensively account for the complex multi-factor coupling effects in concrete shrinkage behavior,14 critical parameters such as water-to-cement ratio and ambient humidity were selected to develop ensemble learning-based concrete shrinkage prediction models using AdaBoost,GBDT,XGBoost,and CatBoost algorithms.The hyperpar ameters of each model were optimized through particle swarm optimization(PSO).Results demonstrate that the PSO-CatBoost model achieved superior predictive performance,yielding R^(2)=0.9524,R_(MSE)=0.1099,and M_(AE)=0.0638 on the test set.SHAP interpretability analysis revealed the five most influential factors as:measurement time,ambient humidity,volume-to-surface ratio,water-to-cement ratio,and drying age.Reduced humidity and volume-to-surface ratio were found to exacerbate shrinkage deformation,while concrete shrinkage increases with water-to-cement ratio elevation within the 0.4~0.6 range.

关 键 词：混凝土收缩 预测模型 集成学习 粒子群优化 机器学习 

分 类 号：TU528[建筑科学—建筑技术科学]