Abstract

Abstract

Under the influence of internal geological conditions and external factors such as cycle and random, the landslide evolution process has typical leaping characteristics. The traditional deep learning model based on the gating mechanism has insufficient prediction ability for the step-type landslide, and the multi-head self-attention can adaptively mine the change degree characteristics of the sequence by focusing on the implicit information of the time-series data at different scales, effectively learn the potential change trend of the data, and improve the prediction ability of the sequence. Based on the impoved variational modal decomposition technique, the cumulative landslide displacements are decomposed into trend, periodic and random terms, and the dynamic time-regularized correlation analysis is carried out for each displacement component and influence factor. The dynamic prediction of each displacement component is carried out by combining the multiple self-attention mechanism and the long and short-term memory network model, and the predicted values of each displacement component are summed up to get the actual prediction results. Taking the Bishui River landslide in the Three Gorges reservoir area as the study area, the cumulative displacement prediction is carried out for the monitoring point ZG118, and the model adaptability is verified by using the monitoring points. The experimental results show that the new model can greatly improve the prediction accuracy for the step data segments caused by the changes of the rainfall and the reservoir water level, and provide a new way of thinking for the study of landslide displacement prediction in the Three Gorges reservoir area.