SLG-Transformer: An Enhanced Simulation and Optimization Model for Low-Penetration Mixed Traffic

With the rapid development of Intelligent Connected Vehicles (ICVs) and autonomous driving technologies, traffic systems are transitioning from human-centered driving to human–vehicle coexistence. In low-penetration scenarios, traditional traffic simulation models, such as IDM, Krauss, MOBIL, and LC2013, face limitations in capturing heterogeneous interactions between autonomous vehicles (AVs) and human-driven vehicles (HVs) and in achieving system-level multi-objective optimization. To address these challenges, this study proposes an enhanced simulation model, SLG-Transformer, which integrates sparse perception, hierarchical interaction, and global coordination modules. The framework enables vehicle-level adaptive parameter tuning and system-level multi-objective optimization, and seamlessly interfaces with the SUMO simulation platform to form a closed-loop simulation–optimization mechanism. Experimental results demonstrate that SLG-Transformer outperforms classical rule-based and deep reinforcement learning baseline models in average speed, average waiting time, number of stops, and CO₂ emissions across different AV penetration rates, with particularly significant improvements under low- to medium-penetration conditions. In terms of average speed, SLG-Transformer outperforms the best traditional model by 6.9% at 10% penetration, and reduces the waiting time to 2.5 s at 20% penetration. Ablation studies further verify the critical roles of sparse perception, local interaction, and global coordination in enhancing traffic efficiency and reducing environmental impact. This work provides a novel methodology for mixed-traffic simulation and intelligent traffic optimization under low penetration.