Abstract: As the deformation monitoring data of concrete dam has evident non-linear and non-stationary characteristics, and the data sequence contains a certain amount of noise, it is easy to lead to low accuracy predicted by the model. Aiming at the above issues, the deformation prediction model established for concrete dam, which is referred to as the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) - permutation entropy (PE) - long short-term memory (LSTM), is proposed. Using the capability of CEEMDAN to adaptively decompose non-linear signals, this model decomposes original deformation data into a set of intrinsic mode functions (IMF) with different frequencies and obvious differences in complexity, which reduces the mutual influence of distinct scale information in the sequence. Then, the IMF quantities with similar complexity are combined and reorganized. Finally, LSTM models are respectively constructed for several re-organized sequences to perform prediction, and then the predicted results are added to obtain the final prediction. This study is based on modeling and analyzing the horizontal displacement monitoring data of a concrete dam. The results demonstrate that the CEEMDAN-PE-LSTM model shows significant improvement in accuracy compared to the conventional model, and is able to better predict nonlinear data sequence. As compared to the LSTM model, the mean average error (MAE), mean average percentage error (MAPE) and relative mean square error (RMSE) are reduced by 76.43%, 75.55% and 74.73%, respectively, suggesting that the proposed model can better exploit the variability of non-linear and non-stationary data by decomposing and reorganizing the original sequence to obtain distinct scale features with improved prediction accuracy, and can be effectively applied to the deformation prediction of concrete dams.