Abstract: Accurate and reliable prediction of dam deformation is of paramount importance for ensuring structural safety, operational stability, and long-term performance monitoring of hydraulic infrastructure. However, deformation monitoring sequences often exhibit complex nonlinear characteristics due to the combined influence of water pressure, temperature, time-dependent effects, and environmental factors, and are frequently contaminated by measurement noise and stochastic disturbances. Traditional prediction models, including statistical regression methods and conventional machine learning algorithms, often struggle to fully capture these intricate patterns and have difficulty mining and eliminating noise information using first mode decomposition. To overcome these limitations, this study proposes a prediction model of dam deformation based on a bi-directional gated recurrent unit network. The proposed methodology introduces a comprehensive two-stage decomposition-reconstruction approach to address the noise contamination problem in dam deformation sequences. In the first stage, the original deformation signal is decomposed using symplectic geometry modal decomposition, a novel signal processing technique that overcomes empirical parameter selection issues inherent in traditional methods and effectively separates the signal into a series of symplectic geometry components and a residual component. Recognizing that the residual may still contain valuable information, the second stage employs improved complete ensemble empirical mode decomposition with adaptive noise for further decomposition. This two-level decomposition strategy ensures more thorough noise separation while preserving critical deformation features across different frequency bands. A correlation test between the modal components and the measured sequence is then conducted using the maximum information coefficient (MIC), a robust statistical measure capable of detecting both linear and nonlinear dependencies between variables. Components exhibiting weak correlations, which typically represent noise-dominated signals, are subjected to wavelet threshold denoising. The db4 wavelet base with soft thresholding is employed for this purpose, effectively suppressing random noise while minimizing signal distortion. The selected and processed components are subsequently reconstructed to form an enhanced deformation sequence with a significantly improved signal-to-noise ratio. For the prediction task, a sophisticated bi-directional gated recurrent unit with attention (BiGRU-Attention) network is adopted, combining the strengths of bi-directional sequence processing with adaptive feature weighting. The BiGRU architecture processes the input data in both forward and backward temporal directions, capturing comprehensive temporal dependencies in the deformation-environment relationship. The attention mechanism dynamically assigns importance weights to different time steps and features, enabling the model to focus on the most relevant information for accurate prediction. Key environmental factors, including upstream reservoir water level, cumulative monitoring days, and temperature effects, are incorporated as model inputs to reflect the comprehensive loading conditions on the dam structure. The model's performance was rigorously evaluated using extensive monitoring data from a large double-curvature arch dam in China. The dataset comprised daily measurements from the monitoring point, covering a four-year period. Comparative experiments demonstrated the superiority of the proposed approach: compared with the primary decomposition method, the noise reduction effect of the secondary decomposition method is significantly improved, and the signal-to-noise ratio is enhanced. In addition, compared with other deep learning models, BiGRU-Attention shows better performance and higher prediction accuracy. Moreover, the deformation value predicted by BiGRU-Attention is closer to the measured value, with smaller residuals. Additional validation on the symmetrical monitoring point confirmed the model's strong generalization capability, with prediction accuracy metrics consistently outperforming conventional approaches. The practical implications of this research are significant for dam safety monitoring systems: secondary mode decomposition enables fine decomposition of measured dam deformation sequences, effectively eliminates sequence noise, and enhances the prediction accuracy of the new model. Furthermore, BiGRU-Attention fully explores the nonlinear relationship between dam deformation and environmental quantities, significantly improves prediction performance, and, through the attention mechanism, assigns different weights to input information to highlight key features. This further improves dam deformation prediction accuracy and provides a new method for forecasting the long-term service performance of dams.