Traffic State Estimation Using Hybrid Physics-Regulated Deep Transfer Learning Approach

Estimating traffic states efficiently and accurately is a fundamental problem in transportation engineering for traffic control and operation. In recent years, there has been a growth of interest in using physics-regulated deep learning (PRDL) to tackle such a problem because they provide higher accuracy and do not require extensive training data compared to traditional deep learning (DL) approaches. However, one of the limitations of using PRDL is the long time associated with the training process for different but closely linked tasks. In order to reduce the training time and improve the estimation accuracy with inadequate observation data for traffic state estimation (TSE), this paper presents a hybrid physics-regulated deep transfer learning approach that complements the advantages of transfer learning, PRDL, and DL. Two variants are also presented under the proposed general framework. These transfer learning approaches capture the meaningful insights of general features obtained in the trained models and transfer them to the new but similar environment models. The hybrid architecture with DL training further reduces the computational expense by eliminating the calculations and training associated with physics losses. Computer simulation results show that compared with the existing PRDL approach, the proposed transfer learning approaches significantly improve the estimation accuracy by more than 12% on average and reduce the training time by more than 50% on average. These promising results showcase the capability of the proposed hybrid transfer learning approaches in expediting the application of PRDL in TSE for transportation communities with limited resources.