基于图注意力网络的交通流量时空预测模型

A Spatio-temporal Prediction Model of Traffic Volume Based on Graph Attention Network

准确的交通流量预测是实现城市路网数据管控的重要任务,传统的方法往往忽略了交通流量因素之间的相互影响以及交通网络的时空依赖性。提出了一种基于图注意力网络的时空预测模型用于解决上述问题,命名为ST-GATC。在输入层,同时将多个交通流量变量作为输入,挖掘各因素之间的非线性关系。在模型方面,首先将道路网络建模为时空有向图,使用卷积模块对样本在时间步上的特征进行捕捉。然后采用多头注意力图网络聚合道路网络中节点的邻居信息,使每个注意力机制分别处理一个子空间,为目标节点的不同邻接节点分配不同的注意力权重。最后使用一个全连接层对数据进行输出。结果表明,在模型对照试验中,相较于对照组模型,ST-GATC模型的RMSE降低了约0.571~9.288,MAE降低约0.314~7.678,证明了ST-GATC模型在交通流量预测任务中表现优秀,可以用做交通流量预测系统的预测模型。在参数试验中,分别进行了多组试验,获取模型拟合效果随注意力头数、神经元个数及输入因素个数增长的变化趋势,对参数的最佳数值进行了选取。结果显示,该模型在输入维度为2,注意力头数为3,隐藏层神经元个数为64时具备较为优秀的预测效果。

Accurate traffic volume prediction is an important task to achieve urban road network data management and control. Traditional methods often ignore the interaction between traffic flow factors and the spatio-temporal dependence of traffic networks. A spatio-temporal prediction model named ST-GATC based on graph attention network is proposed to solve the above problems. In the input layer, multiple traffic flow variables are taken as input at the same time, and the nonlinear relationships among these factors are mined. In terms of model, first, the road network is modeled as a spatio-temporal directed graph, and the features of samples at time steps are captured by using convolution module. Then, the neighbor information of nodes in the road network is aggregated by using multi-head attention graph network, so that each attention mechanism can process a subspace separately and assign suitable attention weight to each adjacent node of the target node. Finally, the data are output by using a fully connected layer. The result shows that in the model control experiment, the RMSE of the ST-GATC model is reduced by about 0.571-9.288,and the MAE is reduced by about 0.314-7.678 compared with the control group model, it is proved that the ST-GATC model performs well in the traffic volume prediction and can be used as the prediction model of the traffic volume prediction system. In the parameter experiment, several groups of experiments are carried out to obtain the changing trends of the model fitting effect with the increase of the numbers of attention heads, neurons and the input factors, and the optimal values of the parameters are selected. The result shows that the model has a relatively good prediction effect when the input dimension is 2,the number of attention heads is 3,and the number of neurons in the hidden layer is 64.