Constructing an accurate and dependable displacement forecasting model is a prerequisite for realizing effective early warning systems of landslide disasters. To overcome the drawbacks of previous displacement prediction models for landslides with step-like deformation characteristics, such as the low prediction accuracy of the mutational displacements and the unclear reliability of the prediction results, we propose a novel hybrid interval forecasting model. This model consists of four parts. First, clustering by fast search and find of density peaks is implemented to distinguish the deformation states of the landslide. Second, the ensemble classifier based on the random forest algorithm is established to identify the deformation states. Third, based on the wild bootstrap, kernel extreme learning machine, and back propagation neural network approaches, the ensemble regressors under different deformation states are built. Finally, by combining the ensemble classifier and ensemble regressors, an interval prediction framework is constructed to realize the dynamic interval prediction of landslide displacement. Taking the Baishuihe landslide as an example, the datasets of three monitoring sites from June 2006 to December 2016 are used to verify the accuracy and reliability of the proposed model. The results show that the proposed model can effectively improve the prediction accuracy of mutational displacements, with the root mean square errors of 28.19 mm, 14.21 mm, and 34.44 mm and the R-squares of 0.9827, 0.9955, and 0.9903, respectively. Moreover, the reliability of the prediction results obtained using this model can be expressed in the flexible prediction intervals (PIs) under different deformation states. The coverage width-based criteria of PIs at 90% nominal confidence are 140.38 mm, 86.61 mm, and 173.68 mm, respectively. In conclusion, the proposed model provides a good basis for developing early warning systems for landslides with step-like deformation characteristics.

构建准确，可靠的位移预测模型是实现有效的滑坡灾害预警系统的前提。为了克服以往具有阶梯状变形特征的滑坡位移预测模型的缺陷，如突变位移的预测精度低，预测结果的可靠性不明确等，我们提出了一种新型的混合区间预测模型。该模型包括四个部分。首先，通过快速搜索和发现密度峰值进行聚类，以区分滑坡的变形状态。其次，建立基于随机森林算法的集成分类器，以识别变形状态。第三，基于狂野的引导程序，内核极限学习机，运用反向传播神经网络方法，建立了不同变形状态下的集合回归。最后，结合集合分类器和集合回归器，构造了一个区间预测框架，实现了滑坡位移的动态区间预测。以白水河滑坡为例，以2006年6月至2016年12月三个监测点的数据集为例，验证了该模型的准确性和可靠性。结果表明，提出的模型可以有效提高突变位移的预测精度，均方根误差分别为28.19 mm，14.21 mm和34.44 mm，R平方分别为0.9827、0.9955和0.9903。此外，使用该模型获得的预测结果的可靠性可以表示为不同变形状态下的弹性预测区间（PI）。在90％的标称置信度下，PI的基于覆盖宽度的标准分别为140.38 mm，86.61 mm和173.68 mm。总之，所提出的模型为开发具有阶梯状变形特征的滑坡预警系统提供了良好的基础。