Tfgcn: a time-varying fuzzy graph convolutional network for multi-sensor traffic flow forecasting

Traffic prediction is a pivotal component of intelligent transportation systems (ITS), which can provide effective support for traffic planning and management. Recently, graph convolutional networks (GCNs) have been proposed to model intricate spatio-temporal correlations. However, most GCNs use static graphs, which fail to capture dynamic spatial correlations due to sensor damage. A few studies based on dynamic graph neural networks can model such dynamics but struggle to capture long-term spatio-temporal dependencies because they mainly focus on local and short-term changes in the graph. To overcome these limitations, we propose a time-varying fuzzy graph convolutional network called TFGCN that combines dynamic and static graphs to predict multi-sensor traffic flow. TFGCN uses a gated fuzzy graph to model long-term dynamic spatial correlations adaptively. It also employs a periodic coupled Transformer network that integrates monthly and weekly periodic data to capture global temporal trend information. Extensive experiments conducted on two real-world datasets demonstrate that our proposed model outperforms several state-of-the-art baselines.