Abstract: As urbanization accelerates, the metro has become a crucial part of public transport, essential for alleviating urban traffic pressure and enhancing transportation environments. Accurate prediction of metro passenger flow is significantly valuable for metro operation management, resource optimization, and emergency response. Due to the complex correlations between metro stations and lines, coupled with the large-scale and high-dimensional nature of metro operational data, existing methods, while considering multiple stations and lines, fail to adequately capture their interactions and dependencies in terms of passenger mobility patterns, time, and spatial distributions, resulting in limited prediction accuracy. Therefore, this paper proposes a new model, the Spatio-Temporal Characteristics Integrated Network (STIN) for Metro Passenger Flow, to comprehensively capture the complex correlations between stations and lines in the metro system from a multi-dimensional and multi-level perspective, aiming to improve the accuracy of short-term prediction of metro OD passenger flow. STIN harnesses physical network graphs, similarity network graphs, behavioral relevance network graphs, and both basic and advanced hypergraphs to capture the complex relationships among metro stations comprehensively. The model particularly emphasizes the utilization of hypergraph convolution to consider higher-order dependencies between stations, providing a new perspective on passenger flow patterns and station-to-station dynamics. Tested on a dataset from Hangzhou's metro system over 27 days in 2019, involving three metro lines and 80 stations, STIN demonstrated superior prediction accuracy compared to traditional models such as GCN, STGCN, and ASTGCN. The efficacy of STIN is attributed to its ability to effectively process complex and dynamic metro network data, ensuring high accuracy in real-time passenger flow predictions. This model can be extensively applied to urban traffic management, passenger flow forecasting, and optimization of metro operation schedules, significantly benefiting urban transport systems. While the current implementation of STIN relies on data from a specific region, its adaptability across different metro systems remains to be tested. Future research will aim to enhance the model's generalizability by incorporating diverse datasets and considering additional factors such as weather conditions, holidays, and seasonal variations. This work highlights the potential of deep learning techniques in advancing metro passenger flow prediction and underscores the importance of choosing the right model architecture and tuning parameters for time-series forecasting tasks.

Key words: Metro Passenger Flow Prediction, Deep Learning, Spatio-temporal Data Analysis, Hypergraph Convolutional Networks