Predicting Regional-Level Bus Stop Passenger Flow with a Multigraph Fusion Spatio-Temporal Graph Attention Network

Abstract

Bus stop passenger flow prediction is a key component in developing urban intelligent transit systems. Bus line networks are usually complex nonlinear time-varying systems, which poses a challenge for constructing the spatio-temporal correlations of passenger flows at bus stops in a region. A multigraph fusion spatio-temporal graph attention network (MF_STGAT) is proposed, which uses the spatio-temporal correlation of each bus stop to achieve regional-level passenger flow prediction. First, MF_STGAT enhances the ability to capture spatio-temporal correlations of bus stops by extracting key frames of bus passenger flow time sequence data and encoding the region into two maps (geographic feature map and functional similarity map). Second, a GAT-based graph fusion method is constructed to obtain the spatial feature of bus stops. The spatial feature information is encoded into vectors, and long short-term memory (LSTM) is used to predict the bus passenger flow at each stop in the region. Finally, the IC card data of 12 main bus lines in Beijing, China, are selected to evaluate the model. The results show that MF_STGAT outperforms baseline models. In addition, we further discuss the extent of the improvement in the accuracy of the model prediction through the spatio-temporal modeling approach and the robustness of MF_STGAT. Moreover, we output the attention weights of the model to enhance the interpretability of the spatio-temporal distribution characteristics of the bus passenger flow.