The traditional memory network model has poor prediction accuracy on a non-stationary mutation data segment due to the gate mechanism has difficulty in learning sequence change trends. A multifractal algorithm was introduced to improve the gating control of the gated recurrent unit (GRU), and dynamically update the gating weight by quantifying the change characteristics of the sequence. The state fusion of the recurrent neural network unit was introduced to learn the long-range correlation of data. The landslide cumulative displacement was decomposed into trend term, periodic term and random term through the variational mode decomposition algorithm. The decomposed terms were trained and predicted by using the improved GRU. The monitoring points ZG93 and ZG118 of the Baishuihe landslide in the Three Gorges Reservoir area were selected for simulation experiments. Experimental results show that compared with traditional models, the new model learns the characteristics of the landslide displacement deformation trend better and has a higher prediction accuracy.