面向交通流量预测的多组件时空图卷积网络

流量预测一直是交通领域研究者和实践者关注的热点问题.流量数据具有高度的非线性和复杂性，对其进行精准预测具有很大的挑战，现有的预测方法大多不能很好地捕获数据的时空相关性.提出一种新颖的基于深度学习的多组件时空图卷积网络（MCSTGCN），以解决交通流量预测问题.MCSTGCN通过3个组件分别建模流量数据的近期、日周期、周周期特性，每个组件同时利用空间维图卷积和时间维卷积有效捕获交通数据的时空相关性.在美国加利福尼亚州高速公路流量公开数据集上进行了实验，结果表明，MCSTGCN模型的预测效果优于现有的预测方法.

Multi-component Spatial-temporal Graph Convolution Networks for Traffic Flow Forecasting

Forecasting the traffic flows is a hot issue for researchers and practitioners in the transportation field. It is very challenging to forecast the traffic flows due to the high nonlinearity and complexity of the data, and most of the existing methods cannot effectively capture the spatial-temporal correlations of traffic flow data. In this paper, we propose a novel deep learning based model, multi-component spatial-temporal graph convolution networks (MCSTGCN), to solve the problem of traffic flow forecasting. MCSTGCN employs three components to respectively model the recent, daily and weekly characteristics of traffic flow data. Each component uses graph convolutions in the spatial dimension and convolutions in the temporal dimension to effectively capture the spatial-temporal correlations of traffic data. Experiments on a public California freeway dataset show that the prediction performance of the MCSTGCN model is better than other existing prevalent methods.