To improve the prediction accuracy on the deformation of dam structure, the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm was adopted in this work to decouple the measured sequence into a series of intrinsic mode functions with different time-frequency characteristics, and then the wavelet threshold denoising was used to stabilize the high frequency component for the reconstruction. Afterwards, a dual-stage attention-based long short-term memory network (DA-LSTM) model was introduced to predict the reconstructed deformation sequence. The results show that the denoising processing method combining CEEMDAN algorithm and wavelet threshold denoising can effectively identify and deal with the outliers in the measured data to improve the representation capacity on the dam performance. Moreover, the established model can exploit the hysteresis of dam deformation and enhance the interpretability of the DA-LSTM network, indicating the strong robustness.