Predicting Traffic Speed Under the Impact of Maintenance Downtime with Graph Convolutional Networks



Lu, Yuanjie

Journal Title

Journal ISSN

Volume Title



In this paper, a traffic forecasting approach based on graph convolutional networks is proposed to learn the effect of maintenance downtime on the surrounding area. Since inappropriate construction can lead to traffic congestion, disturbance and accidents, it is important to evaluate the work zone downtime effect. Furthermore, learning the correlation between each roadway is essential because if construction is carried out, the upstream area will not only block the downstream area, but also affect the surrounding areas. In order to evaluate maintenance downtime effect on traffic roads, traffic speed predictions are used by many researchers as the first experiment to quantify the impact of construction work on traffic states. However, in early studies, there are many machine learning models and time-series models used to predict the abrupt changes in the traffic speed and quantify the impact of maintenance work, but these models cannot explain nonlinear relationships between the speed with traffic incidents, nor can they dynamically describe the causes of traffic speed changes, thus it is difficult to assess the impact of work zones in more complex traffic environments. In addition, few studies based on the deep learning approaches are used to measure the impact of traffic construction on the surrounding area, thus it is hard to reference the latest literature based on deep learning to tackle the problem. To predict traffic speed under the impact of construction work, we design a novel model based on the graph convolutional neural network (GCN) to accommodate the spatial-temporal dependencies among traffic states, differentiate the intensity of connecting to neighboring roads and predict the speed under the road maintenance condition. The advantage of using a graph convolutional network is that it can transfer real-time traffic information between each road segment and adjacent segments, and automatically learn the non-structure features under the impacts of traffic incidents. The more nodes and layers the network has, the more information is involved in the calculation because each road segment represents a node and shares its parameters throughout the network. If we use the graph model to construct a node and find the correlation of its adjacent nodes, and then compare the predictive speed according to historical data, we can quantify the influence of traffic incidents in the area, making it possible to assess the economic losses in the event of traffic accidents. In the experiment, we compare our model with four baselines on two real-world datasets which are collected from the road sensor network around Tyson’s Corner and Los Angeles. The result shows that our model is better than other benchmarks and proves that it is feasible to provide more traffic information for graph-based models to improve traffic flow prediction.



Deep learning, Traffic speed, Laplacian matrix, Graph neural network, Maintenance downtime, Spatio-temporal correlation