A traffic prediction framework under extreme weather combined disaster knowledge and deep learning

Studying the spatiotemporal patterns of urban road traffic under extreme weather is a key step to building a climate-resilient city. Although existing researches model and simulate traffic states from different perspectives, the traffic forecasting of the urban road networks under extreme weather is seldom addressed. In this paper, a novel Knowledge-driven Attribute-augmented Attention Spatiotemporal Graph Convolutional Network framework is proposed to predict urban road traffic under wind and rain especially in tropical cyclone disasters. Considering the disaster conditions, we model the external dynamic hazard attributes and static environment attributes, and designed an attribute-augmented unit to encode and integrate these factors into the deep learning model. The model is combined with the graph convolutional network (GCN), the gated recurrent unit (GRU), and the attention mechanism. Experiments demonstrate that the predictability of traffic speed can be greatly increased by supplementing the disaster-related factors, the prediction accuracy reaches 0.79. The proposed approach outperforms baselines by 12.16%-31.67% on real-world Shenzhen’s traffic datasets. The model also performs robustly on different road vulnerabilities and hazard intensities. The model errors are mainly occurred in the early peak with extreme wind and rain and the coastal area in the southeast of Shenzhen because of the greater uncertainty. The framework and findings provide a valuable reference for the decision-making of traffic management and control prior to a disaster to alleviate traffic congestion and reduce the negative impact of disasters.