关键词: 大坝变形监控模型, 特征筛选, 变分模态分解, 长短期记忆神经网络, 最小冗余最大相关性

Abstract: Accurate deformation prediction is of great significance to safe operation and long-term maintenance of dams, but previous methods have low prediction accuracy and lack sufficient information extraction from monitoring data. This paper constructs a relationship of dam deformation components versus their influencing factors through variational mode decomposition on the deformation series, and constructs Long Short-Term Memory neural networks with different structural parameters. Then, we develop a dam deformation analysis model that can realize optimal modeling through integrating the Grey Wolf Optimizer algorithm, the Minimum Redundancy Maximum Relevance method, and other strategies to improve its accuracy from three aspects－front-end decomposition, information extraction, and time series prediction. A case study shows that compared with the conventional monitoring model, this new model is more accurate in the simulations of dam deformation time variations and better in generalization performance, thus useful for dam deformation safety analysis.

Key words: dam deformation monitoring model, feature screening, variational mode decomposition, Long Short-Term Memory neural network, Minimum-Redundancy Maximum-Relevance