Physics Informed Car-Following Control Model for Emission Reduction in Mixed Traffic

Abstract: This paper introduces the Physics-Informed Regularized Stacked Long Short-Term Memory (PIRS-LSTM) model, a novel approach to enhance autonomous vehicle (AV) car-following behavior while significantly reducing nitrogen oxide (NOx) emissions. By integrating NOx estimation into the LSTM loss function, the model learns to predict acceleration that minimizes emissions without compromising realistic car-following behavior. The researchers utilized both homogeneous and mixed traffic data from the OpenACC database, training the model on 18,817 seconds of AV-AV and traditional vehicle (TV)-AV trajectories and evaluating it on a test dataset of 1,881 seconds.
The PIRS-LSTM demonstrated impressive results, achieving a 14.56% reduction in NOx emissions compared to baseline models. It accurately predicted AV acceleration with a root mean square error (RMSE) of 0.0923 m/s², indicating high precision in mimicking realistic car-following behavior. Moreover, the model's inference time of 0.341 seconds suggests its potential suitability for real-time automotive applications, a crucial factor for practical implementation in AV systems.
This research offers a promising approach for developing environmentally friendly and efficient AV control systems. By simultaneously optimizing for realistic car-following behavior and reduced emissions, the PIRS-LSTM model represents a significant step forward in sustainable autonomous driving technology. The model's ability to perform in real-time conditions opens up possibilities for immediate implementation in AV systems, potentially leading to substantial reductions in vehicular emissions without sacrificing performance or safety.