To enhance the durability and safety of transportation infrastructure, accurate prediction of subgrade deformation is essential. A predictive method fusing a bidirectional long short-term memory (Bi-LSTM) network with a self-attention (SA) mechanism was proposed. An improved particle swarm optimization (IPSO) algorithm was employed for hyperparameter optimization to develop an intelligent prediction model for subgrade deformation (settlement and horizontal displacement), designated as IPSO-Bi-LSTM-SA (I-B-L-S). The established model was validated using field-measured data. The results indicate that under identical input conditions, the I-B-L-S model exhibited smaller prediction errors compared to the BP, Bi-LSTM (B-L), and Bi-LSTM-SA (B-L-S) models, and the predicted varying trend of results is closer to the measured data. Based on the predicted results, performance evaluation metrics for subgrade displacement, including mean absolute error, root mean square error, mean absolute percentage error, and the coefficient of determination, were calculated. The established I-B-L-S model demonstrated significant advantages across all four metrics, indicating that superior accuracy and robustness are achieved, and can provide a theoretical reference for subgrade deformation prediction.