Transformer Based Spatial-Temporal Fusion Network for Metro Passenger Flow Forecasting

Passenger flow forecasting is a very critical task for the daily operations of metro system. The rapid development of deep learning methods offers us an opportunity to give an end-to-end solution to system-level prediction. However, complex spatial-temporal correlations of passenger flow data makes it quite challenging. Existing studies tend to model spatial and temporal correlations separately, which may lead to information loss and unsatisfactory prediction performance. Meanwhile, they cannot take full advantage of human knowledge and external information, such as geographical information, metro map information, etc, for modeling. To bridge the research gap, in this study, we propose a well-designed transformer based spatial-temporal fusion network (TSTFN). To cooperate with different types of external information and give additional insights, we first use multiple pre-defined graph structures to construct multi-view GCN for spatial dependence modeling. Then we propose a novel spatial-temporal synchronous self-attention layer to model spatial and temporal correlation simultaneously. Experiments show TSTFN outperforms other state-of-the-art deep learning based methods on both long-term and short-term tasks. The effectiveness of its crucial components has also been verified by using ablation study and analysis.