Abstract: To develop an efficient prediction model for tunneling parameters in full-face hard rock tunnel boring machines (TBMs) that can assist real-time parameter adjustments, a case study is conducted on the Chao′er river to Liaoxi river water-diversion project, and the data from the project are rigorously processed using the 3σ criterion and singular value decomposition method. Following this, the input parameters are selected based on the grey relational analysis method. To enhance the predictive capabilities of the model, a self-attention mechanism (SAM) is integrated into a long short-term memory (LSTM) network, and an Optuna framework is then employed to search the optimal combination of hyperparameters, thus finally establishing an OptunaSAM-LSTM prediction model. The effectiveness of the Optuna-SAM-LSTM model is validated through 10 search trials conducted using the Optuna framework. The results consistently demonstrate the following: (1) The model trained with the hyperparameters identified by Optuna achieves high predictive accuracy and exhibits remarkable stability across different trials. (2) When predicting the penetration rate of the TBM cutterhead, the SAM-LSTM model displays a high degree of fit and low prediction error over the next three time steps, indicating its robustness in forecasting key operational parameters. Comparative analysis with other models, including LSTM, SAM-recurrent neural network (RNN), and RNN, reveals that the SAM-LSTM model significantly outperforms these alternatives in terms of prediction accuracy. Furthermore, models incorporating the SAM consistently achieves higher accuracy than their conventional counterparts, underscoring the value of this approach. (3) The Optuna-SAM-LSTM model demonstrates exceptional performance in the advance prediction of TBM cutterhead penetration rate. Its application in real-world engineering projects can provide operators with ample decision-making time to adjust parameters, thereby ensuring the safety and efficiency of TBM operations.

Key words: tunnel engineering, full-face tunnel boring machine, tunneling parameter prediction, deep learning, Optuna procedure, self-attention mechanism