A high-dimensional abstraction Transformer model for high dam deformation prediction under the interpretability perspective

Abstract

In extreme environments and complex geological conditions, dam deformation is a critical indicator of structural integrity and safety, essential for accurate prediction and early warning. With the rapid growth of dam safety monitoring data, traditional analytical methods face increasing challenges in handling high-dimensional, nonlinear, and time-dependent information. Although data-driven deep learning models have been widely applied to dam deformation prediction, many existing approaches primarily often insufficiently represent deeper time and factor dependencies of environmental variables. In addition, the limited interpretability of most “black-box” models restricts their practical application in engineering decision-making. To address these challenges, this study proposes a dam deformation prediction method that combines multidimensional mechanism fusion with deep learning, quantifying the contribution of environmental factors to dam deformation and providing a mechanistic explanation of the influencing factors. The method introduces time dimensions into the encoder–decoder layers of the Transformer model; on the basis of extracting inter features of water level, temperature, and aging in the hydraulic–seasonal–time model through an attention mechanism, the temporal dependence and correlation of these factors is simultaneously captured and fused, and controls high-dimensional global dependencies through gating structures, thus proposing the time abstraction Transformer (TA-Transformer) for dam safety monitoring. Furthermore, the Shapley additive explanations method is coupled with the proposed time-series prediction framework to quantify the relative contributions of key environmental factors to long-term dam deformation. This interpretability analysis provides insights into the nonlinear relationships between temperature, water pressure, aging effects, and structural response, thereby enhancing the physical understanding of model predictions. Finally, the proposed method is validated using monitoring data from the Dagangshan high-arch dam. The results indicate that the TA-Transformer is capable of capturing both pronounced deformation variations and subtle long-term trends, achieving improved predictive performance compared with several baseline models and exhibiting stable generalization across different scenarios. And the high-precision prediction by the proposed model is dominated by temperature components and assisted by water pressure components for nonlinear changes of dam deformation.