Physics-Guided Multi-Source Transfer Learning for Network-Scale Traffic Flow Prediction

Recent research has shown that some network traffic flow patterns are similar across multiple traffic regions. Identifying and transferring these domain-invariant features can significantly boost model accuracy and robustness, providing new insights into dealing with modeling issues like traffic data insufficiency and dataset shift. However, how to acquire transferable network traffic flow patterns from multiple traffic regions and adapt such knowledge to downstream prediction tasks of target regions remains challenging. To realize domain-invariant traffic flow pattern transfer and provide more robust prediction under insufficient data conditions, we propose a macroscopic fundamental diagram (MFD) guided transfer learning method, namely physics-guided multi-source domain adversarial network (PG-MDAN). First, an MFD similarity measure is proposed to determine what traffic flow patterns are transferable and to what extent they can be transferred. PG-MDAN embeds this physics-informed transferability measure in domain adversarial pre-training for better adaptation ability. Numerical experiments based on two real-world urban network traffic datasets show that PG-MDAN can successfully transfer recurrent and non-recurrent network traffic flow patterns from multiple regions to provide more robust and responsive prediction performance. Finally, extensive sensitivity analysis is conducted, and the results validate that applying such physical regularization can effectively avoid negative transfer and provide a flexible tool to initiate traffic flow pattern transfer in practice.