The geological information cannot be fully dynamically perceived in the shield tunneling process, which makes it difficult to accurately predict the ground settlement. To solve this problem, one kind of dynamical stratum identification model using the adaptive complex stratum changes is proposed in this paper. This method is based on the extreme gradient boosting (XGBoost) using shield construction parameters to implement the inverse deduction of stratum changes. In this way, the changing rule of construction parameters can be clarified when the stratum changes. The fusion model of error back propagation algorithm (BP) and support vector regression (SVR) for ground settlement prediction is formulated to obtain the intrinsic relationship of the ground settlement at different distances from the initial excavation face, stratum conditions and parameters of shield driving. The proposed method is validated by the 590ring data of a metro construction. Compared with the traditional method, it can achieve higher prediction accuracy.