公路工程

为精准预测斜拉桥主梁挠度，兼顾其多尺度波动特性与长短期时序依赖关系，提出了一种基于DWT-CNN-LSTM组合算法数据驱动的斜拉桥主梁挠度预测模型。首先通过离散小波变换对挠度序列进行多尺度分解，分离不同波动成分；利用1D-CNN提取各子带序列的时间局部变形模式；引入残差模块与注意力机制改进LSTM,捕捉序列长期依赖关系；最终通过DWT逆变换重构子带预测结果，得到完整挠度序列。以某双塔斜拉桥为工程背景，开展短时段与长时段跨中挠度监测数据的训练与预测试验。结果表明：改进LSTM的损失曲线下降速率更快、稳定值更低，相较于CNN-LSTM、LSTM模型，训练效率与拟合精度显著更优；DWT-CNN-LSTM能较好地拟合长、短时段实测值的整体演变与局部波动，短时段平均相对误差约为1.61%,长时段平均相对误差约为1.12%;长时段跨中挠度预测的极大、极小值回归曲线拟合决定系数分别高达0.986 5、0.983 8,体现模型对挠度极值特征的精准复现能力；24 h尺度下模型置信区间宽度稳定，基于训练残差构建的区间有效量化长时序预测不确定性，无随时间推移的发散或突变拓宽现象。

To accurately predict the deflection of the main beam of a cable-stayed bridge, taking into account its multi-scale fluctuation characteristics and long short-term temporal dependencies, a data-driven prediction model for the deflection of the main beam of a cable-stayed bridge based on the DWT-CNN-LSTM combination algorithm is proposed. Firstly, the deflection sequence is decomposed into multiple scales using discrete wavelet transform to separate different wave components; extracting temporal local deformation patterns of each sub-band sequence using 1D-CNN;introducing residual modules and attention mechanisms to improve LSTM and capture long-term dependencies of sequences; Finally, the sub-band prediction results were reconstructed through DWT inverse transformation to obtain the complete deflection sequence. Conduct training and prediction experiments on short-term and long-term mid span deflection monitoring data using a certain twin tower cable-stayed bridge as the engineering background. The results show that the improved LSTM has a faster decline rate and lower stable value in the loss curve. Compared with CNN-LSTM and LSTM models, the training efficiency and fitting accuracy are significantly better; DWT-CNN-LSTM can well fit the overall evolution and local fluctuations of measured values in long and short time periods, with an average relative error of about 1.61% in the short time period and about 1.12% in the long time period; the determination coefficients of the regression curves for the maximum and minimum values of long-term mid span deflection prediction are as high as 0.986 5 and 0.983 8, respectively, reflecting the model's ability to accurately reproduce the extreme deflection characteristics; at the 24 h scale, the confidence interval width of the model is stable, and the interval constructed based on training residuals effectively quantifies the uncertainty of long-term forecasting, without any divergence or abrupt broadening phenomenon over time.