Speed Prediction of Multiple Traffic Scenarios with Local Fluctuation

In recent years, the significant effect of knowledge sharing between similar traffic scenarios in traffic speed prediction has received widespread attention. Existing knowledge-sharing-based researchs usually capture spatial-temporal correlation of traffic flows directly through hard or soft sharing. However, such coarse-grained sharing is not sufficient to capture fine-grained local spatial-temporal dynamics. We argue that local fluctuations in traffic flow may be caused by traffic events, weather changes and others, and implicitly reflect some specific road network structure. To this end, we propose a fine-grained knowledge sharing framework that separates local fluctuations in traffic flow so that traffic prediction modelling can learn knowledge related to the road network structure, thereby improving prediction performance. Specifically, our framework consists of temporal and spatial modules to model traffic flow information. (1) Global changes of spatial-temporal dynamics are captured by the self-decomposition module of the spatial module which is directly shared between similar traffic scenarios. (2) Local fluctuations of spatial-temporal dynamics are captured by the graph convolution layer of the spatial module, and we add parameter constraints to the graph convolution parameters, aiming at shortening their parameter differences. In this way, the fine-grained knowledge sharing is achieved. Finally, skip connections are used to converge spatial-temporal correlations for final predictions. Experimental results on two city datasets and two highway datasets show that our proposed framework achieves state-of-the-art prediction performance in terms of Mean Average Error (MAE) and Root Mean Squared Error (RMSE).