Abstract: Surface settlement prediction plays a crucial role in ensuring infrastructure safety and managing construction risks.However,single predictive models often suffer from limited generalization capability,noticeable systematic bias,and poor responsiveness to outliers in surface settlement prediction.This study proposes a coupled machine learning approach for surface settlement prediction.Using settlement monitoring data from a metro station project in Changchun as the engineering background,three machine learning models—CatBoost,GBDT,and AdaBoost—are integrated to predict surface settlement.The results indicate that the coupled model effectively eliminates systematic deviation and suppresses abnormal peak fluctuations.The coefficient of determination (R²)reachs 0.97,while the mean absolute error (MAE)and root mean square error (RMSE)decrease to 0.007 3 and 0.009 6,respectively,demonstrating superior predictive accuracy and stronger generalization capability.The proposed coupled model provides reliable technical support for long-term deformation monitoring and risk early warning in metro projects.It also offers a methodological reference for enhancing the accuracy of surface settlement prediction and ensuring construction safety,with potential applicability to other geotechnical engineering scenarios.

Key words: surface subsidence, machine learning, coupled model, subsidence prediction