Modeling Crowdedness at Public Transport Stations During Special Events: A Comparative Study of Eleven Cities

Special events significantly disrupt normal crowding patterns in urban public transport (PT) systems, yet our understanding of these impacts remains limited. Accurate prediction of crowdedness during special events could enhance PT passenger flow forecasting when integrated into unified frameworks, ultimately improving operations and crowd management—critical components of PT demand management. This study models PT station crowding patterns during planned special events using publicly available opportunistic data, specifically leveraging the Google Popular Times (GPT) popularity index as a proxy for station-level crowdedness. The extensive spatiotemporal coverage of GPT data in urban areas provides a high-quality data source for modeling station crowdedness and inferring trip generation and attraction capabilities of the surrounding areas. To capture the temporal dynamics specific to special events, we propose a time interval index that encodes event-driven variations in crowdedness. Recognizing PT networks’ natural graph structure, we developed a graph neural network augmented with an attention mechanism and a positional embedding-enhanced temporal convolutional network model, named APT-GCN, for station crowdedness prediction. We evaluated our model using football matches as case studies, conducting comparative analyses across 15 clubs in 11 cities. The results demonstrated the model’s precision in forecasting station crowdedness, with the event indicator substantially improving prediction performance during special events.