Prediction of EPB shield advance rate based on NGOoptimized GRU model and LIME analysis

Accurate prediction of shield advance rate ( AR ) is crucial for optimizing construction efficiency and ensuring construction safety. To overcome the low accuracy of traditional machine learning models and the black-box nature of deep learning, this paper takes the earth pressure balance (EPB) shield section of Beijing Metro Line 19 as the engineering case. Following preprocessing and Pearson analysis, 11 key features were selected to construct four prediction models. To address the tendency of empirical hyperparameter tuning to stall at local optima, the North Goshawk Optimization (NGO) algorithm was introduced to optimize the Gated Recurrent Unit (GRU) model, which has better performance. Furthermore, an independent dataset from Shenyang Metro Line 3 was used to verify the cross-project robustness of the proposed framework. Results indicate that the NGO-GRU model achieved optimal performance, with a test set R 2 of 0.934 and MSE of 2.811, significantly outperforming other models. Even on the unseen Shenyang dataset, the model accurately tracks AR dynamic changes. Locally Interpretable Model-Agnostic Explanations (LIME) analysis indicates that penetration contributes most significantly (67%) to AR . Among controllable parameters, cutterhead speed ( CS ) and speed of screw conveyor ( SSC ) exhibit notable contributions of 21.2% and 4.7%, respectively. During construction, CS and SSC may be appropriately increased to enhance AR while balancing cutter wear and earth chamber pressure. This study provides a reference for optimizing shield construction parameters and enhancing excavation efficiency.