The monitoring data of landslide deformation are characterized by non-smooth, nonlinear and random changes, and the cumulative changes of the monitored objects have both monotonous growth trends and short-term fluctuations. The GM(1,1) model can get better results only when the data series are monotonous. Due to the limitations of the model, the prediction accuracy of the GM(1,1) model is limited to a certain extent. An improved algorithm based on the GM(1,1) model and the empirical mode decomposition (EMD-GM(1,1) model) for deformation prediction was presented to improve the forecast accuracy in this paper. Firstly, EMD was used to effectively separate the nonlinear high-frequency and low-frequency components hidden in the deformation sequence; then the moving average method was applied to build a prediction model for high-frequency component, and the GM(1,1) was applied to build the prediction model for low-frequency one according to the characteristics of each component; finally, the predicted value of each component was superimposed. The experimental results indicate that the optimized EMD-GM(1,1) model combines the advantages of the two models to separate effectively the different frequency components of the deformation sequence, which has higher prediction accuracy. Compared with the conventional GM(1,1) model, DGM(2,1) model and the Linear fitting model, the proposed model could satisfactorily describe the landslide deformation prediction practically.