基于TimeGAN和CNN-BiLSTM-Attention的大坝变形预测混合模型  

Hybrid Model for Dam Deformation Prediction Based on TimeGAN and CNN⁃BiLSTM⁃Attention

作　　者：原佳帆 李丹杨 李佳霖 秦学 毛鹏 YUAN Jiafan;LI Danyang;LI Jialin;QIN Xue;MAO Peng(College of Big Data and Information Engineering,Guizhou University,Guiyang 550000,China;Guiyang Engineering Corporation Limited,Power Construction Corporation of China,Guiyang 550000,China)

机构地区：[1]贵州大学大数据与信息工程学院,贵州贵阳550000 [2]中国电建集团贵阳勘测设计研究院有限公司,贵州贵阳550000

出　　处：《人民黄河》2024年第12期127-130,143,共5页Yellow River

基　　金：贵州省科技计划项目(黔科合支撑[2023]一般251);贵州省基础研究计划(自然科学)青年引导项目(黔科合基础[2024]青年095)。

摘　　要：基于历史数据的深度学习模型往往需要跨越数年的大量数据集,为了解决数据不足问题,提出一种将时间序列生成对抗性网络(TimeGAN)与CNN-BiLSTM-Attention相结合的混凝土面板堆石坝变形预测混合模型。首先,利用TimeGAN生成虚拟数据来扩展稀疏的数据集;然后,利用卷积神经网络(CNN)提取大坝传感器数据中的非线性局部特征,运用BiLSTM捕获双向时间序列特征;最后,引入注意力(Attention)机制对BiLSTM层提取的信息特征自动进行权重分配,通过全连接层输出最终预测结果。以贵州省毕节市某混凝土面板堆石坝为例,验证该混合模型的适用性。建立长短期记忆网络(LSTM)、CNN-LSTM、CNN-LSTM-Attention、CNN-BiLSTM-Attention 4种基模型,再分别引入TimeGAN,对比各模型的预测精度。结果表明:基于TimeGAN和CNN-BiLSTM-Attention的混合模型的拟合效果明显优于其他模型,其预测值与监测值最接近。相较于传统单一LSTM模型,混合模型的EMS、ERMS、EMA分别降低了71%、49%、45%,R2提升了20%。Deep learning models based on historical data often require a large dataset spanning several years.In order to address the issue of insufficient data,a hybrid model for predicting the deformation of concrete face rockfill dams was proposed,which combined Time Series Generative Adversarial Networks(TimeGAN)with CNN⁃BiLSTM⁃Attention.Firstly,it used TimeGAN to generate virtual data to expand the sparse dataset.Then,convolutional neural networks(CNN)were used to extract nonlinear local features from dam sensor data,and BiLSTM was used to capture bidirectional time series features.Finally,the attention mechanism was introduced to automatically fit the weight alloca⁃tion of information features extracted by the BiLSTM layer,and the final prediction result was output through the fully connected layer.Taking a concrete face slab dam in Bijie City of Guizhou Province as an example,it verified the applicability of the hybrid model in practical engi⁃neering.It established four basic models of Long Short Term Memory Network(LSTM),CNN⁃LSTM,CNN⁃LSTM⁃Attention and CNN⁃BiL⁃STM⁃Attention,and introduced TimeGAN separately to compare the prediction accuracy of each model.The results show that the mixed model based on TimeGAN and CNN⁃BiLSTM⁃Attention has significantly better fitting performance than other models,and its predicted values are closest to the monitored values.Compared to traditional single LSTM models,its EMS,ERMS and EMA are reduced by 71%,49%and 45%respectively,and R2 is improved by 20%.

关 键 词：TimeGAN CNN BiLSTM ATTENTION 混凝土面板堆石坝 变形预测 

分 类 号：TV641.4[水利工程—水利水电工程]