一种基于EMD-LightGBM模型的地铁隧道盾构姿态预测方法

A prediction method for shield tunneling attitude in subway tunnels based on EMD-LightGBM model

关键词:  地铁隧道  盾构  姿态预测  经验模态分解  轻量级梯度提升机

Key words:  subway tunnel  shield tunneling  attitude prediction  empirical mode decomposition  light gradient boosting machine

摘要:

针对地铁隧道盾构姿态难以控制的问题,以长春某隧道工程为例,基于现场实测数据,构建了一个融合经验模态分解（empirical mode decomposition,EMD）和轻量级梯度提升机（light gradient boosting machine,LightGBM）的盾构姿态预测模型（EMD-LightGBM）。首先,通过特征重要性和相关性分析筛选原始数据集特征。然后,利用EMD技术将数据分解为多个平稳子序列,并组成新数据集。最后,通过该新数据集拟合训练EMD-LightGBM来实现盾构姿态的预测,并且比较了该模型与单纯的LightGBM及融合EMD的反向传播神经网络（backpropagation neural network,BPNN）的预测效果。通过预测精度和预测稳定性两种评价体系来验证EMD-LightGBM模型的优良性能。结果表明:与LightGBM和EMD-BPNN相比,EMD-LightGBM在盾构姿态偏差预测折线图中的表现最佳,其平均绝对误差（mean absolute error,E_MA）和均方根误差（root mean square error,E_RMS）最大分别为2.89 mm和4.13 mm,决定系数R²最小值为0.95；同时,EMD-LightGBM的预测平均绝对误差E_MA和均方误差（mean square error,E_MS）的95%置信区间最大值分别为3.5 mm与25.6 mm²,结合其预测值的绝对误差（absolute error,E_A）和平方误差（square error,E_S）的良好频数分布,都说明了EMD-LightGBM在预测盾构姿态时的高精度和稳定性。研究成果可为类似工程的盾构姿态控制提供一种理论方法。

Abstract:

In response to the problem of difficult control of shield tunneling attitude in subway tunnels, taking a tunnel project in Changchun as an example, a shield tunneling attitude prediction model (EMD-LightGBM) was constructed based on on-site measured data, which integrates empirical mode decomposition (EMD) and lightweight gradient boosting machine (LightGBM). Firstly, filter the features of the original dataset through feature importance and correlation analysis. Then, the data is decomposed into multiple stationary subsequences and combined into a new dataset by EMD. Finally, EMD-LightGBM was fitted and trained by the new dataset to achieve the prediction of shield tunnel attitude, and the prediction performance of the model was compared with that of LightGBM alone and the EMD-BPNN. Verify the excellent performance of the EMD-LightGBM model through two evaluation systems: prediction accuracy and prediction stability. The results showed that compared with LightGBM and EMD-BPNN, EMD-LightGBM performed the best in predicting shield attitude deviation in the line graph, with a maximum mean absolute error (E_MA) and root mean square error (E_RMS) of 2.89 mm and 4.13 mm, respectively, and a minimum coefficient of determination R² of 0.95. Meanwhile, the maximum 95% confidence intervals for the E_MA and mean square error (E_MS) of EMD-LightGBM predictions are 3.5 mm and 25.6 mm², respectively. Combined with the good frequency distribution of its predicted absolute error (E_A) and square error (E_S), it demonstrates the high accuracy and stability of EMD-LightGBM in predicting shield tunnel attitudes. The research results can provide a theoretical method for the attitude control of shield tunneling in similar projects.