Accurately estimating traffic states is a crucial aspect of transportation engineering, enabling effective traffic control and operations. In recent years, Physics-Regulated Deep Learning (PRDL) has gained significant attention due to its ability to achieve higher accuracy while requiring less training data compared to conventional deep learning (DL) approaches. However, a key challenge of PRDL is the lengthy training time required for closely related but distinct tasks.To address this limitation, this paper introduces a hybrid physics-regulated deep transfer learning approach that leverages the strengths of transfer learning, PRDL, and DL to enhance estimation accuracy and reduce computational costs, particularly in scenarios with limited observation data. The proposed framework includes two transfer learning variants designed to extract and transfer essential features from pre-trained models to new but similar traffic environments. This hybrid approach integrates deep learning training, minimizing computational overhead by eliminating physics-based loss calculations during training.Simulation results demonstrate that, compared to traditional PRDL methods, the proposed transfer learning approaches improve estimation accuracy by over 12% on average while reducing training time by more than 50% on average. These findings highlight the potential of hybrid transfer learning techniques in accelerating the adoption of PRDL for traffic state estimation, making it a valuable tool for transportation systems with limited computational resources.
Kumar et al. (Thu,) studied this question.