Investigating Knowledge Transfer in Residual Physics-Informed Neural Networks Using Connected Vehicles Traffic Data

Accurate traffic state estimation is essential for effective traffic management and control in intelligent transportation systems. Traditional data-driven approaches often require large amounts of training data and may not fully capture the underlying physical dynamics of traffic flow. Physics-Informed Neural Networks (PINN) have emerged as a promising solution, incorporating the governing physical laws into the neural network training process. However, PINN models face challenges such as slow convergence and limited data availability in real-world scenarios. To address these issues, this paper presents a novel Residual Physics-Informed Neural Network (Res-PINN) architecture for traffic state estimation using connected vehicle data. The Res-PINN model incorporates residual connections to improve information propagation and convergence speed, enhancing the training process and model performance. Furthermore, a transfer learning approach is explored to leverage knowledge from data-rich scenarios and improve the performance of Res-PINN in situations with limited data availability. The proposed Res-PINN model and transfer learning approach are evaluated using real-world connected vehicle data from a major vehicle telematics aggregator, focusing on traffic speed estimation on two freeway corridors in Melbourne, Australia. The results demonstrate the superiority of the Res-PINN model over traditional PINN models and highlight the effectiveness of transfer learning in enhancing traffic state estimation accuracy. The proposed framework advances the field of traffic state estimation by providing accurate and reliable methods, enabling better traffic management and control strategies for more efficient transportation systems.