The safety and health monitoring of dams has attracted increasing attention. In this paper, a novel prediction model based on variational autoencoder (VAE) and temporal attention-based long short-term memory (TALSTM) network is proposed for the long-term deformation of arch dams. In the proposed model, the convolutional neural network-based VAE is applied to extracting the features of environmental data. The TALSTM is employed to construct the relationship between the dam displacement and extracted features. For verification, an arch dam is taken as an example. Through the comparison among nine baseline prediction models, the proposed model is more stable and effective than other prediction models. Furthermore, the proposed model could capture the features of environmental data accurately and provide better prediction results. Therefore, the proposed model is more suitable for engineering applications.

大坝的安全和健康监控已引起越来越多的关注。针对拱坝的长期变形，提出了一种基于变分自动编码器（VAE）和基于时间关注的长短期记忆（TALSTM）网络的新型预测模型。在提出的模型中，基于卷积神经网络的VAE被用于提取环境数据的特征。TALSTM用于构造大坝位移与提取特征之间的关系。为了验证，以拱坝为例。通过对9个基线预测模型的比较，提出的模型比其他预测模型更稳定，更有效。此外，所提出的模型可以准确地捕获环境数据的特征并提供更好的预测结果。所以，