Temporal–spatial coupled model for multi-prediction of tunnel structure: using deep attention-based temporal convolutional network

Predicting the mechanical behaviors of tunnel and subsurface facilities is an effective way to prevent accidental disasters. However, some drawbacks exist in many traditional prediction models, such as inadequate consideration of impacting factors, low predictive accuracy, and high computational cost. To this end, a coupled model based on deep attention-based temporal convolutional network (DATCN) is proposed for multiple prediction of structural mechanical behavior, where temporal convolutional network and self-attention mechanism are applied to learn temporal dependencies and spatial dependencies respectively. Subsequently, the DATCN model is formalized on a long-term dataset collected using a Structural Health Monitoring System in the Wuhan Yangtze River tunnel. Using three evaluation indicators, a series of data experiments are conducted to obtain the most appropriate parameters involved in the model and the superiority of DATCN over other commonly used models including LSTM, RNN, GRU, LR, and SVR is discussed. Experimental results indicate that future structural behavior shows a strong correlation between spatial dependencies and historical performance, especially that in the last 16 days. Moreover, the predictive capability of DATCN is the best compared to other commonly used models, whose predictive accuracy for the next 10 days is better than 88% and improved by 1.726% at least. Finally, the DATCN model is adopted to predict the structural behavior of the tunnel under extreme conditions as a field application, and the results suggest that the DATCN model is robust and accurate.