关键词: 盾构隧道, 管片上浮, 长时间序列预测问题, Informer模型, SHAP方法

Abstract: Utilizing machine learning to predict the uplift of shield tunnel linings ahead of the cutterhead during construction enables timely adjustments of control parameters, mitigating lining uplift issues. Nevertheless, existing models exhibit limited performance in long sequence time-series forecasting (LSTF) and face challenges in accurately predicting the uplift of multiple lining rings ahead of the shield cutterhead. Considering the impact of shield control, attitude parameters, and geological condition, and utilizing the Boruta algorithm to determine model input features, a shield tunnel segment uplift prediction model based on LightGBM-Informer was proposed. This model incorporates a wavelet transform filter and a complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) method to eliminate noise in time series data. The accuracy and applicability of the proposed model were validated using the monitoring data from subway shield tunnel projects in Nanjing and Xiamen. The results demonstrate that the model exhibits enhanced prediction accuracy in comparison to other models, including recurrent neural network (RNN), long short-term memory (LSTM), gated recurrent unit (GRU), and Transformer. Additionally, it demonstrates robust generalization capabilities across diverse geological condition datasets. As the length of the prediction sequence increases, the performance advantages of the model become more pronounced, accurately predicting the uplift of 1−2 rings of linings ahead of the shield cutterhead. Feature importance analysis based on Shapley additive explanations (SHAP) method indicates that earth chamber pressure and vertical displacement at the shield head and tail have significant impacts on lining uplift. The model provides theoretical guidance for intelligent control of shield tunnel lining construction in complex, water-rich environments.

Key words: shield tunnel, lining uplift, long sequence time-series forecasting problem, Informer model, Shapley additive explanations (SHAP) method