Abstract: The authors present a predictive model for shield tunneling-induced surface settlement using a convolutional neural networks (CNN)-long short-term memory(LSTM) framework enhanced by time series generative adversarial networks(TimeGAN). This model is designed to deliver accurate predictions even with limited data volumes. The model′s effectiveness is validated through its application to the section Ⅱ of the Beiheng tunnel project in Shanghai, China. Key parameters, including construction, geological, and geometric data along with maximum surface settlement are selected to evaluate the prediction performance. Comparative analysis shows that CNN-LSTM outperforms LSTM alone, while the TimeGAN-CNN-LSTM model provides superior accuracy over CNN-LSTM. In addition, experiments adjusting the training and test data ratios reveal that the TimeGAN-CNN-LSTM model significantly improves prediction accuracy, particularly when the training-to-test set ratio is 4-8.

Key words: shield tunnel, surface settlement, convolutional neural networks, long short-term memory networks, time series generative adversrial networks