Multidirectional short-term traffic volume prediction based on spatiotemporal networks

Traffic volume (i.e., the volumes of cars, bikes and crowds) demonstrates real-time traffic hotspots and city transportation movement patterns, whose prediction is of great significance in traffic management and urban construction. The accurate prediction of traffic volume, however, is a challenging problem, and is essentially a spatiotemporal phenomenon affected by complex factors. In this paper, in addition to the spatiotemporal correlations between adjacent space and time, three significant traffic volume prediction observations are concluded: (1) variations in traffic volume in different directions have different regularities; (2) point of interests (POIs) have great impact on the variations of traffic volume in the spatial dimension; and (3) the rule of daily changes significantly affects short-term traffic volume in the temporal dimension. Inspired by the above observations, a novel multidirectional spatiotemporal network (MDSTN) is proposed for short-term traffic volume prediction. Specifically, traffic volume is divided into four directions, i.e., heading upwards, leftwards, downwards and rightwards, and all further analyses are performed in four directions. Meanwhile, traffic volume is combined with POI signals and time signals to augment the spatial and temporal characteristics. In particular, a POI attention mechanism is designed to adaptively extract the contribution of different POIs to traffic volume, and 3D convolutions are adopted to address the temporal variations in traffic volume. In addition, residual units are employed to eliminate gradient vanishing and exploding caused by the network depth increment, and batch normalization is adopted to adaptively remove the correlation between parallel neural units. Comparison experiments with real-life electric bicycle trajectories and the NYCBike dataset demonstrate that MDSTN vastly outperforms all baselines.