Prediction models of tunnel boring machine tunneling performance based on an improved temporal convolutional network

Rational operation parameters of Tunnel Boring Machine (TBM) are crucial for efficient and safe tunneling construction. This study proposes a data-driven Temporal Convolutional Network (TCN) with a modified internal activation function to address the challenges posed by complex dynamic temporal features in TBM excavation performance prediction. A three-stage framework is adopted to accurately predict two key performance indicators: penetration rate (PR) and field penetration index (FPI). Firstly, data is processed based on the three-sigma criterion (3σ), and critical parameters essential for prediction (e.g., torque, cutterhead speed) are selected from a vast array of TBM parameters using the Extremely Randomized Trees algorithm. This enables the development of prediction models for TBM PR and FPI under varying parameters. Secondly, to manage data complexity and volume, the study further introduces the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) technique, known for its strong data analysis, noise reduction, and decomposition capabilities. A hybrid CEEMDAN-TCN model is established for tunneling thrust prediction. Finally, field data from the Rongjiang Guanbu water diversion project show that, compared to the traditional decision tree regression model, the proposed model reduces the mean absolute errors (MAE) in predicting PR, FPI, and thrust by 57.89%, 60.45%, and 3.71%, respectively. Meanwhile, the coefficients of determination (R 2 ) achieve corresponding improvements of 6.15%, 23.25%, and 9.12%, respectively. Accordingly, this research provides a practical dynamic prediction tool for intelligent TBM tunneling under the investigated geological conditions.