Developing hybrid XGBoost model integrated with entropy weight and Bayesian optimization for predicting tunnel squeezing intensity

Abstract

Abstract Tunnel squeezing is a large deformation problem closely related to creep and severely affects tunnel construction safety and efficiency. In this paper, the extreme gradient boosting (XGBoost) model is optimized using a combination of the Bayesian optimization (BO) algorithm and the entropy weight method (EWM) to accurately predict the tunnel squeezing intensity based on a dataset of 139 tunnel case histories. In order to mine the information contained in the prediction indices, the EWM is used first to pre-process the sample data and eliminate the effect of large differences in the input parameters’ values among different dimensions. On the other hand, the BO algorithm is applied to optimize the XGBoost model’s important hyperparameters, thus improving its performance effectively. As a part of the study, the strength-stress ratio (SSR), rock mass quality index in the BQ system ([BQ]), tunnel diameter (D), and support stiffness (K) are selected as inputs to the tunnel squeezing estimation model. Within the study context, the prediction accuracy (Acc) and kappa coefficient (k) of the EWM-BO-XGBoost, XGBoost, BO-XGBoost, Evidence Theory (ET), Random Forest (RF), Support Vector Machine (SVM), and Decision Tree (DT) models are computed and compared. The study results have shown that the Acc (91.7%) and k (0.89) of the EWM-BO-XGBoost model are the highest, which proves its reliability and superiority against other alternatives. In addition, the analysis of the prediction indices’ feature importance showed that the SSR contributes the most to the squeezing intensity, followed by the [BQ] and D, while the K has the least effect on the squeezing intensity.