Abstract: During the construction of deep foundation pit, the horizontal displacement of composite wall is affected by construction conditions, soil properties and surrounding environment. It is of great significance to accurately predict the horizontal displacement of composite wall during construction. Aiming at the problem of insufficient accuracy of single deep learning model for synchronous prediction of multi-depth horizontal displacement of composite wall, this paper takes the deep foundation pit of a subway station in Wuhan rail transit as the engineering background, and proposes a fusion prediction model combining simulated annealing algorithm (SA) and convolutional neural network (CNN). The model uses the actual monitoring data to synchronously predict the multi-depth (19 m-22 m) deformation of the composite wall, and carries out comparative verification. The results show that the SA-CNN model can accurately predict the deformation at different depths, and the prediction effect of the measuring point CX3-21 (depth of 21 m) is the best. The root mean square error (RMSE) is 0.28 mm, the mean absolute error (MAE) is 0.22 mm, and the coefficient of determination (R²) is 0.96. The RMSE of the remaining depths (CX3-19, CX3-20, CX3-22) ranged from 0.26 to 0.30 mm, the MAE ranged from 0.22 to 0.25 mm, and the R² was not less than 0.94. The research results show that the model can effectively improve the prediction accuracy of multi-depth horizontal displacement of composite wall, which can provide reference and method for the deformation prediction of composite wall in similar deep foundation pit engineering.