| 195 | 0 | 65 |
| 下载次数 | 被引频次 | 阅读次数 |
【背景】传统交通信号控制方法无法有效应对交通事故导致的交叉口车辆积压与通行效率骤降问题。【目标】提出一种新的交通信号控制方法提升事故场景下单交叉口的通行效率。【方法】提出一种融合大语言模型探索（LLME）与混合深度Q网络（HDQN）的交通信号控制方法自适应地调整交叉口的相序和相位绿灯时长，提升事故车道和关联车道的通行效率。LLME使用大语言模型根据事故交叉口交通状态推理生成包含相序及相位绿灯时长的混合动作，并与环境交互获取大模型探索经验用于HDQN的前期训练；HDQN在神经网络中引入相位权重及相位绿灯时长分支，通过分组卷积获取混合动作价值，进而输出最优混合动作。【结果】基于SUMO仿真软件构建单交叉口事故场景对所提方法进行测试，实验结果表明：相比于基线方法，应用所提方法时，交叉口平均排队车辆数下降15.99%、车辆平均行驶速度提升11.17%、车辆平均通行时间下降8.32%、车辆平均等待时间下降12.01%、事故车道排队车辆数下降19.62%。另外，在真实单交叉口事故场景下的测试结果表明，所提方法的各项评估指标均优于基线方法。【结论】基于LLME-HDQN的交通信号控制方法能有效提升事故场景下单交叉口的通行效率。
Abstract：[Background] Traditional traffic signal control methods struggle to effectively address vehicle backlog and sudden drops in traffic efficiency at intersections caused by accidents. [Objective] Proposes a novel traffic signal control method to improve the traffic efficiency of a single intersection under accident scenarios. [Method] A traffic signal control method integrating Large Language Model-based Exploration (LLME) and Hybrid Deep Q-Network (HDQN) is proposed to adaptively adjust the phase sequence and green light duration of an intersection， thereby enhancing the traffic efficiency of both accident-affected lanes and related lanes. LLME uses a large language model to infer the traffic conditions at accident-prone intersections， generates hybrid actions specifying both phase sequence and green phase duration， and then interacts with the environment to collect exploratory experience for the preliminary training of the HDQN; HDQN introduces phase weight and green light duration branches into the neural network， extracts hybrid action values through group convolution， and outputs the optimal hybrid action. [Result] The proposed method is tested using a single-intersection accident scenario built in SUMO simulation software. Experimental results show that， compared to baseline methods， the proposed method reduces the average number of queuing vehicles at the intersection by 15.99%， increases the average vehicle speed by 11.17%， reduces the average travel time by 8.32%， and shortens the average waiting time by 12.01%， decreases the number of queuing vehicles on the accident-affected road by 19.62%. Additionally， tests in a real single-intersection accident scenario demonstrate that the proposed method outperforms baseline methods across all evaluation metrics. [Conclusion] The LLME-HDQN based traffic signal control method can effectively improve the traffic efficiency of a single intersection in accident scenarios.
[1] 杜英翠. 城市交叉口进口道交通事故辨识与应急管控方法[D]. 山东理工大学， 2023.
[2] 李辉， 张昱， 韩海龙， 等. 城市道路交叉口交通事故特性分析及信号控制优化设计[J]. 市政技术， 2024， 42(9): 82-89.
[3] 李瑞敏， 唐瑾. 过饱和交叉口交通信号控制动态规划优化模型[J]. 交通运输工程学报， 2015， 15(6): 101-109.
[4] 卢凯， 汪丽， 吴蔚， 等. 信号控制下路段交通承载能力计算及交叉口配时优化模型[J]. 东南大学学报(自然科学版)， 2021， 51(2): 333-341.
[5] WANG J， HANG J， ZHOU X. Signal timing optimization model for intersections in traffic incidents[J]. Journal of Advanced Transportation， 2020， 2020: 1081365.
[6] ZHANG H， GUO S， LONG X， et al. Combined dynamic route guidance and signal timing optimization for urban traffic congestion caused by accidents[J]. Journal of Advanced Transportation， 2023， 2023: 5858614.
[7] ZEINALY Z， SOJOODI M， BOLOUKI S. A resilient intelligent traffic signal control scheme for accident scenario at intersections via deep reinforcement learning[J]. Sustainability， 2023， 15(2): 1329.
[8] ROW T C， LEE S C， CHENG J F. Designing traffic light signals to ease congestion caused by traffic accidents using timed Petri net[J]. Applied Sciences， 2024， 14(23): 11260.
[9] 徐东伟， 周磊， 王达， 等. 基于深度强化学习的城市交通信号控制综述[J]. 交通运输工程与信息学报， 2022， 20(1): 15-30.
[10] 刘智敏， 叶宝林， 朱耀东， 等. 基于深度强化学习的交通信号控制方法[J]. 浙江大学学报(工学版)， 2022， 56(6): 1249-1256.
[11] 叶宝林， 孙瑞涛， 吴维敏， 等. 基于异步优势演员-评论家的交通信号控制方法[J]. 浙江大学学报(工学版)， 2024， 58(8): 1671-1680， 1703.
[12] 叶宝林， 孙瑞涛， 李灵犀， 等. 基于融合状态预测的深度强化学习A2C的交通信号控制[J]. 计算机工程， 2025， 51(5): 33-42.
[13] 叶宝林， 戴本岙， 苗永超， 等. 基于自适应时空同步建模的交通流预测[J]. 计量学报， 2025， 46(6): 802-812.
[14] WEI H， CHEN C， ZHENG G， et al. PressLight: learning max pressure control to coordinate traffic signals in arterial network[C]//Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. Anchorage AK USA. ACM， 2019: 1290-1298.
[15] BOUKTIF S， CHENIKI A， OUNI A， et al. Parameterized-Action Based Deep Reinforcement Learning for Intelligent Traffic Signal Control[J]. Engineering Applications of Artificial Intelligence， 2025， 159(1)， 111-123.
[16] LUO H， BIE Y， JIN S. Reinforcement learning for traffic signal control in hybrid action space[J]. IEEE Transactions on Intelligent Transportation Systems， 2024， 25(6): 5225-5241.
[17] YE B L， WU P， LI L， et al. Multi-intersection traffic signal control based on dynamic spatiotemporal memory enhanced learning[J]. Control Engineering Practice， 2025， 165: 106606.
[18] YE B L， WU P， LI L， et al. Deep reinforcement learning based on spatiotemporal information for network-wide traffic signal coordination control[J]. IEEE Open Journal of Intelligent Transportation Systems， 2025， 6: 1424-1438.
[19] 贺正冰. 大语言模型在道路交通领域应用: 创新与挑战[J]. 交通运输工程与信息学报， 2025， 23(1): 85-92.
[20] MOVAHEDI M， CHOI J. The crossroads of LLM and traffic control: a study on large language models in adaptive traffic signal control[J]. IEEE Transactions on Intelligent Transportation Systems， 2025， 26(2): 1701-1716.
[21] ZOU X C， YANG Y H， CHEN Z et al. Traffic-R1: reinforced LLMs bring human-like reasoning to traffic signal control systems[EB/OL]. (2025-10-22)[2025-12-19]. https://arxiv.org/abs/2508.02344.
[22] GUO Q， LI X H， CHEN J Y， et al. A dual large language models architecture with herald guided prompts for parallel fine grained traffic signal control[EB/OL]. (2025-10-31)[2025-12-19]. https://arxiv.org/abs/2511.00136.
[23] YUAN Z R， LAI S Q， LIU H， et al. CoLLMLight: cooperative large language model agents for network-wide traffic signal control[EB/OL]. (2025-03-14)[2025-11-26]. https://arxiv.org/abs/2503.11739.
[24] PANG A Y， WANG M N， PUN M O， et al. iLLM-TSC: Integration reinforcement learning and large language model for traffic signal control policy improvement[EB/OL]. (2024-07-08)[2025-11-26]. https://arxiv.org/abs/2407.06025.
[25] CHOI S， LIM Y， et al. Optimizing Traffic Signal Control Using LLM-Driven Reward Weight Adjustment in Reinforcement Learning[J]. Journal of Information Processing Systems， 2025， 21(1): 43-51.
[26] DA Long-chao， GAO Min-quan， MEI Hao， et al. Prompt to Transfer: Sim-to-Real Transfer for Traffic Signal Control with Prompt Learning[EB/OL]. (2024-01-20)[2025-11-26]. https://arxiv.org/abs/2308.14284.
[27] VARAIYA P. Max pressure control of a network of signalized intersections[J]. Transportation Research Part C: Emerging Technologies， 2013， 36: 177-195.
基本信息:
DOI：10.19961/j.cnki.1672-4747.2025.12.009
中图分类号:U491.54
引用信息:
[1]裘德志,叶宝林,张先超.融合大语言模型与强化学习的事故场景交通信号控制[J].交通运输工程与信息学报().DOI:10.19961/j.cnki.1672-4747.2025.12.009.
基金信息:
国家自然科学基金项目（61603154）; 浙江省自然科学基金项目（LTGS23F030002）; 嘉兴市应用性基础研究项目（2023AY11034）
2026-02-24
2026-02-24
2026-02-24