Graph Convolutional Network-based Model for Incident-related Congestion Prediction: A Case Study of Shanghai Expressways

Abstract

Traffic congestion has become a significant obstacle to the development of mega cities in China. Although local governments have used many resources in constructing road infrastructure, it is still insufficient for the increasing traffic demands. As a first step toward optimizing real-time traffic control, this study uses Shanghai Expressways as a case study to predict incident-related congestions. Our study proposes a graph convolutional network-based model to identify correlations in multi-dimensional sensor-detected data, while simultaneously taking into account environmental, spatiotemporal, and network features in predicting traffic conditions immediately after a traffic incident. The average accuracy, average AUC, and average F-1 score of the predictive model are 92.78%, 95.98%, and 88.78%, respectively, on small-scale ground-truth data. Furthermore, we improve the predictive model’s performance using semi-supervised learning by including more unlabeled data instances. As a result, the accuracy, AUC, and F-1 score of the model increase by 2.69%, 1.25%, and 4.72%, respectively. The findings of this article have important implications that can be used to improve the management and development of Expressways in Shanghai, as well as other metropolitan areas in China.

RadNet: Incident Prediction in Spatio-Temporal Road Graph Networks Using Traffic Forecasting

“…It then combines the two using Kalman Filtering. More recently, Transformer blocks have been identified to provide accurate temporal models for forecasting, while additionally enabling parallelized computation for efficient training [Vaswani et al, 2017;Tuli et al, 2022]. A recent work, TSE-SC [Cai et al, 2020], sequentially infers the input using a GCNs and then a Transformer encoder.…”

“…For a graph G of a road network, at each timestep t, we encode the input window W (t) to perform both permutations of spatial and temporal inferences as described in Section 1. For spatial inference, we use graph attention networks (GAT) [Veličković et al, 2017] and for temporal inference we use transformer blocks [Vaswani et al, 2017]. Spatial-Inference.…”

“…We aggregate the outputs of the multiple heads in the final layer to ensure same feature size as the input [Veličković et al, 2017]. The stacked matrix of h i ∀i is denoted by…”

“…We also measure the incident diagnosis performance, such that we could identify the links at which incidents occur, rather than just determining there is an incident somewhere in the entire network. To measure the diagnosis performance, we use HitRate@P% [Su et al, 2019b] and NDCG@P% (normalized discounted cumulative gain) [Järvelin and Kekäläinen, 2002] where P% is 100 and 150 for both metrics as seen in prior work [Tuli et al, 2022;Zhao et al, 2020]. Here, these assess how many of the ground-truth links have been included in the top predicted incident candidates.…”

“…Spatio-temporal forecasting is a critical problem for dynamic industrial environments such as transportation systems, supply chains and mobile computation systems [Li et al, 2018;Tascikaraoglu, 2018;Yang et al, 2019]. One such environment is the network of road traffic systems.…”