PSO-based Machine Learning Methods for Predicting Ground Surface Displacement Induced by Shallow Underground Excavation Method

Four hybrid intelligent methods are developed to predict the maximum ground surface settlement ( S max ) induced by shallow underground excavation method (SUEM). Particle swarm optimization (PSO) algorithm with k -fold cross validation is used to determine the optimal hyperparameters or random parameters in the four machine learning (ML) methods, namely that, back-propagation neural network (BPNN), extreme learning machine (ELM), support vector regression (SVR) and random forest (RF). 100 field engineering samples are collected from published papers. In each data sample, the effect of stratum mechanical conditions, tunnel geometric parameters and construction parameters on S max is considered. Correlation laws among parameters are investigated through Pearson correlation coefficient, data distribution histogram and correlation confidence ellipse. The performance of four PSO-based ML methods is comprehensively compared by fitness function, time cost and prediction accuracy in the training and test processes. PSO-RF outperforms PSO-SVR, PSO-ELM and PSO-BPNN in the prediction accuracy for S max owing to larger R 2 , smaller MAE and RMSE . Calculation time that the optimal hyperparameters are determined is the fastest for PSO-RF, and PSO-ELM has the smallest fitness function. The prediction performance of PSO-RF method for construction parameters is also discussed. This work can provide theoretical guidance for design and construction of SUEM.