Predicting passenger flow in the city bus system across the COVID-19 pandemic

Abstract

Accurate passenger flow prediction is essential for urban traffic management. Existing studies either assume stationary mobility patterns or redesign models specifically for pandemic periods, limiting their robustness in different situations. This study proposes a unified deep learning framework called the Prediction of Passenger Flow (PPF) model, which can operate continuously during both pandemic and non-pandemic periods without structural modifications. The proposed architecture consists of two complementary modules. First, the Flow module employs residual fully convolutional networks with a novel heuristic layout of districts to capture recent and periodic temporal dependencies of passenger flow in the study area. Second, a District module integrates contextual and pandemic-related features through spatial residual learning, and a parametric fusion mechanism then adaptively balances the influences of all factors. Extensive experiments conducted on a real-world bus ridership dataset demonstrate that the framework maintains stable performance despite drastic mobility fluctuations during severe outbreak periods, and consistently achieves low RMSE, MAE, and MAPE across different pandemic stages. The results validate that combining residual spatial modeling with pandemic-aware contextual encoding provides a robust and generalizable solution for urban passenger flow forecasting.