基于改进特征筛选的随机森林算法对锂渣混凝土强度的预测研究  

Using Random Forest with Improved Variable Selection to Predict the Compressive Strength of Concrete with Lithium Slag

作　　者：魏令港 黄靓[1] 曾令宏[1] WEI Linggang;HUANG Liang;ZENG Linghong(School of Civil Engineering,Hunan University,Changsha 410082,China)

机构地区：[1]湖南大学土木工程学院,长沙410082

出　　处：《材料导报》2024年第9期156-161,共6页Materials Reports

基　　金：2021年产业技术基础公共服务平台-重点原材料行业碳达峰、碳中和公共服务平台项目(2021-H029-1-1)。

摘　　要：本工作提出了特征变量筛选结合特征变量相关性的方法,对构建的锂渣混凝土28 d抗压强度数据库进行优化,分别建立了随机森林模型和深度神经网络模型用于测试数据库,并以相关系数(R)、均方根误差(RMSE)和平均相对误差(MAE)三个指标对模型的预测结果进行对比分析。结果表明,预测锂渣混凝土的28 d抗压强度时,采取改进的特征变量筛选方法能够有效提高模型的预测效果,此外,特征变量筛选的前后随机森林(RF)模型的预测效果明显优于深度神经网络(DNN)模型。This work proposed the method of feature variable screening combined with feature variable correlation to optimize the constructed 28-day compressive strength database of lithium-slag concrete,and established random forest model and deep neural network model for testing the database,respectively,and compared the prediction results of the models with three indicators:correlation coefficient(R),root mean square error(RMSE)and mean relative error(MAE).The results show that the improved feature variable screening method can effectively improve the prediction effect of the model when predicting the 28-day compressive strength of lithium-slag concrete,and in addition,the prediction effect of the before-and-after random forest(RF)model with feature variable screening is significantly better than that of the deep neural network(DNN)model.

关 键 词：随机森林(RF) 深度神经网络(DNN) 特征变量筛选 锂渣混凝土 抗压强度 

分 类 号：TU528[建筑科学—建筑技术科学] TP312[自动化与计算机技术—计算机软件与理论]