Multi-target prediction and dynamic interpretation method for displacement of arch dam with cracks based on A-DSRSN and SHAP

To address the challenges of accuracy and interpretability in displacement prediction for arch dams with cracks, this study proposes a deep stacked residual shrinkage network based on attention mechanisms (A-DSRSN), combined with the multi-target prediction method leveraging maximum correlated stacking of single-target (MCSST) and the interpretive method of SHapley Additive exPlanations (SHAP), constructing a multi-target prediction and dynamic interpretation method for displacement of arch dams with cracks. Initially, factors influencing displacement are preliminarily selected using hydrostatic-air temperature–time (HT_AT) and hydrostatic-temperature-crack-time (HTCT) models. Elastic net (ElaNet) and principal component analysis (PCA) are subsequently employed for feature selection and extraction. A-DSRSN is then constructed by integrating convolutional block attention modules and residual shrinkage blocks. Using the A-DSRSN model, a multi-target displacement prediction method based on MCSST is established. Furthermore, the integration of the A-DSRSN model with SHAP analysis develops global, local, and dynamic interpretation methods for prediction results. Finally, a case study demonstrates the validity of the proposed method by comparing it with existing baseline approaches. The findings reveal that the A-DSRSN model outperforms baseline methods in prediction accuracy, while the MCSST method enhances overall predictive precision. The interpretation method in this paper reveals the process of model prediction, accurately identifying dominant influencing factors during water level rise and periods of water level and temperature decline. This research provides a novel method for dam displacement prediction, which has significant practical application value for long-term health monitoring and safety management of dams.