Extended Multi-Component Gated Recurrent Graph Convolutional Network for Traffic Flow Prediction

Traffic flow prediction is a difficult undertaking in transportation systems, due to the intricate periodicity and real-time dynamics for traffic data, spatial-temporal dependency for road networks, existing prediction approaches fail to yield satisfactory results. We propose a traffic flow prediction method named Extended Multi-component External Interactive Gated Recurrent Graph Convolutional Network (EMGRGCN). The extended multi-component (EMC) module is incorporated into the prediction model to address the periodic temporal diffusion problem. Then, we introduce an encoder-decoder architecture that incorporates attention mechanism to capture spatial-temporal dependencies. Specifically, an External Interactive Gated Recurrent Unit (EIGRU) is utilized to capture crucial temporal features. EIGRU and graph convolutional network are combined in the encoder to extract spatial-temporal correlation, and EIGRU and convolutional neural network based decoder transforms the spatial-temporal characteristics into a sequence to predict future traffic flows. Experiments on public transportation datasets PEMSD8 and PEMSD4 demonstrate that EMGRGCN model achieves the best performance.