Developing two robust hybrid models for predicting tunnel deformation in squeezing prone grounds

Abstract

Listen

Translate article

The ground deformation under squeezing conditions has a considerable impact on the tunnel excavation process. In this study, we utilize a combination of particle swarm optimization (PSO) and the Entropy Weight Method (EWM) optimization algorithms to optimize the hyperparameters of Extreme Gradient Boosting (XGBoost) and Random Forest (RF) algorithms. This optimization process aims to accurately predict tunnel deformation in squeezing prone grounds, utilizing a dataset derived from 263 tunnel case histories. In the initial step, the database underwent processing utilizing the DBSCAN method to eliminate outlier data. Subsequently, XGBoost and RF algorithms were applied within the Python environment. The optimal hyperparameters for each model were then determined through two methods: PSO and EWM. The overburden depth (H), tunnel diameter (D), support stiffness (K), and rock mass quality index (Q) serve as inputs for the tunnel deformation estimation models. The performance evaluation of the developed models, based on 10-fold cross-validation, using three statistical indices (Normalized Root Mean Square Error/NRMSE, Coefficient of Determination/R2, and Variance Accounted For/VAF), shows that the proposed XGBoost-PSO model accurately predicts tunnel deformation with acceptable performance (R2 = 0.922, NRMSE = 0.344, and VAF = 91.675). Moreover, the results regarding the importance of the input parameters suggest that the rock mass quality index (Q) holds the most significant influence, while the tunnel diameter (D) exhibits the least impact on tunnel deformation in squeezing prone grounds. This study can be a reference for predicting tunnel deformation in the squeezing prone grounds, as well as estimating the severity of the squeezing phenomenon based on deformation. It aids in estimating project costs, selecting an appropriate support system, and determining the necessary methods to mitigate the risk of squeezing.