To overcome the drawbacks of previous displacement prediction models for step-like landslides, such as poor performance in predicting mutational displacement and unclear reliability of prediction results, this paper proposes a new hybrid method of landslide displacement prediction intervals. Firstly, the combination of SOM network and K-means clustering is implemented to divide the deformation states of step-like landslides into steady state and mutational state. Secondly, on the basis of expanding the mutational state samples through the comprehensive application of the engineering geology analogy method and the adaptive synthetic sampling algorithm, the random forest algorithm is used to establish an ensemble classifier for recognizing the landslide deformation states automatically. Finally, based on the Bootstrap-KELM-BPNN model, an interval prediction framework considering the dynamic switching of landslide deformation states is constructed to realize the dynamic prediction of landslide displacement. Taking Baishuihe landslide, a typical step-like landslide in the Three Gorges Reservoir Area, as an example, the dataset of XD01 monitoring point from June 2006 to December 2016 are explored to verify the effectiveness, accuracy and reliability of the proposed method.