In order to improve the accuracy of dam deformation prediction, a concrete dam deformation prediction model with optimized variational mode decomposition (VMD) and gated recurrent unit (GRU) was proposed based on the idea of decomposition-reconstruction, in which the deformation signal processing technology was used to perform time-frequency decomposition on the measured deformation and the deep learning networks was combined to predict and reconstruct the decomposed signals. Grey wolf optimization (GWO) optimized VMD was used to decompose the raw data into a set of optimal intrinsic mode components (IMF), and GWO optimized GRU network was used to perform rolling prediction on each IMF component. By overlaying the prediction results of each component, displacement sequence prediction results were obtained, solving the problems of poor decomposition effect caused by VMD manual parameter selection and the impact of GRU manual parameter selection on training speed, usage effect, and robustness. The prediction results of engineering examples show that the model has low prediction error and good prediction accuracy and robustness.