Integrated Prediction of Regional Traffic Situation Based on Multi-Task Spatial-Temporal Network

Some recent evidence suggests that there is a certain correlation between parking saturation and traffic flow in a region, and the trends of the two are affected by changes in both themselves and the other. The traditional single-task prediction often neglects this point, which becomes a bottleneck for further improvement of the prediction accuracy. In this paper, we propose a multi-task spatial-temporal prediction model, which uses multi-channel graph convolutional network (GCN) to fuse the spatial features of the traffic network and the parking network, and then extracts the joint temporal features from the fused spatial features through gated recurrent unit (GRU), so as to realize the integrated and simultaneous prediction of traffic flow and parking saturation. The experiment results show that the multi-task prediction is better than the single-task prediction in terms of accuracy, especially when road traffic and parking interact with each other more closely. Through the experiment, the influence of the correlation between traffic flow and parking saturation on the prediction accuracy is observed for the first time.