Abstract
The displacement evolution of concrete dams serves as a key indicator of their structural safety. Establishing an accurate and reliable model for predicting displacement is essential for effective dam monitoring. Nevertheless, current multi‐point forecasting approaches often overlook the interdependencies among deformation drivers and lack robust validation techniques to assess generalization capability and stability. This shortcoming hinders the accurate representation of deformation behavior under complex loading scenarios. To overcome these issues, this research introduces a spatiotemporal prediction model for concrete dams that combines hybrid clustering with adaptive decomposition and optimization strategies. Initially, the SOM‐K‐means method is employed to clusters monitoring points, followed by spatial correlation analysis to uncover interpoint relationships. Clustering performance is quantitatively evaluated using a composite assessment technique. During model development, hydrostatic pressures are derived through finite element simulation, and sensitivity analysis is applied to gauge the influence of environmental variables on deformation. Furthermore, the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and Mean Impact Value (MIV) techniques are employed to decompose and select deformation features. Tests show that the proposed model achieves superior predictive accuracy within clustered zones ( R 2 > 0.98, compared to Transformer: 0.9673 and CNN‐BiLSTM: 0.9501). Validation across multiple dam types further confirms the framework’s broad applicability and resilience. By incorporating spatiotemporal analysis, this method enables regionalized health monitoring and integrates data fusion under physical constraints, thereby significantly improving noise resistance and establishing a new benchmark for health prediction in high concrete dams.
Funding Information
  • National Natural Science Foundation of China (52369026)
  • National Natural Science Foundation of China (52069029)

This publication has 35 references indexed in Scilit: