文章编号： 1672-5328（2026）02-0048-11

陈纲梅，周勇，祝佳祥，魏鸿坤，林涛
(深城交科技集团股份有限公司，广东深圳518057)

摘要： 针对中国城市交叉口信号相位设计普遍依赖常规固定相位与基于专家经验的配置的局限性， 突破传统单一算法的局限，提出一种融合生成对抗网络(Generative Adversarial Networks, GAN)与可 解释决策树的智能相位优化方法，并构建了“数据扩增—特征映射—方案推荐”的全流程智能化框 架。该方法创新性地将交叉口静态设施条件(车道渠化、几何布局等)与动态流量特征(转向比例、流 量波动等)深度耦合，采用基于Gumbel-Softmax 改进技术的GAN模型解决交通样本稀缺问题，将实 际采集的159 组交叉口样本高效扩增至15 104 组有效训练数据；进而基于分类与回归树模型算法构 建承担“特征-相位”映射功能的决策树模型，通过信息增益优化节点分裂策略，实现多维度交通 特征与相位方案的精准匹配。在北京市和桐乡市的6 个不同类型交叉口的实证应用表明：本算法使 优化后交叉口的平均排队长度缩短12.3%，平均停车次数减少11.5%。研究成果为城市道路交叉口 动态信号相位优化提供了一种可工程化实践的解决方案。

关键词： 交通控制；信号相位方案；生成对抗网络(GAN)；决策树；AI与交通协同

中图分类号： U491.2

文献标识码：A

An Intelligent Recommendation Method for Intersection Signal Phase Plans Based on the Integration of Generative Adversarial Networks and Decision Trees

Chen Gangmei, Zhou Yong, Zhu Jiaxiang,Wei Hongkun, Lin Tao
(Shenzhen Urban Transport Planning Center Co., Ltd., Shenzhen Guangdong 518057, China)

Abstract： To address the limitations of traditional signal design at urban intersections in China, which mainly relies on fixed phases and expert-based configurations, this paper proposes an intelligent phase optimization method that integrates Generative Adversarial Networks (GAN) with an interpretable decision tree, and develops a full-process intelligent framework encompassing data augmentation, feature mapping, and plan recommendation. The proposed method innovatively combines static intersection conditions, such as lane channelization and geometric layout, with dynamic traffic characteristics, including turning proportions and flow fluctuations. A GAN model based on an enhanced Gumbel-Softmax technique is adopted to address the problem of limited traffic samples. It expands 159 observed intersection samples into 15,104 effective training samples. Furthermore, a decision tree model using the classification and regression tree (CART) algorithm is constructed to perform "feature-phase" mapping. The node splitting strategy is optimized based on information gain, which enables accurate matching between multidimensional traffic characteristics and signal phase plans. Empirical applications at six different types of intersections in Beijing and Tongxiang show that the proposed method reduces the average queue length by 12.3% and the average number of stops by 11.5% after optimization. The results provide a practical and implementable solution for dynamic signal phase optimization at urban intersections.

Keywords： traffic control; signal phase plans; Generative Adversarial Networks (GAN); decision trees; AItraffic integration