Abstract: To achieve the dynamic identification of landslide evolutionary states and high-precision displacement prediction, a landslide deformation prediction model based on Dynamic Clustering (DCPF) was proposed.This model consists of a Temporal Clustering Module (TCM) and a Channel Correlation Modeling (CCM) module.The TCM was utilized to adaptively extract the principal deformation patterns from landslide displacement sequences, while the CCM characterized the correlations among multiple monitoring variables.This integration enabled the unified modeling of both landslide deformation state identification and deformation prediction.Taking Baijiabao and Bazimen landslides in the Three Gorges Reservoir area as case studies, the validity and applicability of DCPF were evaluated, and the results were compared with simulations from SVR and LSTM models.The findings indicated that in the Baijiabao landslide case, the coefficient of determination for the DCPF model consistently remained above 0.98 across all monitoring points.Notably, at the large-deformation monitoring point ZD3, the Mean Absolute Error of its prediction was reduced by 22.8% compared to the LSTM model.Furthermore, the DCPF maintained extremely high robustness during the complex, long-term evolution of the Bazimen landslide.These research verified the advantages of DCPF model in handling the complex non-linear deformation and multivariate coupling characteristics of landslides, providing a high-precision analytical tool for the early warning of landslide disasters in reservoir areas.