An interpretable and high-precision method for predicting landslide displacement using evolutionary attention mechanism

Precise and reliable displacement prediction is essential for preventing landslide disasters, but the evolution of landslides is a dynamic process influenced by diverse factors at different stages. Despite advances in the application of machine learning models to landslide displacement prediction, these models struggle to dynamically capture triggers during the prediction process. This limitation not only fails to capture the characteristics of the short-term fast deformation area, thus affecting the overall prediction accuracy, but also fails to establish a connection between the data relationships and the physical mechanism, thereby limiting the understanding of the physical mechanism of the landslide and resulting in low reliability of the prediction results. In this study, we establish a new model for landslide displacement prediction that combines double exponential smoothing (DES), variational mode decomposition (VMD), and evolutionary attention-based long short-term memory (EA–LSTM). The prediction process is as follows: (i) VMD is used to extract trend, periodic, and random displacement from cumulative displacement; (ii) DES is utilized for forecasting trend displacement, and periodic and random displacements are predicted by EA–LSTM; and (iii) these individual predictions are combined to produce the total displacement prediction. The proposed model is validated using monitoring data collected from the Baishuihe and Bazimen landslides in the Three Gorges Reservoir area. The results indicate that, compared with other models, the proposed model demonstrates higher predictive accuracy. In addition, the real-time dynamic weights of historical information revealed by the model on different time stamps are consistent with the actual historical evolution of landslides. These results verify that the proposed model is a promising tool for the high-quality prediction of landslides and can inform landslide treatment-related decision-making.