At present, it is difficult for traditional models and single models to fully capture the complexity and diversity of dam deformation data, resulting in limited predictive performance and interpretation ability. In order to solve the above problems, an efficient and interpretable dam deformation prediction method was proposed by combining and optimizing various prediction models. First, the Least Absolute Shrinkage and Selection Operation (LASSO) was used for efficient screening among environmental variables, both simplifying model input and explaining the reliability of factor selection. Then, Long Short-Term Memory (LSTM) network was employed to predict dam deformation, and Attention Mechanism was introduced to enhance the extraction of important information. Finally, Bagging (bootstrap aggregating) algorithm was used to integrate multiple model prediction results to further improve the accuracy, stability and generalization ability of the overall prediction. Taking a roller-compacted concrete gravity dam as an example, the model built has a high prediction accuracy, and the average MAE, MSE and RMSE at each measuring point are 0.042mm, 0.004mm and 0.053mm respectively. The comparative analysis with various commonly used models shows that the coupled model can capture the dynamic change of dam deformation more accurately, which provides a simple and efficient method for the study of prediction models.