采用机器学习方法预测连续刚构桥预拱度研究

Prediction of the Pre-camber of Continuous Rigid Frame Bridge Using Machine Learning

为在连续刚构桥悬臂浇筑施工阶段快速预测各悬浇段预拱度,收集土木垴大桥、庄窝大桥及南石大桥各悬臂浇筑段的预拱度及影响因素构建高维数据集,采用缺失值填充、归一化等数据预处理技术对数据集进行处理,基于梯度提升回归、极端梯度提升、支持向量机回归、随机森林及决策树5种机器学习算法,建立连续刚构桥预拱度预测模型。应用训练好的模型对西郊大桥悬浇段进行预测。结果表明:极端梯度提升在边跨预测效果最好,平均绝对误差0.97 mm、均方根误差1.28 mm,训练集确定性系数0.998,测试集确定性系数0.944,对西郊大桥边跨预测最大误差为3.9 mm;梯度提升回归在中跨预拱度预测效果最好,平均绝对误差1.4 mm、均方根误差1.63 mm,训练集确定性系数0.995,测试集确定性系数0.989,对西郊大桥中跨预测最大误差3.2 mm。研究成果满足施工要求,未来可进一步扩充数据集,提高预测精度。

In order to quickly predict the pre-camber of each cantilever section in its casting stage of continuous rigid frame bridge, the pre-camber and influencing factors of each cantilever cast section of Tumunao Bridge, Zhuangwo Bridge and Nanshi Bridge are collected to construct a high-dimensional data set. The data are preprocessed by data preprocessing techniques such as missing value filling and normalization. Based on five machine learning algorithms of gradient boosting regression, extreme gradient boosting, support vector machine regression, random forest and decision tree, the prediction model of pre-camber of continuous rigid frame bridge is established. The trained model is used to predict the cantilever casting sections of Xijiao Bridge. The results show that the extreme gradient boosting has the best prediction effect on side span, the mean absolute error is 0.97 mm, the root mean square error is 1.28 mm, the coefficient of determination of training set is 0.998, the coefficient of determination of test set is 0.944, and the maximum error of side span prediction of Xijiao Bridge is 3.9 mm. Gradient boosting regression has the best prediction effect on mid-span pre-camber, with the mean absolute error of 1.4 mm, root mean square error of 1.63 mm, coefficient of determination of training set of 0.995, coefficient of determination of test set of 0.989, and the maximum prediction error of mid-span of Xijiao Bridge of 3.2 mm. The research results meet the construction requirements, and the data set can be further expanded in the future to improve prediction accuracy.