摘要
【目标】为实现更加准确的高速公路短时交通流预测,提出一种基于ETC门架系统多模态数据的XGBoost-MSTAN交通流预测模型。【方法】首先,融合ETC系统中的图像、文本等多模态数据,在数据预处理的基础上,建立多模态交通数据集。其次,构建XGBoost模型进行初步短时流量预测,快速捕获数据中的非线性关系,生成初步预测结果。最后,利用MSTAN网络对XGBoost预测的残差进行建模,并结合Attention机制提取复杂的时空特征,进一步修正初步预测的误差,从而优化高速公路短时交通流的预测精度。【结果】XGBoost-MSTAN模型在不同预测时间和交通流条件下均表现出较高的准确性和稳定性。在未来5 min短时预测中效果最佳,交通流准确率超过90%,整体平均速度误差低于2.9 km/h。随着预测时间的延长,模型准确率逐渐下降,但整体误差仍控制在合理范围内。在未来120 min的预测中,车流量准确率仍超过80%,整体平均速度误差小于5.1 km/h。【结论】使用ETC门架系统多模态数据构建的XGBoost-MSTAN交通流预测模型,具有良好的预测性能和较强的泛化能力,可为高速公路交通管理部门制订科学合理的主动交通管理与控制政策提供参考。
[Objective]To achieve more accurate expressway short-term traffic flow prediction,the XGBoost-MSTAN traffic flow prediction model based on multimodal data from ETC gantry system was proposed.[Method]First,multimodal data,including images and texts from ETC system,were fused to construct a multimodal traffic dataset based on data preprocessing.Then,the XGBoost model was built for preliminary short-term traffic flow prediction to quickly capture nonlinear relations among data,and generate initial prediction result.Finally,the MSTAN network was used to model the residuals predicted with XGBoost.Attention mechanism was employed to extract the complex spatiotemporal features,further correcting the errors of initial prediction.[Result]XGBoost-MSTAN model performs well in terms of accuracy and stability in different prediction times and traffic flow conditions.The model aims to achieve the best performance in short-term forecasts,predicting 5 minutes into the future.traffic flow prediction accuracy exceeding 90%,and average speed error below 2.9 km/h.The accuracy gradually decreases as the prediction time increases,while the overall error remains within a reasonable range.For upcoming 120-minute prediction,the accuracy remains over 80%,and the average speed error is below 5.1 km/h.[Conclusion]XGBoost-MSTAN traffic flow prediction model,constructed with multimodal data from ETC gantry system,has excellent predictive performances and strong generalization ability.It can provide references for expressway traffic administration departments to formulate scientific and reasonable active traffic control policies.
作者
郭凤斌
潘崇柯
张思楠
张可可
郭庆锋
GUO Fengbin;PAN Chongke;ZHANG Sinan;ZHANG Keke;GUO Qingfeng(Gansu Highway Construction Management Group Co.,Ltd.,Lanzhou,Gansu 730030,China;Jiaoke Transport Consultants Ltd.,Beijing 100191,China;National and Local Joint Engineering Research Center for Intelligent Vehicle Road Collaboration and Safety Technology,Tianjin 300222,China;TongDun Technology Co.,Ltd.,Hangzhou,Zhejiang 310000,China)
出处
《公路交通科技》
北大核心
2025年第11期38-46,共9页
Journal of Highway and Transportation Research and Development
