The prediction method for ground surface settlement of pipe jacking tunnels based on a spatiotemporal transfer learning network

Surface settlement induced by ground disturbances during pipe-jacking tunnel construction poses a persistent risk, with progressive deformation potentially leading to surrounding structural collapse. Accurate settlement prediction provides advance risk warning and enables proactive deformation risk control through construction parameter optimization. Conventional deformation control is based on simplified analytical models, numerical simulations, and field monitoring, while the effectiveness is constrained by theoretical assumption and sparse field monitoring data. The Long Short-Term Memory-Convolutional Neural Network model with Transfer Learning (LSTM-CNN-TL) is proposed to achieve settlement prediction under data-scarce conditions. The LSTM module captures long-range temporal dependencies in settlement sequences, while the CNN extracts spatial correlations and inter-feature interactions across monitoring points. Transfer learning is implemented through historical data pretraining followed by fine-tuning on target project data, enabling cross-project domain adaptation and improved generalization with limited monitoring input. The LSTM-CNN-TL model is evaluated on a soft soil pipe-jacking project. Comparative experiments against multiple baseline models across various evaluation metrics demonstrate its consistent superiority in prediction accuracy, robustness, and adaptability under limited monitoring conditions.