<p>Surface deformation during urban shield tunneling typically follows a nonlinear pattern involving initial heave followed by settlement, posing significant risks to surrounding infrastructure. Accurately capturing this evolving process remains a challenge for conventional prediction models. This study proposes a novel hybrid machine learning method—TCN-SA—that effectively models the full dynamic evolution of surface deformation. The hybrid method combines Temporal Convolutional Networks (TCN) for learning spatiotemporal patterns, Particle Swarm Optimization (PSO) for automatic hyperparameter tuning, and a Self-Attention (SA) mechanism for capturing inter-feature dependencies. Validation using field data from a Suzhou shield tunneling project demonstrates that the proposed model successfully predicts the complete deformation curve, including the transition from uplift to settlement. Compared to RNN, LSTM, and BPNN baselines, it achieves MAE reductions of 33%, 19%, and 17%; RMSE reductions of 22%, 20%, and 19%; and <i>R</i>² improvements of 13%, 7%, and 6%, respectively. Ablation experiments confirm the individual contributions of PSO and SA. The model also shows strong transferability for multi-step prediction along the tunnel alignment. By providing real-time predictions of surface behavior, this approach offers practical guidance for proactive monitoring and adaptive adjustment of shield tunneling operations.</p>