Abstract: Landslide displacement is a key indicator for evaluating slope stability and implementing early warning measures. However, under the influence of cyclic reservoir water level fluctuations, displacement often exhibits step-like patterns, posing significant challenges for accurate modeling and prediction. To address this, we propose an interpretable machine learning framework for landslide displacement forecasting. The framework first employs an improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) to decompose displacement signals into high-frequency cycles and low-frequency trends, effectively mitigating mode mixing while preserving multi-scale features. Then, a Bidirectional Gated Recurrent Unit (BiGRU) model is used to predict each component, leveraging bidirectional context and a lightweight gating mechanism to capture both long-term dependencies and abrupt changes triggered by rainfall. Finally, SHapley Additive exPlanations (SHAP) are applied to interpret the model outputs, identifying key drivers such as historical and current reservoir levels, cumulative rainfall, and recent displacement trends, with site-specific differences across monitoring points. Case studies demonstrate that ICEEMDAN improves RMSE, MAE, MAPE, and R² by over 20% compared to traditional decomposition methods (EMD, EEMD, CEEMDAN). The BiGRU model achieves high prediction accuracy (e.g., R² = 0.992 and MAE = 3.617 mm at YY209), while SHAP enhances the transparency and physical interpretability of the predictions. Overall, the proposed framework combines high accuracy with interpretability, offering a promising approach for reservoir landslide early warning and risk management.