Prediction of TBM tunneling speed based on TCN-LSTM-Attention model

In order to predict the TBM tunneling speed more accurately, based on the analysis of advantages and disadvantages of common machine learning models, this paper proposed a TCN-LSTM-Attention model based on spatio-temporal feature extraction. Firstly, the original data were preprocessed by using isolated forest (IF) and ensemble empirical mode decomposition (EEMD) to eliminate outliers and high-frequency noise. Secondly, the temporal convolutional neural network (TCN) was used to extract spatial features, and then the long short-term memory network (LSTM) was used to extract temporal features. Finally, the spatio-temporal features were weighted by the attention mechanism to realize the accurate fitting of the nonlinear relationship. The engineering application results showed that: ① Compared with the untreated data, the performance of the model was significantly improved after data pretreatment by IF and EEMD, for the Ⅱ, Ⅲ_a, Ⅲ_b, Ⅳ and Ⅴ grade surrounding rocks, the determination coefficient (R²) of the model was increased by 47.69 %, 43.77 %, 42.79 %, 42.25 % and 36.88 % respectively. ② The prediction accuracy of the proposed model on all levels of surrounding rock was better than that of RF, SVR and LSTM, and the attention mechanism can effectively improve the model performance. ③ Attention weight analysis showed that similar historical information in time and space had a greater impact on the prediction of future information. ④ In the data set of surrounding rock combination at all levels, the model also achieved good performance, with mean square error (MSE) of 3.743 3 (mm/min)² and R² of 96.22 %.