时空融合的堆石坝变形预测模型及在安全监测中的应用

Spatial-temporal fusion model for deformation prediction of rockfill dams and its application in safety monitoring

变形预测是堆石坝安全监测与健康诊断的重要手段,现有研究多根据堆石坝监测数据建立单测点预测模型,未充分考虑测点之间相关性进行整体建模,且现有模型难以对漂移数据进行长期精准预测。本文考虑堆石坝变形序列的时序依赖性和测点之间的协同相关性,提出了基于图卷积和循环神经网络、引入概率预测与全过程训练的时空融合变形预测模型。该模型首先采用图卷积网络对多测点特征进行自适应汇聚,然后利用循环神经网络中细胞状态与隐层记忆沿时间轴的传递性,实现对时空信息的挖掘与融合,最后通过线性层得到概率预测参数,提高了模型对监测数据噪声的鲁棒性。采用全过程训练方式,提高模型对影响因子与累积变形量内在关系的学习能力,实现对漂移数据的长期精准预测。最后以水布垭面板堆石坝为例,进行了模型对比实验与消融实验,介绍了该模型在堆石坝安全监测和健康诊断中的三种具体应用。结果表明,本文模型有效融合了时空信息,在预测精度方面显著高于现有模型,解决了现有模型对大坝整体变形规律学习能力差、漂移数据预测精度低的问题,可用于堆石坝变形长期预测、测点异常检测与缺损数据补全。

Deformation prediction is the key for safety monitoring and health assessment for rockfill dams.Current research mostly focuses on single-point deformation prediction models,neglecting the multi-point correlation for the overall modeling.Besides,it is challenging for current models to achieve long-term accurate prediction of drift deformation data.Considering temporal dependence of time series and spatial correlation between multipoint for the deformation of rockfill dams,a spatial-temporal fusion model based on Graph Convolutional Network(GCN)and Recurrent Neural Network(RNN)is proposed for deformation prediction,introducing probabilistic prediction and full-process training.Firstly,the model adaptively performs multipoint features fusion using GCN.Then,the transmissibility of cell states and hidden memories along the time axis in RNN is utilized to realize the mining and fusion of spatial-temporal information.Finally,the parameters of the probabilistic prediction are obtained as linear layer output to improve the model’s robustness against noise in monitoring data.In order to enhance its ability to understand the intrinsic relationship between influencing factors and cumulative deformation,the model adopts a full-process training and inference technique,which realizes long-term accurate prediction for drift deformation data.Taking Shuibuya concrete-faced rockfill dam as a study case,we conduct comparison and ablation experiment,then present three specific applications of this model in safety monitoring and health assessment for rockfill dams.Our results demonstrate that the model successfully integrates the spatial-temporal information,significantly improving prediction accuracy compared to current models.It addresses the challenges of learning the general law properly and predicting drift deformation data accurately of rockfill dams,and can be applied for long-term deformation prediction,anomaly detection and missing data completion of measurement points.