基于图卷积网络的交通预测方法研究

Graph convolutional neural network based traffic prediction

由于交通预测问题的时空复杂性,在智能交通系统中完成预测是一项具有挑战性的任务。虽然交通预测的时间依赖性已经得到很好的研究和讨论,但由于空间依赖的变化较大,特别是在城市复杂交通环境中,对空间依赖的交通预测研究相对较少。该文提出一种新的图卷积预测网络模型,并将其应用于两个具有不同几何约束的城市交通网络。首先,该模型利用多重加权邻接矩阵对速度数据进行图形卷积运算,将速度限制、距离和道路角度等特征组合在一起。其次,对组合特征进行空间隔离降维运算,以学习特征之间的依赖关系,并将输出的大小降低到计算可行水平。然后,将多权图卷积网络的输出应用于具有长短期记忆单元模型,以学习时间依赖。最后,将所提出的预测网络应用于城市核心区和城市混合区两个交通网络,其性能不仅优于其余六种比较模型,而且降低了城市混合区交通网络的预测方差。结果表明,所提出的预测网络能够在不同的空间复杂度下提供稳健的交通预测性能,这在城市交通预测中具有很强的优势。

Due to the spatial and temporal complexity of the traffic prediction problem,accomplishing prediction in intelligent transportation systems is a challenging task.Although the time dependence of traffic prediction has been well studied and discussed,relatively little research has been done on spatially dependent traffic prediction due to the large variability of spatial dependence,especially in complex urban traffic environments.In this paper,a new graph convolution prediction network model is proposed and applied to two urban traffic networks with different geometric constraints.First,the model performs graph convolution operations on speed data using multiple weighted adjacency matrices to combine features such as speed limits,distances,and road angles.Second,spatially isolated dimensionality reduction operations are performed on the combined features to learn the dependencies between the features and reduce the size of the output to a computationally feasible level.Then,the output of the multi-weighted graph convolutional network is applied to a model with long and short-term memory units to learn temporal dependencies.Finally,the proposed prediction network is applied to two traffic networks in urban core and mixed urban areas,and its performance not only outperforms the remaining six comparative models,but also reduces the prediction variance of the traffic network in mixed urban areas.The results show that the proposed prediction network can provide robust traffic prediction performance under different spatial complexities,which is a strong advantage in urban traffic prediction.