The displacement and deformation of landslides are affected by many factors, which is a typical non-linear change process. The neural network can deal with linear problems well. However, the traditional serialization network will cause gradient disappearance and sequence forgetting due to sequential calculation. This paper proposes a combined prediction model that combines singular spectrum analysis (SSA) and improved attention mechanism. First, SSA is used to decompose the displacement time series and eliminate noise, reduce random fluctuation factors and increase the amount of effective information, and then use the improved attention mechanism to predict the decomposition sub-sequence, and superimpose the results to obtain the final prediction value. Experiments show that the combined model has better effects and higher accuracy than traditional sequence models, such as RNN and LSTM.