基于动态时空图卷积的多传感器路网交通流预测

Multi-Sensor Road Network Traffic Flow Prediction Based on Dynamic Spatio-Temporal Graph Convolution

摘要: 准确的交通预测为城市发展提供规划支持，然而交通流预测精度取决于多重影响因素，道路之间的复杂程度与时间逻辑上的变化规律，现有方法无法合理分析路网交通模式的动态时空相关性，为解决这些问题本文提出一种结合动态自适应(Dynamic Self-adapting)、时空注意力机制(TS At-tention)、基准自适应机制(Benchmark Adaptive Mechanism)与空洞卷积(Dilated Convolution)的切比雪夫图卷积神经网络(GCN)。该模型采用时空注意力机制提取时间与空间动态相关性，结合切比雪夫图卷积神经网络获取交通流数据空间依赖关系，同时，将GCN的输出作为输入，该网络引入空洞卷积扩展感受野范围和提取时间和周期依赖关系，增加残差模块以构建时空残差网络，最后多模块融合预测。

Abstract: Accurate traffic prediction provides planning support for urban development. However, the accuracy of traffic flow prediction depends on multiple influencing factors, the complexity between roads and the temporal logic change rule, and the existing methods cannot reasonably analyze the dynamic temporal and spatial correlation of road network traffic patterns. To solve these problems, this paper proposes a novel adapting method that combines Dynamic Self-adapting, TS Attention, Benchmark Adaptive Mechanism and Dilated convolution Convolution of Chebyshev graph convolutional neural networks (GCN). This model uses spatio-temporal attention mechanism to extract the dynamic correlation between time and space, and combines Chebyshev graph convolutional neural network to obtain the spatial dependency of traffic flow data. At the same time, the output of GCN is taken as the input, the network introduces the cavity convolution to extend the receptive field range and extract the time and period dependence, and adds the residual module to construct the spatiotemporal residual network. Finally, multiple modules are fused and predicted.