To improve the stability and interpretability of ground surface settlement prediction under small-sample and incremental-data conditions, a mechanism-driven feature grouping and layered additive residual learning method was proposed. Input variables were grouped into geological-environmental, shield operation, and grouting and segment attitude parameters, and a layered learning path was constructed according to geological control, construction disturbance, and grouting and segment-attitude compensation. The method was validated using field data from the super-large-diameter shield construction of the Beiheng Passageway in Shanghai. The results show that the grouped structure improves the prediction accuracy of several tree-based models across different construction sections and enhances the stability of incremental training. The best overall performance is achieved with the geology→device→grout order. Under a fixed 200-ring sample condition, higher R² and lower RMSE, MAE, and runtime are obtained with Bagging-A than with long short-term memory (LSTM) and Transformer. It is concluded that the method can provide a reference for real-time settlement prediction, construction parameter organization, and rapid model updating in soft-soil shield tunneling.