文章摘要
郭海锋*,许宏伟**,周子盛**.基于GRU的密集连接时空图注意力网络的城市交通预测[J].高技术通讯(中文),2024,34(5):463~474
基于GRU的密集连接时空图注意力网络的城市交通预测
Urban traffic prediction based on densely connected spatial-temporal graph attention network of GRU
  
DOI:10. 3772 / j. issn. 1002-0470. 2024. 05. 003
中文关键词: 交通预测; 时空特征; 神经网络; 门控循环单元(GRU); 密集连接; 图注意力网络(GAT)
英文关键词: traffic prediction, spatio-temporal features, neural network, gated recurrent unit (GRU), densely connected, graph attention network (GAT)
基金项目:
作者单位
郭海锋* (*浙江工业大学网络空间安全研究院杭州 310014) (**浙江工业大学信息工程学院杭州 310023) 
许宏伟**  
周子盛**  
摘要点击次数: 195
全文下载次数: 183
中文摘要:
      城市道路拓扑结构的复杂性、交通流量的实时变化以及多元的外部环境等因素给交通预测带来了极大的困难。现有方法对交通路网的时空特征挖掘性不足,缺乏对外部因素的考虑,为此本文提出了一种基于门控循环单元(GRU)的时空图注意力密集连接网络,通过门控循环单元来捕获路网数据的动态规律,并以图注意力密集连接网络来提取路网复杂的空间结构特征,建立城市交通网络对时空的依赖关系。针对外部客观因素,采用独热编码的方式对城市各路段发生的交通事件进行数据建模,增强交通网络的信息属性。以杭州申花路及周围共309个路段为例,对所提出模型的预测能力和可行性进行验证。实验结果表明,模型预测精度最高达到了81.64%,与传统数学模型和主流的神经网络模型对比,预测精度较ARIMA提高了35.42%,较图注意力网络(GAT)和GRU神经网络分别提高了17.45%和3.02%。实验证明该方法可以适应复杂的交通流进行长期的交通预测任务,同时也能增强交通管理能力,减少交通拥堵成本。
英文摘要:
      Due to the complexity topology of urban traffic network, the real-time change of traffic flow and external environmental factors, there are huge difficulties in traffic prediction. In view of the inadequacy of existing methods in mining the spatio-temporal features of road network and the insufficient consideration of external factors, a spatial-temporal network of dense graph attention network based on gated recurrent unit (GRU) (DG-GRU) is proposed. The function of gated recurrent unit is used to capture the dynamic changes of road network data. Densely connected graph attention network (GAT) is used to extract the complex spatial structure characteristics of the road network. They can establish the dependence of urban traffic network data on time and space. Considering the influence of external factors, the one-hot encoding is used to model the traffic events that occur in urban sections to enhance the information attributes of transportation network. Taking Shenhua Road and its surrounding sections in Hangzhou as an example to verify the predictive ability and feasibility of the network. The experimental results illustrate that the prediction accuracy of the method is up to 81.64%. Compared with traditional mathematical model and mainstream neural network model, the prediction accuracy of DG-GRU is 35.42% higher than that of ARIMA. Compared with graph attention network (GAT) and GRU neural networks, its prediction accuracy is improved by 17.45% and 3.02%, respectively. Experimental results show that the model in this paper can adapt to complex traffic flow and carry out long-term traffic forecasting tasks. Meanwhile, it can enhance traffic management ability and reduce the costs traffic congestion.
查看全文   查看/发表评论  下载PDF阅读器
关闭

分享按钮