Spatio-temporal Graph Convolutional Neural Network for traffic signal prediction in large-scale urban networks

This research aims at tackling the traffic signal problem for large-scale networks via a deep learning approach. Our ultimate goal is to construct an automatic traffic management system, where human operators supply commands, and the system realizes them via executing appropriate signal plans (SPs) or green durations in the intersections. The current paper considers the first step to achieve this goal. In this paper, two models that can handle spatio-temporal graphical data are developed based on Graph Convolutional Neural Network. The developed models can be utilized either for traffic prediction tasks or for decision-making, e.g. of green times in intersections, given fixed cycle time steps. Different dataset and features are considered. In the first model, prediction of speed data is examined, while in the second model green times and speed are predicted. The large-scale urban network of Tel Aviv is considered, where data features such as speed are extracted from an array of Bluetooth sensors located at the network signalized intersections, while its signal plans represent the traffic operators’ commands. The obtained results show that: (i) including signal plan IDs and/or temporal features (month, year, day, etc.) in speed or green time duration prediction tasks can improve the performance; (ii) considering fixed cycle time steps enhances the prediction compared with non-cycle-time steps; and (iii) including Bluetooth features in green times prediction task resulted with a slight degradation in performance. • Spatio-temporal graph models improve traffic signal prediction in urban networks. • Enhanced GCN model predicts green times using signal plans as command inputs. • Signal plan IDs and temporal features boost prediction of speed and green times. • Experiment shows signal plan data and green times offer insights for autonomous traffic control.