Abstract: Previous zonal deformation prediction models lack the capability of capturing spatial differences in the trend, periodic and fluctuating components of dam deformation. This paper describes an improved zonal deformation prediction model to solve this problem. First, we adopt a variational mode decomposition algorithm to split dam displacements into trend, periodic and fluctuating components, and determine the representative environmental and load factors using hierarchical clustering of the principal components, so that these factors can be decomposed into the trend, low- and high-frequency components according to their physical meanings. Then, an optimized dynamic time warping algorithm based on a shape-based distance is used to divide the displacement components at the measured points into different deformation zones; for these zones, a sequence of their centroids is calculated to capture shared characteristics. The zonal data sets of the centroid sequences and their strongly related components of the dominant influencing factors can be established. Finally, we construct an improved zonal deformation prediction models using three machine learning algorithms－random forest, least squares support vector machine, and boosted regression tree－and an improved hydrostatic-thermal-time model. These improved models are verified against the measurements of Xiluodu super-high arch dam. The verification shows satisfactory results in accuracy and well explains the spatiotemporal correlation and differences in the trend, periodic and fluctuating components of dam displacements.

Key words: super-high arch dam, deformation, prediction, spatiotemporal correlation, variational mode decomposition, machine learning