Abstract: Taking the Bazimen landslide in the Three Gorges reservoir area as an example, a new landslide displacement prediction method is proposed to solve the problems of the static machine learning model in the periodic displacement prediction and the difficulty in the high-frequency random displacement prediction. Based on the idea of time series decomposition, particle swarm optimization (PSO) is used to optimize the parameters of variational mode decomposition (VMD), and the displacement time series is decomposed into trend term, periodic term and random term. The trend term is mainly affected by the internal factors of landslide, and the Fourier curve is used to fit and predict. The periodic term is caused by external factors. The causes are analyzed based on Granger causality test, and a nonlinear autoregressive neural network with exogenous inputs (NARX) with high sensitivity to the historical state of time series is introduced for prediction. The random term frequency is high and the influencing factors cannot be determined, thus one-dimensional gated recurrent unit (GRU) is used for prediction. Finally, the predicted displacement of each component is superimposed and reconstructed to realize the prediction of landslide cumulative displacement. The results show that the (PSO-VMD)-NARX-GRU landslide displacement dynamic prediction model has higher accuracy, and the prediction accuracy of each displacement component is obviously higher than that of BP neural network, support vector machine (SVM) and conventional autoregressive model ARIMA in static models, which can provide reference for step landslide displacement prediction.

Key words: landslide displacement prediction, particle swarm optimization, variational mode decomposition, Granger causality test, nonlinear autoregressive neural network with exogenous inputs, gated recurrent unit