Traffic Flow Prediction with Random Walks on Graph and Spatiotemporal Bidirectional Attention Transformer

Abstract

Traffic flow prediction is crucial in intelligent transportation systems. Considering the severe disruptions caused by traffic accidents or congestion, a time series model is developed for traffic flow prediction based on multiple random walks on graphs (MRWG) and the bidirectional spatiotemporal attention mechanism (BSAM), which can adapt to both normal and exceptional situations. The MRWG mechanism is applied to capture spatial features of urban areas during traffic accidents and congestion, especially the spatial dependencies among neighboring regions. Further, a local position attention module is applied to acquire the spatial correlations between different regions to investigate their impact on the global area, while a local temporal attention module is adopted to extract short-term periodic time correlations from traffic flow data. Finally, a spatiotemporal bidirectional attention module is applied to simultaneously extract both the temporal and spatial correlations of the historical traffic flow data in order to generate the output prediction. Experiments have been conducted on NYCTaxi and NYCBike datasets with abnormal events, and the results indicate that the developed model can efficiently predict traffic flow in abnormal events, especially short-term traffic disruptions, outperforming the baseline methods under both abnormal and normal conditions.