A Cross-City Traffic Flow Prediction Framework Incorporating Holidays and Weather Factors

Although graph neural networks are widely used in urban traffic flow prediction research, they often face challenges related to data scarcity. Cross-city transfer learning presents a promising and cost-effective solution to this problem. This study introduces TransGTR-MCA (Meteorological and Calendrical Aware), an advanced model for cross-city traffic flow prediction based on the TransGTR framework. TransGTR-MCA utilizes a Graph Structure Learning framework that creates adaptive graph structures dynamically, without depending on predefined adjacency matrices. This method enhances the ability to capture complex spatial dependencies, especially in cross-city contexts. A key innovation of the model is its incorporation of external factors like weather and holidays, which allows for a thorough consideration of the unique climatic and cultural traits of different cities. This integration aims to improve the transfer learning process across various urban settings. To assess the model's effectiveness, experiments were carried out using three benchmark datasets for traffic flow prediction: METR-LA, PEMSD7M, and SZ-TAXI. The results demonstrate that the TransGTR-MCA model outperforms others in short-term predictions. Specifically, when transferring knowledge from the U.S. METR-LA dataset to the Chinese SZ-TAXI dataset, the model achieved a notable 2.73% reduction in error for 5-minute predictions, highlighting its potential for cross-cultural urban traffic forecasting.