Abstract: Aiming at the inherent limitation that conventional traffic signal control systems are difficult in adapting to dynamic traffic flow, as well as the problems such as insufficient model generalization capability of existing reinforcement learning methods and high deployment complexity of large language models (LLM), a lightweight large language model-based traffic signal control system (L3M-TCS) was proposed. Firstly, an instruction-based fine-tuning dataset specifically designed for traffic signal control was constructed. Secondly, through fine-tuning and parameter quantization techniques, the LLM was compressed into a lightweight architecture suitable for roadside devices. Finally, a verification system architecture with real-time feedback mechanisms was designed to validate its effectiveness and reliability in real-world traffic environments. Research results demonstrate that: compared to the traditional fixed-timing scheme, L3M-TCS reduces traffic delays by 60.6% and queue lengths by 50.2%. Compared to reinforcement learning methods, L3M-TCS reduces the delay at untrained intersections by 64.7%, while providing natural language-based explanations for decision-making. The proposed model achieves an inference speed of 11.41 tokens/s on roadside device side, with memory footprint compressed to 81.7% of the original model, while maintaining control decision latency within 2 500 ms.

Key words: traffic engineering; intelligent transportation; traffic signal control; large language model; command data set; domain-specific fine-tuning

摘要： 针对传统交通信号控制系统难以适应动态交通流量的固有缺陷，且现有强化学习方法存在的模型泛化能力不足、大语言模型部署复杂度高等问题，提出了基于轻量化大语言模型(LLM)的交通信号控制系统(L3M-TCS)。首先构建面向交通信号控制的指令微调数据集；其次采用微调与参数量化技术，将大语言模型压缩为适用于路侧设备的轻量化架构；最后设计具有实时反馈机制的验证系统架构，验证其在真实交通环境中的有效性和可靠性。研究表明：相较传统固定配时方案L3M-TCS降低了60.6%的交通延误与50.2%的排队长度；相比强化学习方法，L3M-TCS在未训练交叉口的延误降低了64.7%，同时可提供基于自然语言的决策依据解释；模型在路侧设备端的推理速度达到11.41 tokens/s，内存占用压缩至原模型的81.7%，且控制决策延迟在2 500 ms以内。

关键词: 交通工程；智能交通；交通信号控制；大语言模型；指令数据集；领域化微调