Abstract:The uncertainty of current landslide geological model, prediction model and real-time monitoring data is common in landslide system and known by the researchers in the geotechnical field, causing the difficulties to efficiently describe transient characteristics of the landslide evolution. However, the landslide displacement prediction as one of the effective methods for landslide forecasting becomes the important foundation of the landslide risk assessment and corresponding reinforcements. To overcome the deficiency of static predicting models for landslide, a novel forecasting method named evolutionary Support Vector Machines (SVM) combined Simplex-Niche-Genetic algorithm is proposed and conducted in this paper, avoiding the artificial control of the kernel function and parameters compared with traditional SVM model. And also, the Simplex-Niche-Genetic algorithm was applied to deal with the local and global optimization, showing the model generalization and good convergence compared with traditional BP network model. In order to improve the prediction reliability of the dynamic landslide displacement model, deep investigation was extended to the dynamic displacement prediction of reservoir landslide, and interval prediction method was introduced and the cumulative displacement was decomposed into trend terms and periodic terms, and main controlled factors including six components caused by rain infiltration and reservoir water variation and landslide displacements at two different time periods are determined to form the dynamic displacement prediction model, and then realize the real-time displacement prediction of the landslide cumulative displacement. Finally, take a reservoir landside as a case study, the displacement prediction covering the landslide displacement curve completely indicates better precision than other prediction model. Therefore, the proposed evolutionary SVM model considering displacement time-series can explain the uncertainty of the landslide displacement caused by dynamic variation of controlled triggering factors, and also obtain the real-time landslide displacement quickly and correctly, which can offer a new idea for the forecasting and the early warning of reservoir stepped landslides.