关键词: 智能交通, 粒流协同控制, 驾驭工程, 智能网联车辆, 集疏运道路作业区, 临时集卡专用道

Abstract: This study proposes a Harness Engineering-based large language model approach for granular-flow cooperative control. Connected and automated passenger cars (CAV-Cs) are treated as controllable particles to guide the traffic flow upstream of a temporary exclusive container truck lane (ECTL) in a road work zone, thereby creating the stable operating conditions for human driven trucks and connected and automated truck platoons (CAV-TPs) to merge into the ECTL. First, to address the heterogeneity of traffic information and context loss in the continuous decision-making process of large language models, a structured representation and encapsulation module, including vehicle position, speed, merging demand, and collision risk, is developed for key information to support the generation of continuous control decision. Second, to address the potential infeasibility and insufficient safety constraints of actions generated by the large language model, a Harness Engineering safety-checking module is constructed to verify lane legality, speed boundaries, safe gaps, and collision risk constraints for high-level cooperative control intentions. Finally, the verified high-level actions are transformed into executable commands, including lane changing, acceleration, deceleration, and maintaining. Taking the port collector-distributor road work zone of the Donghai Bridge in Shanghai as the study site, a mixed-traffic simulation scenario was developed by using the open-source traffic simulation platform SUMO to evaluate the proposed method. The simulation results show that, compared with the uncontrolled scenario, the proposed method reduces the total delay cost in the work zone by 11.4%~21.82%, which is lower than that from the dynamic speed control in most scenarios. The numbers of controlled vehicles and control actions, as well as memory usage, are 14.0%~19.8%, 6.5%~ 9.6%, and 14.7%~23.0% of those required by dynamic speed control, respectively, which indicate that the proposed method requires a lower intervention scale and fewer computational resources. Sensitivity analysis shows that when the penetration rates of CAV-TP and CAV-C are 75% and 40%, respectively, a relatively favorable balance between delay cost and collision risk can be achieved.

Key words: intelligent transportation, granular flow cooperative control, Harness Engineering, connected and automated vehicles, port collector-distributor roads work zones, temporary exclusive container truck lane