Abstract: In order to improve the accuracy of foundation pit deformation prediction, a GCN- Informer prediction model based on graph convolutional network (GCN) and Informer algorithm is proposed. The weighted adjacency matrix was constructed according to the location of each monitoring point, and GCN was used to extract spatial features from each time series data. Informer was used to learn the time features, integrate the spatial- temporal information, and input the fully connected layer to obtain the prediction results. The model was applied to the vertical displacement prediction of the top of the retaining wall of the construction foundation pit of a station in Shanghai. The results show that compared with the time series prediction models Informer, Transformer, GRU and LSTM, the Mean Absolute Error (MAE) is decreased by 16.94%, 35.53%, 50.42% and 48.93%, and Mean Absolute Percentage Error (MAPE) is decreased by 15.84%, 32.37%, 48.52% and 47.09%, respectively. The Root Mean Square Error (RMSE) is reduced by 14.53%, 37.57%, 46.31% and 46.89% respectively, indicating a high prediction accuracy. The GCN- Informer model can provide a reference for the deformation prediction of the vertical displacement of the top of the retaining wall of similar foundation pits.