In order to reduce the influence of non-stationarity of dam deformation data on the prediction model and improve the prediction accuracy，the dam deformation monitoring data were first decomposed into several subsequences using variational mode decomposition （VMD），and then the spatial features of the deformation data were extracted and predicted by using convolutional neural networks （CNN），which were combined with bi-directional long short-term memory （BiLSTM） for time domain modeling and prediction，and the hyper-parameters of the model were optimized by introducing dung beetle optimizer （DBO），and finally the predicted values of each subsequence were superimposed and reconstructed to obtain the final dam deformation prediction. The results show that compared with BiLSTM，VMD-BiLSTM and VMD-CNN-BiLSTM，the combined VMD-DBO-CNN-BiLSTM model is optimal in all evaluation indexes，and shows higher model accuracy and stability in dam deformation prediction.