Abstract: In order to solve the problem of feature factor redundancy and insufficient capture of periodic laws in dam deformation prediction, a dam deformation prediction model that combined features screening and multi-scale features enhancement was established in this paper. Firstly, the maximum information coefficient (MIC) was used to screen out the environmental factors highly related to the dam deformation, which can effectively remove redundant variables and simplify the model input. Secondly, the complete ensemble empirical mode decomposition adaptive noise (CEEMDAN) was used to adaptively decompose the deformation data, which effectively reduced the influence of nonlinearity and non-stationarity, and extracted the intrinsic mode function with clear physical meaning. Finally, a frequency-time enhanced attention block was proposed and embedded into the Transformer model. The frequency domain information was captured by discrete cosine transform (DCT) to realize multi-scale feature extraction and enhancement of data. Engineering experiments were carried out based on the deformation monitoring data of Shangyoujiang Dam in Jiangxi Province. The results showed that the model constructed in this paper can achieve excellent prediction results. The model′s R² reached 0.999 1, RMSE was 0.041 3 mm, and MAE was 0.031 8 mm. Compared with Transformer, LSTM, GRU and TCN models, R² was improved by 0.015 9, 0.019 2, 0.018 0 and 0.016 9, respectively. Especially at the peak and fluctuation node positions, the model constructed in this paper showed higher accuracy and stability. In addition, in the deformation prediction experiments of different monitoring points, this model still maintained high prediction accuracy, which verified its effectiveness and practical application value in the field of dam safety monitoring.