Abstract: Affected by seasonal rainfall fluctuations and the traction of adjacent points, the landslide displacement shows a step-like change trend. In order to predict the displacement effectively, the attention mechanism is introduced into bidirectional long-short term memory (Bi-LSTM) neural network prediction model in this paper. Firstly, a landslide monitoring cumulative displacement time series model is established to decompose the landslide cumulative displacement into a trend item and a period item. Then, the correlation coefficient among the landslide factor, the trend item and the period item are analyzed, and the polynomial regression is used to fit the trend item. For the period item prediction, the attention mechanism-based Bi-LSTM neural network prediction model is constructed. Taking a landslide data in Chongqing as an example, the experiment result shows that the proposed model has better robust generalization ability and can capture the correlation between different time series data effectively. The average absolute error of prediction accuracy is 0.088 mm, and the average mean square error is 0.042 mm. Compared with common long-short term memory neural network, the model proposed in this paper has higher prediction accuracy.

Key words: rainfall landslide, deformation prediction, time series decomposition, attention mechanism, bidirectional long-short term memory neural network