公路隧道交通流预测是合理优化隧道运营管控方案的重要技术基础。针对交通流数据的非线性、时空耦合性以及降噪处理中扰动信息丢失等问题，提出了一种结合趋势建模与残差补偿的公路隧道交通流预测模型。首先，基于对交通流数据的趋势与统计特征分析，将原始数据划分为客车与货车两类；其次，结合平稳性检验结果，对非平稳数据采用高斯平滑方法进行降噪处理；然后，将平稳数据与降噪数据输入融合卷积神经网络、双向长短期记忆网络和时序自注意力机制的主干网络，以提取趋势特征，并引入时序卷积网络对残差数据进行建模，以恢复降噪处理中丢失具有时序结构的扰动特征；最后，通过融合趋势特征与扰动特征，生成最终预测结果。选取秦岭终南山公路隧道的小时流量数据进行模型训练和测试，实验结果表明：在对总体流量的连续预测中，集成深度学习模型预测的均方根误差为42.29±5.66 pcu，加权平均绝对百分比误差为4.18±1.06%，与其他预测模型的最优结果(64.28±7.84 pcu、6.57±1.08%)相比，两类误差分别降低了34.20%、36.37%。在消融实验中，残差补偿模块对模型性能影响最显著，可使两类误差分别降低26.97%、35.00%。研究结果为公路隧道智能交通系统的动态流量预测提供了理论基础。

Traffic flow prediction for highway tunnels is a crucial technical foundation for rationally optimizing tunnel operation and management strategies. To address the nonlinearity, spatiotemporal coupling of traffic flow data, and the loss of disturbance information during noise reduction, a highway tunnel traffic flow prediction model combining trend modeling and residual compensation is proposed. Firstly, based on the analysis of the trend and statistical characteristics of traffic flow data, the original data is divided into two categories: passenger cars and trucks. Secondly, combined with the stationarity test results, the Gaussian Smoothing method is applied to denoise the non-stationary data. Then, both the stationary and denoised data are input into a backbone network that integrates the Convolutional Neural Network, the Bidirectional Long Short-Term Memory network and the Temporal Self-Attention mechanism to extract trend features, and the Temporal Convolutional Network is introduced to model the residual data to restore the disturbance features with temporal structure lost during denoising. Finally, the trend features and disturbance features are fused to generate the final prediction results. The hourly traffic volume data of the Qinling Zhongnanshan Highway Tunnel is used for model training and testing. The experimental results show that in the continuous prediction of the overall traffic volume, the Root Mean Square Error of the prediction by the Integrated Deep Learning Model is 42.29±5.66 pcu, and the Weighted Mean Absolute Percentage Error is 4.18±1.06%. Compared with the best results of other prediction models (64.28±7.84 pcu, 6.57±1.08%), the two types of errors are reduced by 34.20% and 36.37%, respectively. In the ablation experiment, the residual compensation module has the most significant impact on the model performance, reducing the two types of errors by 26.97% and 35.00%, respectively. The research results provide a theoretical basis for dynamic traffic flow prediction in intelligent transportation systems of highway tunnels.