Arrival information-guided spatiotemporal prediction of transportation hub passenger distribution

Abstract

Understanding the spatiotemporal distribution of hub passenger flow is essential for optimizing both hub and urban transportation operations. However, predicting spatiotemporal distribution of transportation hub passenger flow encounters is challenging due to complex factors influencing its dynamics. This paper proposes a deep learning model, the Deep Spatiotemporal Graph Attention Network (DSTGAT), to predict the spatiotemporal distribution of hub passenger flow in urban areas. The DSTGAT consists of two modules: a spatiotemporal passenger flow prediction module and a passenger flow correction module. The spatiotemporal prediction module integrates Graph Attention Networks (GATs) and Gated Recurrent Units (GRUs) to capture the spatial and temporal dependencies in passenger flow, considering factors such as land function, adjacency, distance to the hub, and weather conditions. The passenger flow correction module uses Dynamic Time Warping (DTW) to identify the similarity of historical arrival passenger flows. Based on this similarity, it selects the most similar passenger flow distribution for prediction correction. A case study using data from Beijing Daxing International Airport in China demonstrates the superior performance of the DSTGAT compared to baseline models. The model exhibits robust predictive accuracy, particularly in regions with high passenger flow fluctuations and during holiday periods. The study highlights the importance of considering external factors and arrival passenger flow in achieving accurate hub passenger flow predictions.