To achieve higher accuracy in predicting surface settlement resulting from shield tunnel construction, this study proposes a Bidirectional Gated Recurrent Unit model integrating a self-attention mechanism with a slime mould algorithm optimization approach, designated SA-BiGRU. The model incorporates engineering geological parameters, spatial characteristic parameters, and key shield construction parameters. Based on the interval of Haixi Village Station-Zaohu Station of Qingdao Metro Line 9 Phase I Project, the model is trained and tested by using the measured values of tunnel settlement within the defined interval together with the corresponding input parameters. Comparison experiments with BiGRU, GRU, LSTM, and BiLSTM models show that the SA-BiGRU model has the optimal prediction accuracy, with MAE, RMSE, R², and MAPE indexes of 0.13 mm, 0.36 mm, 0.993, and 6.4%, respectively. Meanwhile, the adopted SMA algorithm can effectively explore the hyperparameter space to determine the optimal configuration, and its optimization effect (MAE: 0.54 mm, RMSE: 0.46 mm, R²: 0.99, MAPE: 12.40%) is significantly better than that of the commonly used grid search, PSO, and GA. It is illustrated that the proposed SA-BiGRU model provides an innovative solution for high-exactness prediction of ground subsidence. In practical application, the model parameters are adaptable to engineering accuracy demands, allowing for improved predictive performance, which provides a new solution framework and theoretical basis for the prediction of surface settlement in shield tunnels.