Machine Learning-Based Prediction Method of Subway Tunnel Settlement under the Influence of Adjacent Construction

To predict the settlement of existing subway tunnels caused by adjacent construction, a data generation method based on numerical simulation was proposed and a series of prediction models based on machine learning algorithms were established. The influence of the training set size on the prediction effect was analyzed and the results show that LightGBM has the best performance among all models based on the decision tree and the performance is positively correlated with the number of the training set. For the prediction of tunnel settlement involving multilayer soils with a complex constitutive model, the reasonable number of the training set is above the thousand level. Besides, the tunnel settlement is determined by the soil property, tunnel property, and spatial relationship between tunnels with a ratio of the importance of 3:1:1. The prediction results are associated with 24 key features that account for 90% of the contribution of all features, with 10 core features contributing more than 50%.