A Closed-Loop Feature Identification Framework for Key Parameter Screening and Multi-Step Prediction of Shield Tunnelling Attitude Deviation

Accurate attitude deviation prediction is key to segment quality during shield tunnelling, yet existing data-driven approaches typically treat feature selection and predictive modelling as independent steps, limiting interpretability and accuracy. This study proposes TransFeatNet, a key-parameter identification framework integrating multi-source feature selection, Transformer-based temporal encoding, threshold adaptive optimisation and attention-guided feature analysis into a closed-loop process. On this basis, a TransFeatNet-guided multi-step prediction method is developed using the Temporal Convolutional Network (TCN), Gated Recurrent Unit (GRU) and Decomposition Linear (DLinear) as backbones under a unified input and data split. Validation on the Zheng–Yan section of Beijing Metro Line 22 shows that TransFeatNet compresses 31 candidate parameters to the top-10 key variables for both horizontal and vertical deviation tasks, with cumulative weights of 0.6544 and 0.6642, markedly higher than the 0.4704 and 0.4854 of a Transformer-only baseline. Compared with Transformer-guided inputs, TransFeatNet further reduces RMSE by 12.50–20.94% and raises R2 by 0.045–0.057. The optimal TransFeatNet + TCN combination achieves R2 of 0.9210 and 0.8930 on an independent section, indicating stable cross-section performance under the current engineering data conditions. The framework provides a compact, interpretable input representation and shows clear engineering potential for attitude early warning and assisted control.