Integrating Time Series Decomposition and Deep Learning: A SOO-VMD-CNN-TimeXer Framework for Landslide Cumulative Displacement Prediction in Alpine Regions

The cumulative displacement of landslides in alpine regions is jointly affected by rainfall, temperature variation, freeze–thaw cycles, and other factors, and usually exhibits nonlinear, non-stationary, and multi-scale fluctuation characteristics. To improve the accuracy of landslide displacement prediction under complex environmental conditions, this study takes the Taker Tubek Village landslide in Gongliu County, Xinjiang, China, as the study object. Cumulative displacement data from GNSS02 and GNSS03, together with daily rainfall and daily mean temperature, were used to construct a SOO-VMD-CNN-TimeXer hybrid prediction model. First, SOO was employed to adaptively optimize the VMD parameters, and the cumulative displacement series were decomposed into multiple IMF components. Then, CNN was used to extract local fluctuation features, while TimeXer was applied to model long-term temporal dependencies and the effects of exogenous variables. Finally, the predicted results of all components were reconstructed to obtain the cumulative displacement prediction. The results show that the proposed model achieved high prediction accuracy at both GNSS02 and GNSS03. The MSE, MAE, MAPE, and R2 values were 0.0158, 0.0960, 0.0112, and 0.9464 for GNSS02, and 0.0483, 0.1590, 0.0203, and 0.9946 for GNSS03, respectively, outperforming LSTM, Informer, iTransformer, Crossformer, and other models. The results indicate that the SOO-VMD-CNN-TimeXer model can effectively characterize the cumulative displacement evolution of landslides in alpine regions and provide technical support for landslide deformation trend forecasting and disaster early warning.