Machine learning models to predict the tunnel wall convergence

Abstract

Deformation and damage of the rock mass induced by excavation operations can result in the tunnel wall convergence. Accurate prediction of the convergence is crucial for ensuring safe, economical, and efficient construction of tunnel projects. In this study, six popular and reliable machine learning (ML) models, including back-propagation neural network (BPNN), general regression neural network (GRNN), extreme learning machine (ELM), kernel extreme learning machine (KELM), least squares support vector machine (LSSVR) and random forest (RF) were selected to predict the tunnel wall convergence. The jellyfish search optimizer (JSO) was utilized to find the optimal hyperparameters of these models. A total of 142 sets of highway tunnel convergence data were collected for model training and testing, six parameters including overburden thickness (H), rock mass rating index (RMR), rock mass quality index (Q), rock mass cohesion (Crm), deformation modulus of rock mass (Erm), and internal friction angle of rock mass (φrm) were used as input variables and the mean convergence of tunnel wall (Y) was considered as the output variable. Mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R2), and variance accounted for (VAF) as evaluation indicators were used to compare the predictive performance of the proposed models. The results from various evaluation methods demonstrate that the overall performance of the JSO-RF model is superior, with MAE, RMSE, R2, and VAF values of 4.288, 6.219, 0.917, and 89.480 during the training phase, and 3.141, 3.958, 0.939, and 93.164 during the testing phase. When compared with RF, multivariate regression (MVR), and empirical equations, the JSO-RF model also exhibits better predictive performance. The importance results of all input variables showed that φrm, Erm, and H were more important input variables than other variables.