Traffic Flow Forecasting with Spatial-Temporal Graph Convolutional Networks in Edge-Computing Systems

Traffic forecasting is one of important functions in Intelligent transportation systems (ITSs) and is of great significance to user experience and urban traffic control. Edge computing, has been recognized as a promising technique for real-time and accurate traffic flow forecasting. In this paper, we propose Spatial-Temporal Graph Convolutional Networks in an edge-computing system (STGCN-EC) for traffic flow forecasting. Firstly, we model the road network as a graph and partition it into multiple subgraph according to the spatial correlation of the area so that traffic flow forecasting can be individually performed by each edge node. Secondly, by taking the geographic information and temporal similarity of the traffic flow into account, we propose a spatial-temporal Graph convolutional network to efficiently capture the spatial-temporal features for traffic flow prediction in each subgraph. In addition, we adopt transfer learning to share models among different edge nodes to further improve training efficiency. Simulation results on real- world dataset demonstrate that the proposed approach is able to improve prediction accuracy and training efficiency in an edge-computing system.