Deformation represents the most direct indicator of an arch dam's operational state. However, existing multi-point deformation prediction models struggle to accurately characterize the time-varying interdependencies among monitoring points and remain limited in both spatio-temporal feature extraction and model interpretability. To address these gaps, this study develops a refined dynamic relational framework from two complementary perspectives: the heterogeneous environmental-load–driven responses of individual monitoring points, and the nonlinear coupling among their deformation behaviors. Building on this foundation, we formulate a Joint Spatio-Temporal Diffusion Block (JST-Block) to synchronously capture multi-scale spatio-temporal correlations. Furthermore, an integrated interpretability module is introduced to provide quantitative attributions from three hierarchical levels: input features, edge-level dependencies, and global dynamic graph structures, thereby elucidating the intrinsic mechanisms linking environmental loads, structural responses, and deformation prediction. • A physical-informed dynamic graph framework integrates causal responses and dynamic coupling. • A Joint Spatiotemporal Diffusion Block (JST-Block) captures multi-order spatial diffusion and temporal evolution within a unified layer. • Multi-level interpretability quantifies feature contributions, edge sensitivity, and graph-topology evolution. • The model outperforms traditional ML, DL, and GNN baselines in multi-point displacement prediction. • A distressed high arch dam case shows strong generalization and physically consistent interpretations.
Liu et al. (Fri,) studied this question.