Uncertainty-Aware Forecasting of Shield Tunneling Parameters Using Variational Bayesian Neural Networks

Abstract Optimizing the tunneling parameters of shield machines is crucial for ensuring project safety and efficiency. However, traditional deterministic prediction models struggle to address uncertainties in complex geological conditions. This paper proposes a probabilistic prediction framework for tunneling parameters based on Bayesian neural networks (BNNs). Compared with conventional machine learning models, the proposed approach not only provides point estimates for shield tunneling parameters but also generates complete posterior prediction distributions, thereby quantifying the uncertainty of predictions. A BNN model structure comprising 10 input features, a single hidden layer, and 7 output nodes is adopted, and variational inference (VI) is employed to achieve efficient posterior inference. Validation using real shield construction data sets demonstrated the model’s superior performance in predicting multiple critical parameters, including cutterhead rotational speed, torque, and total thrust: the coefficient of determination ( R 2 ) reached 0.963 with 70% training data, and the 90% prediction interval coverage consistently was near or above 90%, confirming the effectiveness of both prediction accuracy and uncertainty quantifications. The proposed model maintains robust performance even with limited training data. Sensitivity analysis of prior distributions confirmed that a standard normal prior constitutes a robust choice in this study. The proposed BNN framework provides a reliable data-driven approach for intelligent control and risk decision-making in shield tunneling.