基于CNN-BiLSTM-Attention的重力坝稳定时变安全系数预测模型  

Prediction Model for Time-varying Safety Factor for Gravity Dam Stability Based on CNNBiLSTM-Attention

作　　者：曹宇鑫 张瀚[1,2] 尹金超 李亚楠 CAO Yuxin;ZHANG Han;YIN Jinchao;LI Yanan(State Key Laboratory of Hydraulics and Mountain River Engineering,Sichuan University,Chengdu 610065,China;College of Water Resources&Hydropower,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学水力学与山区河流开发保护国家重点实验室,四川成都610065 [2]四川大学水利水电学院,四川成都610065

出　　处：《人民珠江》2025年第4期1-8,共8页Pearl River

基　　金：四川省科技厅重点研发项目(2022YFS0535)。

摘　　要：在高水压和高渗压等复杂工况作用下,准确把握重力坝安全系数的时变规律并进行有效预测,对于大坝运行状态的科学管控至关重要。为此,基于深度学习理论的CNN-BiLSTM-Attention方法,提出以上游水位、坝顶顺河向位移、时效为自变量,抗滑稳定系数为因变量的耦联预测模型。通过对某坝高148.0 m的重力坝工程分析,模型的拟合平均绝对误差(Mean Absolute Error,MAE)和均方误差(Root Mean Square Error,RMSE)为1.12×10-3和1.66×10-3,预测误差MAE、RMSE分别为3.08×10-3和3.53×10-3,与传统统计回归方法相比,预测精度提高了51.80%和45.44%,与SVM(Support Vector Machine)算法相比,预测精度提高了16.08%和10.18%,显示出对有限元计算结果曲线更好的吻合度,预测精度优势也较为明显。Under complex working conditions such as high water pressure and high seepage pressure,accurately grasping the timevarying law of the safety factor of gravity dams and effectively predicting it are crucial for the scientific control of the dam's operation status.To this end,a coupled prediction model is proposed based on the CNN-BiLSTM-Attention method of deep learning theory,with the upstream water level,the riverward displacement at the dam crest,and time-dependent effects as independent variables,and the anti-sliding stability coefficient as the dependent variable.Through the analysis of a gravity dam project with a height of 148.0 meters,the model demonstrates a mean absolute error(MAE)and root mean square error(RMSE)of 1.12×10-3 and 1.66×10-3,respectively,prediction errors MAE and RMSE of 3.08×10-3 and 3.53×10-3,respectively.Compared to traditional statistical regression methods,this model has increased the prediction accuracy by 51.80%and 45.44%,and when compared to the SVM algorithm,the prediction accuracy has increased by 16.08%and 10.18%,respectively.This indicates that the proposed model has a better alignment with the finite element calculation result curves and a more remarkable advantage in prediction accuracy.

关 键 词：CNN-BiLSTM-Attention 重力坝 预警指标 预测模型 

分 类 号：TV3[水利工程—水工结构工程]