Dynamic Graph Convolution and Interaction Network for Traffic Flow Forecasting

The forecasting of traffic flow is crucial for optimizing urban traffic operations, improving public transport services, and reducing environmental pollution. However, due to the complex correlation and nonlinearity of traffic flow over time and space, accurately capturing this dynamic spatio-temporal dependence could be challenging. Most existing approaches could be struggling to fully understand the impact of real-time interactions between different geographic locations and to integrate dynamic data over long time scales effectively. To address these challenges, we propose an efficient Dynamic Graph Convolution and Interaction Network (DGCINet). This method enables the simultaneous capture of temporal and spatial dependencies by embedding the graph convolution network into an interactive learning structure to achieve effective long-term traffic flow forecasting. We also leverage a novel dynamic graph convolution method, using merged real-time generated graphs from adaptive and learnable adjacency matrix, to capture the spatial correlation of real-time changes in traffic networks. Furthermore, we have integrated a spatio-temporal adaptive Transformer that could extract global and local features simultaneously. On four real traffic flow datasets, DGCINet's prediction performance is significantly better than the other nine baseline methods, improving the average prediction precision by 6.3%.