Multidynamic Temporal Representation Graph Convolutional Network for Traffic Flow Prediction

1. We propose a novel traffic flow prediction model (MDTRGCN), a dynamic spatial dependency learning approach that propagates node hidden states on the basis of dynamic spatial relationships to capture dynamic spatiotemporal features, enabling effective long-term predictions.2. We construct a dynamic graph builder and dynamic graph convolutions, and through the multidimensional fusion module, we integrate auxiliary hidden states with the main hidden states in both spatial and temporal dimensions to uncover dynamic spatiotemporal relationships.3. We design a temporal representation learning method that pretrains through a masked reconstruction task to obtain compressed and contextual temporal representations of subsequences, capturing periodic features in long historical sequences.4. Extensive experimental results on two real-world datasets demonstrate that the MDTRGCN outperforms baseline methods in terms of prediction accuracy.