基于多变量时间序列和LSTM网络的面板缝变形预测  

Prediction of slab joint deformation based on multivariate time series and LSTM network

作　　者：谭瀛 马刚[1] 徐建华 程翔 冷天培 周伟[1] TAN Ying;MA Gang;XU Jianhua;CHENG Xiang;LENG Tianpei;ZHOU Wei(State Key Laboratory of Water Resources and Hydropower Engineering Science,Wuhan University,Wuhan 430072,China;Hubei Energy Group Loushui Hydropower Co.,Ltd.,Enshi 445000,China;Chinese Antarctic Center of Surveying and Mapping,Wuhan University,Wuhan 430072,China)

机构地区：[1]武汉大学水资源与水电工程国家重点实验室,湖北武汉430072 [2]湖北能源集团溇水水电有限公司,湖北恩施445000 [3]武汉大学中国南极测绘研究中心,湖北武汉430072

出　　处：《人民长江》2022年第10期198-204,共7页Yangtze River

基　　金：国家杰出青年科学基金项目(51825905);国家自然科学雅砻江联合基金重点项目(U1865204)。

摘　　要：混凝土面板是面板堆石坝的主要防渗结构,其接缝部位的变形对大坝防渗和安全稳定至关重要。混凝土面板接缝系统受水位、堆石坝变形和温度等多个因素的共同影响,其变形发展演化是一个动态非线性的复杂过程。首先进行面板接缝变形与各影响因子之间的相关性分析,筛选出对面板接缝变形影响较大的多个变量的时间序列。然后通过滑动时间窗口构造时序数据并将其输入长短期记忆神经网络(LSTM),以面板接缝变形作为网络输出,训练得到基于多变量时间序列和LSTM网络的面板接缝变形预测模型。最后将模型预测结果与向量自回归模型(Vector Autoregression, VAR)、单变量LSTM模型及三次指数平滑模型(Holt-Winters)的预测结果进行对比分析。对比结果表明:多变量LSTM模型的预测精度较高,对开展面板接缝变形的实时动态预测有一定工程实用价值。Concrete face slab is the main seepage prevention structure of concrete face rockfill dams, and the deformations of joints are very important to the dam’s seepage prevention, safety and stability.The joint system of concrete slabs is affected by multiple factors, such as water levels, deformation of rockfill dams, and temperature.Therefore, the deformation and evolution of the joint system is a dynamic and nonlinear complex process.We use multiple time series of the slab joint deformation and its influencing factors and Long Short-Term Memory(LSTM)to establish a prediction model for the slab joint deformation.First, we analyze the correlation between the slab joint deformation and its influencing factors, in which the time series of the influencing factors that have a greater impact on the slab joint deformation will be selected.The selected time series data is constructed through sliding time windows and is input into the LSTM network.Taking the joint deformation as the network output, the prediction model of slab joint deformation based on multivariate time series and LSTM network is trained.The predicted results are compared with results of Vector Autoregression(VAR), univariate LSTM model and Holt-Winters model.The comparison results show that the multivariable LSTM model has higher prediction accuracy, and has certain engineering practical value for real-time dynamic prediction of slab joint deformation.

关 键 词：面板堆石坝 面板接缝 变形预测 多变量时间序列 LSTM 

分 类 号：TV641.4[水利工程—水利水电工程]