During the daily operation and management of an earth-rock dam

data monitoring often suffers from difficulties caused by environmental data distortion and even data sequence gaps or interruptions. Owing to this challenge

traditional prediction methods have struggled to conduct a scientific and reliable analysis of the dam￠s deformation behaviors and health conditions. This paper presents a novel prediction method for earth-rock dam deformation

based on a single-time-series optimization model and the multi-scale combination theory. We consider the strong time-dependence of these deformation behaviors

and construct a Transformer single-time-series training model that features an excellent capability of global dependency learning to capture the autocorrelation of target variables accurately. And

variational mode decomposition is adopted to implement frequency-domain preprocessing of the training samples to reduce cross-interference from multi-frequency components and noise within the data. Further

we use the Kepler optimization algorithm to optimize the decomposition parameters and the historical information volume input for sub-sequence training

and thereby achieve a deformation prediction model for earth-rock dams under the condition of environmental data distortion. Case studies demonstrate this new method presents satisfactory prediction performance and strong generalization capability of handling non-stationary and low-quality monitoring data

showing a promising potential for practical deformation analysis of earth-rock dams under complicated monitoring conditions.