Citation

Li J, Xie N, Zhang K, Guo F, Hu S, Chen XM. Transp. Res. C Emerg. Technol. 2022; 141: e103719.

Copyright

(Copyright © 2022, Elsevier Publishing)

DOI

10.1016/j.trc.2022.103719

PMID

unavailable

Abstract

Network traffic flow prediction on a fine-grained spatio-temporal scale is essential for intelligent transportation systems, and extensive studies have been carried out in this area. However, existing methods are mostly data-driven, with stringent requirements on the amount and quality of data. The collected network-scale traffic data are expected to be complete, sufficient, and representative, containing most traffic flow patterns in the road network. Unfortunately, it is very rare that sufficient and representative traffic data across the whole road network in several consecutive weeks are available for model calibration. In real-world applications, data insufficiency and dataset shift problems are prevalent, resulting in the 'cold start' issue in traffic prediction. To deal with the challenges above, this paper develops a two-stage physics-informed transfer learning method for network-scale link-wise traffic flow knowledge transfer under MFD-based physical constraints. In the first stage, the road network is partitioned and similar traffic regions are identified according to the physical invariants and MFD characteristics. In this way, the network-scale link-wise traffic flow pattern transfer between similar regions can be initiated under the assumption that regions with similar aggregated traffic flow patterns are more likely to share comparable link-wise traffic flow features. In the second stage, we propose our knowledge transfer architecture Deep Tensor Adaptation Network (DTAN) to bridge traffic flow knowledge in source and target regions via the parallel Siamese network structure, and further reduce domain discrepancy by imposing two distribution adaptation regularizations. A real-world traffic dataset on the urban expressway network of Beijing is used for numerical tests. The experiment results show that the proposed framework can leverage the trade-off between specific regression task performance in a single region and generalized domain adaptation capacity across multiple regions. The data insufficiency, dataset shift, and heavy computational cost problems are alleviated by improving model transferability. Finally, extensive empirical analysis is carried out to explore traffic flow pattern transferability and its relation to network traffic properties.

Language: en

Keywords

Domain adaptation; Macroscopic fundamental diagram (MFD); Network partition; Traffic prediction; Transfer learning