Real-time Prediction Method of Shield Tunneling Attitude under Complex Geological Conditions

Accurate prediction of shield tunneling attitude under complex geological conditions is critical for ensuring the safety and efficiency of tunnel construction. Excessive deviations in tunneling attitude may lead to significant engineering risks, such as tunnel alignment deviations, structural collisions, or ground subsidence. This study, based on the Qingdao Jiaozhou Bay Second Subsea Tunnel project, proposes a method for real-time prediction of shield tunneling parameters in complex geological conditions by comparing the performance of three mainstream machine learning models: XGBoost, Transformer, and LSTM. The proposed method follows a systematic process. During the data preprocessing phase, stoppage data and anomalies were removed, and discrete wavelet transform (DWT) was applied for noise reduction. Bayesian optimization (BOA) was used to fine-tune the hyperparameters of each model to enhance prediction accuracy. Experimental results demonstrate that the BOA-XGBoost model achieved an average prediction accuracy of 0.98, with errors for all five output parameters remaining below 2%. BOA-XGBoost outperformed BOA-LSTM and BOA-Transformer across multiple accuracy evaluation metrics. However, BOA-LSTM exhibited faster and more stable convergence, achieving the highest overall prediction efficiency-33.44% higher than that of XGBoost and 96.79% higher than that of Transformer. This study highlights the importance of data preprocessing techniques and provides an in-depth analysis of the strengths and weaknesses of these three models in real-time shield tunneling attitude prediction under complex geological conditions.