摘要
针对桥梁风荷载静力三分力现有时程预测精度不足的问题,提出了一种攻角融合卷积-Transformer(AFConv-Transformer)模型。该模型采用一维卷积网络来提取局部高频特征,利用Transformer编码器捕捉全局时序依赖,将攻角作为物理约束进行多模态融合,从而解决传统模型的相位偏差问题。然后,基于某大跨钢箱梁的风洞试验数据,生成860组样本集,对模型进行验证。消融试验结果表明,攻角融合有助于消除预测的相位偏差,卷积与Transformer编码器模块的协同作用是保证模型有效性的基础。在测试集上,所提模型的平均绝对误差、均方根误差和决定系数分别为0.354 7、0.654 3和0.976 8;相较于经典的攻角融合卷积-长短期记忆(AFConv-LSTM)模型,训练耗时从147.50 s降至65.60 s,效率提升55.5%。该研究为桥梁抗风设计中的气动力智能预测提供了一种高效可靠的新方法。
To address the issue of insufficient accuracy in the current time⁃history prediction of static three⁃component forces on bridges under wind loads,an angle⁃fused convolutional⁃transformer(AFConv⁃Transformer)model is proposed.A one⁃dimensional convolutional network is used to extract local high⁃frequency features and a transformer encoder is utilized to capture global time⁃series dependencies.Multi⁃modal fusion is conducted by taking the angle of attack as a physical constraint to resolve the phase de⁃viation problem of traditional models.Then,860 sample sets are generated based on wind tunnel test data from a large⁃span steel box girder to validate the model.The results of the ablation experiments show that the fusion of the angle of attack is benefical for eliminating phase deviation in the prediction,while the synergistic effect between the convolutional and transformer encoder modules forms the foundation of the model’s effec⁃tiveness.On the test set,the mean absolute error,the root mean square error,and the coefficient of determi⁃nation of the proposed model are 0.3547,0.6543,and 0.9768,respectively.Compared with the classic angle⁃fused convolutional⁃long short⁃term memory(AFConv⁃LSTM)model,the training time decreases from 147.50 to 65.60 s,marking a significant efficiency improvement of 55.5%.This research provides an effi⁃cient and reliable new method for intelligent prediction of aerodynamic forces in bridge wind⁃resistant design.
作者
孙洪鑫
罗臻懿
燕飞
张明
欧阳鹭伟
SUN Hongxin;LUO Zhenyi;YAN Fei;ZHANG Ming;OUYANG Luwei(School of Civil Engineering,Hunan University of Science and Technology,Xiangtan 411201,China)
出处
《东南大学学报(自然科学版)》
2026年第2期268-276,共9页
Journal of Southeast University:Natural Science Edition
基金
国家自然科学基金资助项目(52478514)。
关键词
三分力时程预测
桥梁抗风气动力
物理约束融合
攻角融合卷积-Transformer
训练效率优化
three⁃component force time⁃history prediction
bridge aerodynamics
physical constraint fusion
angle⁃fused convolutional⁃transformer(AFConv⁃Transformer)
training efficiency optimization
