Abstract: People have always been concerned about the velocity and safety of shield tunneling, which have a close relationship with stratum conditions and operation parameters. However, it is difficult to predict the shield performance using traditional methods due to the complicated shieldground interaction, impeding the development of the rapid tunneling technology. In this paper, a new prediction method for shield advancing speed based on the AttentionResNetLong Shortterm Memory(LSTM) hybrid model is proposed to address this problem. Compared with traditional neural networks, such as LSTM and GRU, the proposed model has introduced the Attention mechanism, which can update the weight matrix adaptively according to various stratum conditions during longdistance shield tunneling. As a result, the model can modify the weight matrix itself when confronted with different tasks. This ability effectively increases the prediction accuracy, which has been validated through realtime prediction of shield advancing speed in the eastern route of the ChinaRussia gas pipeline project. The method can analyze the correlations between input and output parameters during shield tunneling. Moreover, it has good adaptability, which is of great significance in selecting the operation parameters for the velocity and safety of shield tunneling.

Key words: shield tunnel, artificial intelligence, hybrid neural network, property prediction, advance rate