Landslide displacement prediction based on time series and long short-term memory networks

Accurate landslide displacement prediction is an important component to realize landslide warning. For the displacement prediction of landslides, how to improve the prediction accuracy has been the focus on such research problems. Therefore, in this paper, the TS-SSA-LSTM landslide displacement prediction model is proposed as an example for landslides in Anxi County, Quanzhou City, Fujian Province. First, the cumulative displacement is decomposed into trend and period terms by time series. Then, the MIC coefficients are utilized for correlation test to screen the input terms of the prediction model. Then, the prediction models of trend and period terms are established by LSTM model, and SSA algorithm is introduced in this process for optimization. Finally, after obtaining the prediction results of the trend and period terms, the displacement data are reconstructed according to the basic theory of time series to obtain the final displacement prediction results. By comparing the RMSE, MAE and R\(^2\) of the TS-SSA-LSTM landslide displacement model with those of the traditional RNN and LSTM landslide displacement models, it is shown that the established landslide displacement prediction models have higher accuracy and stability. From the prediction results of TS-SSA-LSTM, the deviation from it and the real monitoring value is small and the fitting degree is high, which indicates that the TS-SSA-LSTM types proposed to this paper can effectively predict the displacement change of the landslide in the study case.