STFGCN: Spatio-Temporal Fusion Graph Convolutional Networks for Subway Traffic Prediction

Abstract

Metro passenger flow prediction is a crucial and challenging task in the intelligent transportation system of subways. It serves as the foundation for achieving intelligent transportation in subway systems and holds significant importance in practical applications. Although much progress has been made, accurate traffic flow prediction still faces challenges. To address this, we propose the Spatio-Temporal Fusion Graph Convolutional Network (STFGCN) for predicting metro passenger flow. Specifically, we employ Discrete Cosine Transform (DCT) to replace the Fourier Transform used in traditional models, which avoids the Gibbs phenomenon associated with the Fourier Transform. We combine DCT with channel attention to form periodic trend-enhancing attention, thereby enhancing the expressive power of the model. Furthermore, we introduce trend similarity-aware attention to capture the evolutionary trends of time series and adopt a dynamic correlation graph convolutional network to dynamically adjust spatial correlation strengths based on changes in different time periods. Experimental results on the Hangzhou Metro’s inbound and outbound passenger flow datasets demonstrate that the STFGCN model exhibits significant superiority over baseline models and shows excellent performance in metro passenger flow prediction. Compared to the CorrSTN model, STFGCN achieves improvements of 22.15%, 16.9%, and 0.6% in MAE, RMSE, and MAPE, respectively.