To improve the accuracy of real-time prediction of deep foundation pit deformation by existing machine learning models and enhance the dynamic learning ability of the model for complex temporal dependencies, a gated recurrent network prediction model based on attention mechanism and sand cat swarm optimization (SCSO) algorithm was constructed. The attention mechanism was coupled in the gated recurrent unit (GRU) in the model to fully explore the deep temporal correlations of deformation monitoring data, and effectively capturing key features that affect real-time deformation of the foundation pit at different time steps. At the same time, the SCSO algorithm was used to optimize the parameters of the GRU-Attention model, further improving the predictive performance of the model. The analysis of the engineering case shows that compared with traditional prediction models, the proposed model has higher prediction accuracy, significantly enhanced generalization ability, and applicability, providing a technical reference for real-time warning and safety assessment of foundation pit deformation.