基于自适应时空图网络的交通预测

Traffic Prediction Based on Adaptive Spatio-Temporal Graph Network

在面对复杂的交通状况时,能够精准的提取出其中的时空依赖关系,使交通预测的准确率提高是构建智能交通系统的重要环节。针对交通图的时空特性,提出了一种基于自适应时空图网络的交通预测模型(Ada-STGN)。模型通过多层叠加门控残差时间卷积结构(GRes-TCN)处理时间依赖关系;使用自适应多头空间注意力模块(Ada-SAN)对不同子空间依赖模式进行联合建模获取空间依赖关系。分别在METR-LA和PEMS-BAY两个真实的公共交通数据集上进行了预测实验,在与基准模型比较后可知,Ada-STGN减小了预测误差,具有更好的预测性能。

Accurately extracting spatiotemporal dependencies in complex traffic situations and improving the accuracy of traffic prediction is an important step in building intelligent transportation systems.In this paper,a traffic prediction model based on adaptive spatiotemporal graph network(Ada-STGN) is proposed.The model deals with the time dependence by using a multi-layer stacking gating residual time convolution(GRes-TCN) structure Adaptive multi-head spatial attention module(Ada-SAN) is used to model different subspace dependence patterns and obtain spatial dependence relations.The prediction experiments were carried out on two real public transport datasets,METR-LA and PEMS-BAY,respectively.Compared with the baseline model,Ada-STGN reduces the prediction error and has better prediction performance.