Time series prediction of shield construction ground settlement based on the fusion method of VMD-CNN-BiLSTM model and attention mechanism
GUO Junhao1, YAN Mengyao1, LI Jie1, ZHOU Lei1, ZHAO Xing1, HUANG Fu2, REN Ziheng2
(1. China Railway 16th Bureau Group Corporation Limited Beijing 100018, China; 2. School of Civil and Environmental Engineering, Changsha University of Science & Technology, Changsha 410114, Hunan, China)
Abstract: In surface settlement prediction induced by shield tunneling, traditional empirical formulas and conventional single machine learning models often fell short in adequately capturing the nonlinear spatiotemporal characteristics embedded in multi-source parameters. To address this challenge, a deep learning-based predictive model was proposed in this paper for effectively representing the dynamic interdependencies of stratum disturbance responses under complex geological conditions. Specifically, a hybrid VMD–CNN–BiLSTM–Attention framework was constructed by synergistically integrating Variational Mode Decomposition (VMD), Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and the Attention Mechanism. Within the proposed framework, VMD was first applied to decompose the raw settlement time series into multiple intrinsic mode functions, thereby facilitating noise reduction and latent pattern extraction. Subsequently, CNN was employed to extract spatial correlation features from both shield operational parameters (e.g., thrust, torque) and geotechnical properties (e.g., elastic modulus, cohesion, permeability coefficient). The high-level feature representations were then processed by the BiLSTM module to model bidirectional long-term temporal dependencies, while the attention mechanism was utilized to dynamically assign adaptive weights to salient time steps, thus enhancing both predictive accuracy and model interpretability. The model was trained and validated using field-monitored surface settlement data obtained during the shield-driven construction of the eastern extension of Changsha Metro Line 6, where the tunnel was successfully underpassed a critical airport runway. Comparative evaluation against in-situ measurements demonstrated that the proposed model achieved a low root mean square error (RMSE) on the test set and a high coefficient of determination (R2> 0.96), which significantly outperformed conventional approaches in prediction accuracy. The results confirmed that the VMD–CNN–BiLSTM–Attention framework provided a robust and reliable tool for real-time settlement forecasting, thereby offering strong theoretical support for settlement control and early risk warning in shield tunneling projects.
郭军浩, 颜梦瑶, 李杰, 周雷, 赵兴, 黄阜, 任子衡. 基于VMD-CNN-BiLSTM模型与注意力机制融合的盾构施工诱发地表沉降时序预测[J]. 隧道与地下工程灾害防治, .
GUO Junhao, YAN Mengyao, LI Jie, ZHOU Lei, ZHAO Xing, HUANG Fu, REN Ziheng. Time series prediction of shield construction ground settlement based on the fusion method of VMD-CNN-BiLSTM model and attention mechanism. Hazard Control in Tunnelling and Underground Engineering, 0, (): 1-14.
