To improve the prediction accuracy of soil settlement induced by shield tunnel construction, a deep learning model was proposed that combined the self-attention (SA) mechanism and multi-head self-attention (MHSA) mechanism separately with the bidirectional long short-term memory (Bi-LSTM) model, effectively capturing the spatiotemporal features and key information within the data. Using the time-series data from multiple sensors as inputs, the model employed a multi-layer bidirectional network architecture and attention mechanisms to capture the vital data features and their internal self-correlation. Based on the actual soil settlement data from a shield tunnel project, hyperparameters such as the number of hidden units and the number of attention units were optimized through cross-validation, and the predictive effects on soil settlement for the Bi-LSTM model before and after the introduction of various attention mechanisms were compared. Results show that the MHSA-Bi-LSTM model achieved optimal performance, with its total mean absolute percentage error (1.27%) showing approximately a 46% decrease over the SA-Bi-LSTM model (2.53%). Both models maintained high prediction accuracy for soil settlement across various engineering scenarios without parameter recalibration, exhibiting total mean absolute percentage errors of 9.06% for the MHSA-Bi-LSTM model and 14.82% for the SA-Bi-LSTM model, indicating strong generalization capability.