融合交通事件信息的短时交通流参数预测  

Prediction of Short-term Traffic Flow Parameter based on Traffic Event Information

作　　者：张静 陈诺 鲍钱涵 李斌 金盛[3,5] ZHANG Jing;CHEN Nuo;BAO Qian-han;LI Bin;JIN Sheng(Zhejiang Highway and Water Transport Engineering Consulting Group Co.Ltd.,Hangzhou 310000,China;China Wenzhou Maritime Safety Administration,Wenzhou 325002,China;College of Civil Engineering and Architecture,Zhejiang University,Hangzhou 310058,China;Hangzhou Wengine Network Technology Co.Ltd.,Hangzhou 310053,China;Center for Balance Architecture,Zhejiang University,Hangzhou 310058,China)

机构地区：[1]浙江公路水运工程咨询集团有限公司,杭州市310000 [2]中华人民共和国温州海事局,温州市325002 [3]浙江大学建筑工程学院,杭州市310058 [4]杭州恒生芸擎网络科技有限公司,杭州市310053 [5]浙江大学平衡建筑研究中心,杭州市310058

出　　处：《公路》2024年第10期241-250,共10页Highway

基　　金：浙江省交通运输厅科技计划项目,项目编号202212;国家自然科学基金,项目编号72361137006;浙江省杰出青年基金,项目编号LR23E080002。

摘　　要：为提高交通流参数预测精度,提出了一种融合交通事件信息的短时交通流参数预测模型(Incident Fusion Dynamic Spatial Temporal Network,IF-DSTN)。首先,IF-DSTN模型引入加州算法进行路网交通流事件状态检测计算,考虑道路事件发生在时空上对于整个路网交通流的影响,提高模型在事件发生等异常状态下的响应能力。其次,模型基于Transformer架构以及图卷积网络架构分别构建空间注意力模块(Spatial Transformer,ST)和时间注意力模块(Temporal Transformer,TT),模块组合成时空注意力模块共同提取交通流数据中的时空相关性。然后,模型融合事件检测的数据以及时空注意力模块数据,经过卷积操作得出短时交通流参数的预测结果。最后,设计实验讨论了IF-DSTN模型在路网交通流预测精度上对比其他优秀模型的优势。结果表明:对于一段布设有200余个检测点的高速公路,本文模型预测结果平均绝对百分比误差为4.418%,均方根差为16.084辆/h。相较于其他未考虑交通事件检测信息的模型有较大地提升,并在未经计划的交通事件发生时具有明显优势,强调了事件影响下的路网时空关系变化,模型能够在交通事件影响下实现更准确地预测。In the paper,an Incident Fusion Dynamic Spatial-Temporal Network(IF-DSTN)that incorporates traffic event information is proposed.Firstly,the IF-DSTN model incorporates the California algorithm to detect and assess the state of traffic flow events within the road network.This accounts for the spatial-temporal influence of road events on the overall traffic flow and augments the model's responsiveness during abnormal states,such as incidents.Secondly,the model encompasses a Spatial Transformer Module (ST) and a Temporal Transformer Module (TT), leveraging the Transformer andGraph Convolutional Network architectures respectively. These modules are integrated into a Spatial-Temporal Attention Module, collectively extracting spatial-temporal correlations from traffic flow data.Subsequently, the model amalgamates the data from event detection with that from the spatial-temporalattention module, culminating in the prediction of short-term traffic flow parameters through convolutionaloperations. Finally, an experiment is devised to assess the superiority of the IF-DSTN model in comparisonto other exemplary models concerning traffic flow prediction accuracy. The results demonstrate an averageabsolute percentage error of 4. 418% and a root mean square error of 16. 084 vehicles/hour for a segment ofexpressway featuring over 200 inspection points. Notably, in cases of unforeseen traffic incidents, the IFDSTNmodel exhibits marked advantages over models that neglect traffic incident detection information,underscores the shifts in the spatial-temporal relationships within the road network induced by incidents,resulting in more precise predictions under their influence.

关 键 词：交通工程 交通流参数预测 深度学习 高速公路交通流 时空相关性 

分 类 号：U491.112[交通运输工程—交通运输规划与管理]