Abstract: Under the influences of the geological conditions of landslide and external periodic and random factors, the evolution process of the landslide has typical abrupt characteristics. The conventional deep learning model based on gating mechanism has insufficient ability to predict step mutation sequence, while by mining the potential information of time series data with different scales, multi-head self-attention can adaptively find the alteration trend of sequences and improve the prediction ability of landslide sequence. Based on variational mode decomposition, the cumulative displacement of landslide was decomposed into trend displacement, periodic displacement and stochastic displacement. Dynamic time warping algorithm was used to analyze the correlation between each component and influencing factors. These displacement components were dynamically predicted by a modified model integrated with long short-term memory (LSTM) neural network and multi-head self-attention mechanism, and the predicted values of each component were added to obtain the cumulative displacement prediction result. The monitoring point ZG118 of Baishuihe landslide in the Three Gorges Reservoir area was taken as an example to predict the cumulative displacement in this paper. The monitoring points ZG93 and XD01 were used for model adaptability verification. The results show that the new model can greatly improve the prediction accuracy for the step data mutation caused by the changes of rainfall and reservoir water level, providing a new idea for the prediction of landslide displacement in the Three Gorges Reservoir area.

Key words: landslide displacement prediction, variational mode decomposition, dynamic time warping, multi-head self-attention mechanism, long short-term memory (LSTM) neural network