Cross-City Pretraining Transfer Learning Model for Traffic Flow Prediction

Accurate traffic flow prediction plays a pivotal role in intelligent transportation systems (ITS). While deep learning-based approaches have demonstrated remarkable success in this domain, their performance heavily depends on the availability of large-scale training data. However, many cities face challenges in collecting sufficient traffic flow data due to privacy concerns and substantial storage requirements. Consequently, conventional traffic flow prediction models often suffer from performance degradation when applied to cities with limited data availability, primarily due to spatially unbalanced data distributions. To overcome this limitation, we propose a novel pre-trained framework for cross-city traffic flow prediction, termed PTCC. Different from existing methods that focus solely on optimizing performance for data-rich cities, our framework innovatively transfers spatiotemporal knowledge from data-abundant cities to enhance prediction accuracy in data-scarce scenarios. The proposed PTCC framework comprises three key components: 1) A pre-trained module that learns long-term temporal patterns from traffic flow data in source cities and generates comprehensive segment-level representations; 2) A discrete graph learning structure that captures node dependencies from contextual segment-level representations; 3) A spatiotemporal prediction module that effectively transfers the acquired knowledge to facilitate accurate traffic flow forecasting in target cities. We conduct extensive experiments to validate the framework’s effectiveness, training the model on METR-LA and PEMS-BAY datasets, and evaluating its performance on PEMS04 and PEMS08 datasets. The experimental results demonstrate that our pre-trained frame-work significantly outperforms existing methods, establishing its superiority for traffic flow prediction in cities with limited data availability.