A Physics-Informed LSTM Car-Following Model for AVs: Minimizing Emissions in Mixed Traffic

The growing integration of autonomous vehicles (AVs) equipped with adaptive cruise control (ACC) into mixed traffic environments brings about both advantages and challenges. While AVs offer safety and efficiency benefits, their presence can introduce traffic flow instability, potentially leading to increased emissions. Nonetheless, limited research exists on the development of AV car-following models with a specific focus on emission reduction. This study introduces a novel car-following model tailored for AVs in mixed traffic scenarios, harnessing the power of a physics-informed Long Short-Term Memory (LSTM) deep neural network. The model optimizes a multi-objective loss function, with a dual mission: to accurately mimic the real-world positions of AVs over time and to reduce emissions effectively. In pursuit of the latter goal, the model integrates the Vehicle-Specific Power (VSP), as a proxy of emissions, into its physics-based equations. This study utilizes the OpenACC dataset. A robust training dataset is drawn from OpenACC encompassing 18,817 seconds of AV-AV interactions, and traditional vehicle-AV (TV-AV) interactions. The model performance is evaluated on an independent test dataset spanning 1,881 seconds. Comparative analysis against baseline models, including basic LSTM and ACC models, highlights the superior performance of the proposed model. In particular, the Root Mean Square Error (RMSE) and Symmetric Mean Absolute Percentage Error (SMAPE) performances of the model are remarkable; underscoring its capacity to capture intricate temporal dependencies, elevate car-following dynamics, and, most importantly, reduce emissions in the evolving landscape of mixed traffic scenarios.