Abstract: ［Objective］With the citizens′ daily increasing demand for the diversification of cultural and recreational activities, the number of large-scale events held in cities is continuously increasing. The large passenger flows induced by these events can disrupt the normal operation of urban rail transit and pose safety risks. Therefore, it is necessary to extract passenger flow characteristics from historical data and propose passenger flow prediction methods to provide scientific decision-making support for rail transit operation authorities during large-scale events. ［Method］Starting from the exploration and analysis of multi-source influencing factors, information such as origin stations, venues-station distances, date types, weather types, event types, event scale, ticket sales volume, and exit passenger numbers is collected as feature variables. Considering that the amount of independent variable information available is increasing and its accuracy improves over time, a rolling data-driven XGBoost prediction model is designed. Based on passenger travel trajectories, a spatiotemporal distribution analysis method for large passenger flows during the crowd dispersal of large-scale events is proposed. The accuracy of the proposed prediction method is verified using the case of a large-scale event held at a venue close to station in a certain city. ［Result & Conclusion］The proposed passenger flow prediction method achieves high prediction accuracy at the OD (origin-destination) level, line level, and entry volume level. It can provide a decision-making basis for station passenger flow control and network operation adjustments.