Objectives The dynamic coupling of rainfall and reservoir water level triggers step-like deformation and creep-mutation transitions in Three Gorges Reservoir landslides. Existing prediction methods can't adequately capture the lagged attenuation prediction method for reservoir bank landslides.

Methods To address these issues, we propose a novel dynamic displacement prediction method for reservoir bank land-slides. The proposed method integrates a new lagged-attenuation-based dominant factor screening strategy with a combined model utilizing grey wolf optimizer (GWO), bidirectional temporal convolutional network (BITCN), bidirectional gated recurrent unit (BIGRU), and attention mechanism. The model aims to simultaneously capture time-series features and high-dimensional dynamic spatial features. Using the Xinpu and Baishuihe landslides in the Three Gorges Reservoir area as case studies at different temporal scales, we first decompose cumulative landslide displacement into trend and periodic components, using complete ensemble empirical mode decomposition with adaptive noise. Trend displacement is predicted by the double exponential smoothing method. Subsequently, we conduct an in-depth analysis of the dynamic response of landslide deformation to rainfall and reservoir level variations, explicitly considering their lagged attenuation effects on displacement evolution. Building on this, we employ a comprehensive approach combining recursive feature elimination with cross validation-extreme gradient boosting, classification and regression tree, and the maximal information coefficient to screen for the dominant factors controlling displacement evolution, identifying the antecedent precipitation index as the core factor. Finally, the GWO-BITCN-BIGRU-Attention model is applied to predict the periodic displacement component.

Results Case studies on the Xinpu and Baishuihe landslides demonstrate that the proposed method outperforms existing approaches. It achieves higher correlation coefficients and lower root mean square error.

Conclusions Model evaluations and ablation experiments confirm that the method effectively characterizes the time-varying nature of lagged attenuation effects. It also captures long-term evolutionary dynamics across different time scales. This leads to a substantial improvement in displacement prediction accuracy.