Rolling Window Deep Learning for Hard-rock TBM Penetration Rate Prediction

Penetration rate prediction for hard rock tunnel boring machines (TBMs) remains challenging because rock mass conditions and machine operating regimes vary continuously along the tunnel alignment. Conventional static prediction models may not adequately represent such non-stationary construction conditions, particularly when geotechnical measurements are incomplete or obtained with delay during excavation. This study aims to develop a construction-phase prediction framework for ring-scale TBM penetration rate in a granite tunnel drive by integrating geotechnical data completion, sequence deep learning, and rolling-window model evaluation. A dataset of 1,000 consecutive rings was compiled, including boring-only penetration rate, thrust, torque, cutterhead rotational speed, rock mass type, and uniaxial compressive strength (UCS). Missing UCS measurements were completed using inverse distance weighting within a block model representation, resulting in estimated UCS values of approximately 33–177 MPa for rock masses dominated by massive and fractured granite. Three sequence deep learning models, namely long short-term memory (LSTM), gated recurrent unit (GRU), and temporal convolutional network (TCN), were evaluated using root mean square error (RMSE), mean absolute error (MAE), and a symmetric ±10% tolerance band adapted from accuracy-band concepts used in AACE-based project controls. The proposed rolling protocol used 100-ring validation and 100-ring test blocks to assess predictive performance under changing ground conditions. The results show that the optimized GRU model provided the most robust overall performance, achieving a mean test RMSE of approximately 0.229 m/h and a mean within-band compliance of approximately 54% across rolling folds. These findings indicate that rolling-window sequence learning can provide a practical and adaptable framework for construction-phase TBM performance prediction under evolving geological and operational conditions.