A Segment Flotation Prediction Model for Shield Tunnel Construction Based on a Hybrid Neural Network

• Published in: Geotechnical and geological engineering Vol. 42; no. 7; pp. 5539 - 5556
• Main Authors: Wang, Xu, Zhang, Jiabing, Song, Wuyue, Guo, Fanglu, Yao, Changqing
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.09.2024; Springer Nature B.V
• Subjects: Algorithms; Buoyancy; Civil Engineering; Construction; Dredging; Earth and Environmental Science; Earth Sciences; Errors; Excavation; Flotation; Geotechnical Engineering & Applied Earth Sciences; Grouting; Hydrogeology; Mathematical models; Neural networks; Original Paper; Parameters; Prediction models; Predictions; Principal components analysis; Regression models; Segments; Terrestrial Pollution; Tunnel construction; Tunneling shields; Tunnels; Waste Management/Waste Technology; Shield tunnel; Numerical simulation; Hybrid neural network; Prediction model; Segment flotation
• ISSN: 0960-3182; 1573-1529
• DOI: 10.1007/s10706-024-02845-x

During shield tunneling construction, the buoyancy of segments occasionally occurs due to the influence of complex factors such as stratum excavation and simultaneous grouting at the shield tail. Referencing the construction of the tunnel section from Guangfu Avenue Station to Chongqing East Station of Chongqing Rail Transit Line 27 crossing the Kuxi River, a shield tunnel construction period segment buoyancy prediction model based on ISSA-XGBoost-GRU-SVM is proposed. First, the principal component analysis method is employed to consider the influence of shield excavation parameters on segment buoyancy, and a set of new features highly correlated with the principal components is selected as input. Second, to address the difficulty in manually configuring parameters and the randomness of the XGBoost model, an improved sparrow algorithm is used to optimize the model hyperparameters. Finally, leveraging the advantages of SVM in feature extraction and GRU neural networks in time series prediction, the ISSA-XGBoost-GRU-SVM hybrid neural network model is constructed by updating the weights of the prediction results using the reciprocal of the errors. The results show that the mean regression values at the top and bottom measurement points of the proposed model are 0.9, and the root mean square errors at the top and bottom measurement points decrease significantly compared to those of traditional algorithms. Compared to the numerical simulation results, the proposed model can better describe the changing characteristics of segment buoyancy during the accelerated uplift phase. Therefore, the algorithm of the proposed model in this paper has greater predictive effectiveness and accuracy.