融合特征因子筛选的拱坝变形深度学习预测模型  

Deep learning prediction model for arch dam deformation by incorporating feature factor screening

作　　者：刘桓辰 朱静 郭梦京[2] LIU Huanchen;ZHU Jing;GUO Mengjing(Shanxi Province Institute of Water Resources and Electric Power Investigation and Design,Xi’an 710001,Shaanxi,China;Institute of Water Resources and Hydroelectric Engineering,Xi’an University of Technology,Xi’an 710000,Shaanxi,China)

机构地区：[1]陕西省水利电力勘测设计研究院,陕西西安710001 [2]西安理工大学水利水电学院,陕西西安710000

出　　处：《水利水电技术(中英文)》2025年第3期123-134,共12页Water Resources and Hydropower Engineering

基　　金：国家自然科学基金项目(41807156);陕西省教育厅重点实验室科研计划项目(18JS073)。

摘　　要：【目的】变形是库水、温度和材料特性等多因素耦合作用下大坝整体服役性态的直接表征,建立精确、高效的预测模型对于掌握坝体变形趋势和评估大坝风险具有重要意义。【方法】针对传统预测模型精度低、适应性差和抗噪能力弱等问题,将哈里斯鹰算法(HHO)、变分模态分解(VMD)、随机森林算法(RF)和长短时记忆神经网络(LSTM)相结合,提出了一种混凝土拱坝变形深度学习预测模型。首先,通过引入Tent混沌映射、能量随机性递减策略改进HHO算法,利用IHHO-VMD方法分解拱坝变形数据序列得到若干不同频率的模态分量(IMF);其次,利用RF算法计算变形特征因子的贡献率,筛选预测模型最优输入因子集合;最后,采用LSTM模型对各IMF分量进行学习和预测,重构各分量预测值得到最终的变形预测值。【结果】仿真信号分解结果表明:与现有信号分解方法相比,采用IHHO-VMD方法可以实现信号最优分解。通过某工程实例分析,所提模型预测4个测点位移时,平均RMSE、MAE、R^(2)和MAPE为0.3976 mm、0.3275 mm、0.9918和1.5194%。【结论】相较于其他组合模型,所提模型的4种评价指标结果均为最优,表明该模型具有预测精度高、泛化能力好和鲁棒性强等优势。[Objective]Deformation is a direct characterization of the overall serviceability of dams under the coupling of reservoir water,temperature and material properties,etc.The establishment of an accurate and efficient prediction model is of great significance in grasping the deformation trend of dams and assessing the risk of dams.[Methods]Aiming at the problems of low accuracy,poor adaptability and weak noise immunity of traditional prediction models,a deep learning prediction model for concrete arch dam deformation is proposed by combining the Harris Hawk algorithm(HHO),Variational Modal Decomposition(VMD),Random Forest algorithm(RF),and Long-Short-Term Memory neural network(LSTM).First,the HHO algorithm is improved by introducing Tent chaotic mapping,energy randomness decreasing strategy,and the arch dam deformation data sequence is decomposed to obtain a number of modal components(IMF)with different frequencies using the IHHO-VMD method.Secondly,The RF algorithm is utilized to calculate the contribution of deformed characteristic factor and to screen the optimal set of input factors for the prediction model;.Finally,the LSTM model is used to learn and predict each IMF component,and the final deformation prediction is obtained by reconstructing the predicted values of each component.[Results]The simulated signal decomposition result show that compared with the existing signal decomposition method,the optimal signal decomposition can be realized by using the IHHO-VMD method.Analyzed by a project example,the proposed model predicts the displacement of four measurement points with average RMSE,MAE,R^(2) and MAPE of 0.3976 mm,0.3275 mm,0.9918 and 1.5194%.[Conclusion]Compared with other combined models,the result of the four evaluation indexes of the proposed model are optimal,indicating that the model has the advantages of high prediction accuracy,good generalization ability and robustness.

关 键 词：混凝土拱坝变形 哈里斯鹰算法 变分模态分解 随机森林算法 长短时记忆神经网络 水利工程 变形 

分 类 号：TV698.11[水利工程—水利水电工程]