基于ACNN-VAR的特高拱坝变形深度学习预测模型  

Deformation Prediction Model of Super High Arch Dam Based on ACNN-VAR Deep Learning

作　　者：钟桥俊 赵二峰[1,2,3] 胡灵芝 刘峰 宋桂华 ZHONG Qiao-jun;ZHAO Er-feng;HU Ling-zhi;LIU Feng;SONG Gui-hua(The National Key Laboratory of Water Disaster Prevention,Hohai University,Nanjing 210098,China;National Engineering Research Center of Water Resources Efficient Utilization and Engineering Safety,Hohai University,Nanjing 210098,China;Yunnan Key Laboratory of Water Resources and Hydropower Engineering Safety,Kunming 650051,China;PowerChina Kunming Engineering Corporation Limited,Kunming 650051,China;Shanghai Investigation,Design and Research Institute Co.,Ltd.,Shanghai 200335,China)

机构地区：[1]河海大学水灾害防御全国重点实验室,江苏南京210098 [2]河海大学水资源高效利用与工程安全国家工程研究中心,江苏南京210098 [3]云南省水利水电工程安全重点实验室,云南昆明650051 [4]中国电建集团昆明勘测设计研究院有限公司,云南昆明650051 [5]上海勘测设计研究院有限公司,上海200335

出　　处：《水电能源科学》2025年第4期163-167,共5页Water Resources and Power

基　　金：国家自然科学基金项目(52079046);云南省水利水电工程安全重点实验室开放课题基金(202302AN360003)。

摘　　要：变形是大坝服役性态的直观表现,对其进行分析与预测是科学诊断大坝健康的关键措施。为解析特高拱坝变形性态,利用基于注意力机制优化的卷积神经网络(ACNN)对实测数据进行局部时间特征提取,引入变分自编码器(VAE),将回归器嵌入到VAE构建变分自回归器(VAR),提出基于ACNN-VAR的特高拱坝变形深度学习预测模型,该模型综合注意力机制、CNN神经网络、VAE生成模型和回归器,深度挖掘特高拱坝变形性态的特征信息,生成潜在特征向量,实现深层次变化特征提取。实例应用结果表明,建立的预测模型能够准确模拟实测值的年周期变化及局部波动,具有较高的预测精度和稳定的鲁棒性,为特高拱坝变形监测提供了新思路。Deformation is the intuitive manifestation of the service behavior of dams.The analyzing and predicting deformation is a key measure for scientifically diagnosing the health of dams.In order to analyze the deformation behavior of ultra-high arch dams,this paper utilizes an Attention Convolutional Neural Network(ACNN)optimized based on attention mechanism to extract local temporal features from measured data.Then,the Variational Auto-encoder(VAE)is introduced,embedding the regressor into VAE to construct the Variational Auto Regressor(VAR),proposing an ACNNVAR based deep learning prediction model for deformation of ultra-high arch dams.This model integrates attention mechanism,CNN neural network,VAE generative model,and regressor,deeply mining the characteristic information of ultra-high arch dam deformation behavior,generating latent feature vectors,and achieving deep-level change feature extraction.Application examples demonstrate that the established prediction model can accurately simulate the annual cyclic variation and local fluctuations of measured values,with high prediction accuracy and stable robustness,providing new ideas for the deformation monitoring of super high arch dams.

关 键 词：变分自编码器 注意力机制 深度学习 潜在特征向量 

分 类 号：TV698.11[水利工程—水利水电工程]