Real-time prediction method of shield tunneling attitude under complex geological conditions

Abstract Accurate prediction of shield tunneling attitude is crucial for preventing engineering risks such as tunnel misalignment, structural collisions, and ground subsidence in underground construction projects. This study develops a real-time prediction method for shield tunneling parameters using the Qingdao Jiaozhou Bay Second Subsea Tunnel as a case study. We systematically compare three machine learning approaches (XGBoost, Transformer, and LSTM) enhanced with discrete wavelet transform (DWT) for noise reduction and Bayesian optimization (BOA) for hyperparameter tuning. Results show that BOA-XGBoost achieved the R 2 is 0.98 with the Mean Absolute Percentage Error (MAPE) errors below 2% for all five output parameters (horizontal/vertical deviations at shield head/tail and roll angle). While BOA-XGBoost outperformed other models in accuracy, BOA-LSTM demonstrated superior computational efficiency—33.44% faster than XGBoost and 96.79% faster than Transformer. Under complex geological conditions, deep learning models (BOA-LSTM and BOA-Transformer) showed better adaptability than BOA-XGBoost, particularly in mixed soil-rock strata. This study provides practical guidance for selecting appropriate prediction models based on geological conditions and computational constraints in tunnel engineering projects.