Abstract: To predict the time characteristics of data more accurately in foundation pit engineering, two single time series neural network models are combined, the convolutional neural network (CNN) and long short-term memory network (LSTM), as well as the gated recurrent unit (GRU), to establish a hybrid time series neural network model CNN-LSTM and CNN-GRU. An excavation project of a foundation pit adjacent to an existing station in Hangzhou is selected, and a rolling prediction method is used to create a dataset of surface settlement caused by excavation of the foundation pit in the adjacent subway stations. The predicted results are evaluated by three evaluation indexes: mean absolute error (MAE), mean relative error (MAPE) and root mean square error (RMSE). The results demonstrate that the CNN-GRU has the best prediction effects, followed by the CNN-LSTM, GRU and LSTM. Compared with the LSTM model, the CNN-LSTM hybrid network model reduces the three evaluation indexes by 24.4%, 53.8% and 4.1%, respectively, and the CNN-GRU hybrid network model decreases by 13.9%, 49.1% and 1%, respectively, compared with the GRU model.