摘要
交通流量预测在智慧交通建设中发挥着至关重要的作用。为了充分挖掘交通网络中结点之间的空间相关性,本文提出一种基于异质图的深度时空模型STREGCN。首先,本文提出将交通网络抽象为异质图,增强图的表达能力,进而充分捕获交通网络的空间相关性。其次,本文采用基于线性门控单元的一维因果卷积去充分提取交通流量的时间相关性。最后,本文设计时空卷积模块的输出经过全连接层获得最终的交通流量预测结果。本文在开源的交通数据集PEMSD8进行了预测区间为5分钟和30分钟的交通流量预测实验。实验结果表明STREGCN模型与大多数基线模型相比,在未来短期和长期的交通流量预测任务上都有更好的表现。
Traffic flow forecasting plays a crucial role in the construction of intelligent transportation. In order to fully exploit the spatial correlation between nodes in a traffic network, this paper proposes a deep spatio-temporal model STREGCN based on heterogeneous graphs. Firstly, this paper proposes to abstract the traffic network as a heterogeneous graph to enhance the graph’s expressiveness, and then fully capture the spatial correlation of the traffic network. Secondly, this paper uses one-dimensional causal convolution based on linear gating units to fully extract the temporal correlation of traffic flows. Finally, this paper designs the output of the spatio-temporal convolution module to obtain the final traffic flow prediction results after a fully connected layer. In this paper, traffic flow prediction experiments with prediction intervals of 5 and 30 minutes are conducted on the open-source traffic dataset PEMSD8. The experimental results show that the STREGCN model performs better than most baseline models for both short- and long-term future traffic flow forecasting tasks.
出处
《计算机科学与应用》
2023年第3期399-409,共11页
Computer Science and Application
