Spatiotemporal attention based multi-graph convolutional network for passenger congestion delay short-term prediction

The development of bus rapid transit (BRT) is currently in a rising phase, and at the same time, the demand for resident travel is continuously increasing. During peak periods, a large volume of commuter traffic converges at stations in a short time. Due to the limited carrying capacity of vehicles, some passengers are forced to remain on platforms waiting for the next or even more buses. This delays the overall travel time and causes crowding. Precisely predicting the passenger congestion in advance has become an urgent problem to solve. Addressing the issue that existing studies do not adequately consider spatial correlations, temporal correlations, and external factors, which leads to low prediction accuracy, a spatiotemporal attention multi-graph convolutional network (STA-MGCN) is proposed for short-term forecasting of passenger congestion delays. By establishing a spatiotemporal attention mechanism, the model adjusts the input data to effectively capture the dynamic spatiotemporal correlations in traffic data. It encodes spatial and semantic correlations and models them as multi-location graphs, which are then fused through a multi-graph convolutional network. A bidirectional recurrent layer module considers the backward and forward states of congestion delay time series, ultimately fusing multi-source data to generate the final prediction results. Taking the Xiamen BRT network and Xiamen subway network passenger congestion delay data as examples, experimental results show that the proposed STA-MGCN method has higher prediction accuracy compared to baseline models.