Dynamic Spatio-Temporal Graph Fusion Network modeling for urban metro ridership prediction

Predicting urban metro ridership holds significant practical value for optimizing operational scheduling and guiding individual travel planning. Understanding the complexity of metro ridership influenced by temporal and spatial factors and modeling the intrinsic correlation between station inflow and outflow are current research focuses and challenges. To address these issues, we develop a Dynamic Spatio-Temporal Graph Fusion Network (DSTGFN) to comprehensively extract and exploit spatio-temporal features. We first use non-negative tensor decomposition technique to construct the dynamic virtual graph, revealing deep patterns of interaction between metro inflows and outflows. Then, to overcome the limitations of stacked designs in processing spatio-temporal features, we adopt a multimodule parallel learning design to construct four feature extraction modules: temporal, static spatial, dynamic spatial, and dynamic spatio-temporal modules. Each module is dedicated to refining key features from the fluctuations of metro ridership. Finally, we establish a feature fusion module to compress the outputs of four modules into three aspects: static spatio-temporal, static and dynamic spatial, and dynamic spatio-temporal features, and further fuse them through a convolutional layer and a fully connected layer to predict the inflow and outflow of passengers at each station across the entire metro network. Empirical validation using Hangzhou and Shanghai metro datasets demonstrates the significant advantages of DSTGFN in terms of prediction accuracy, robustness and overall performance. Hence, our model can serve as a data-driven tool for metro management authorities, enabling them to make timely and effective dynamic adjustments to ensure optimal levels of metro services.