Enhancing Spatial-Temporal Awareness via Graph Convolutional Networks and Transformers for Traffic Flow Forecasting

Deep learning has significantly advanced time series forecasting, leading to numerous new methods. In the context of time series prediction problems, traffic flow prediction is particularly representative due to the high nonlinearity and variability. Existing traffic flow prediction methods often lack the extraction of spatial-temporal topological features at different scales. To address the aforementioned issues, we propose a novel framework that combines periodic spatial-temporal awareness with graph convolutional networks and transformer (STA-GCNT). Firstly, the model can find the most important topological feature in the data, the persistent homology of the data. Secondly, the nodes are helped to obtain a larger field of perception through capturing the cross-space correlation between nodes in different spatial-temporal graphs. By focusing on the remaining time steps within the sliding window at each time step, a long-range bidirectional time dependency across multiple time steps is established. Finally, we aggregate all the features from different scales for fusion. Our results demonstrate that STA-GCNT outperforms 8 state-of-the-art methods on three time series datasets.