交通运输系统工程与信息

交通运输系统工程与信息 ›› 2026, Vol. 26 ›› Issue (2): 113-124.DOI: 10.16097/j.cnki.1009-6744.2026.02.011

• 智能交通系统与信息技术 • 上一篇    下一篇

城轨车站进站大客流自适应预警智能体研究

宋绪扬1,2 ，刘方达2 ，庞文丰2 ，王芳盛*3 ，万怀宇1   

  1. 1. 北京交通大学，计算机科学与技术学院，北京100044；2.佳都科技集团股份有限公司，中央研究院，广州510653； 3. 上海工程技术大学，城市轨道交通学院，上海201620
  • 收稿日期:2025-12-10 修回日期:2026-01-09 接受日期:2026-01-15 出版日期:2026-04-25 发布日期:2026-04-20
  • 作者简介:宋绪扬（1996—），男，新疆克拉玛依人，博士后研究员。
  • 基金资助:
    国家自然科学基金 (60634010, 60776829)。

Adaptive Early Warning Intelligent Agent for Inbound Large Passenger Flows at Urban Rail Transit Stations

SONG Xuyang1,2, LIU Fangda2, PANG Wenfeng2, WANG Fangsheng*3, WAN Huaiyu1   

  1. 1. School of Computer Science and Technology, Beijing Jiaotong University, Beijing 100044, China; 2. Central Research Institute, PCI Technology Group Co Ltd, Guangzhou 510653, China; 3. School of Urban Rail Transit, Shanghai University of Engineering Science, Shanghai 201620, China
  • Received:2025-12-10 Revised:2026-01-09 Accepted:2026-01-15 Online:2026-04-25 Published:2026-04-20
  • Supported by:
    National Natural Science Foundation of China (60634010, 60776829)。

PDF

摘要： 针对城市轨道交通进站大客流预警中阈值固化、响应滞后及适配性不足等问题，本文提出一种融合大语言模型（LLM）的自适应预警智能体方法。该方法构建“离线聚类-实时预警-LLM自适应修正”技术框架：首先，采用“层次聚类预处理+k-means优化”两阶段算法挖掘历史客流规律，生成包含日期类型和峰值特征等信息的典型客流模式库；其次，基于模式匹配，计算客流极值、变化率和累积量3类异常指标，利用四分位距法动态确定各模式下的预警阈值，并划分3级预警等级；最后，引入LLM构建自适应预警智能体框架，基于RAG（检索增强生成）技术实现预警参数的自然语言交互式优化与动态修正。以青岛地铁五四广场站373d进站客流数据为样本的实证表明：系统成功识别5类典型客流模式，并在五一假期某日高客流场景中，将原本可能触发的24次预警降低为7次有效预警，显著降低误报率；同时，案例通过语言指令支持参数调整和预警灵敏度变化，实现28.6%的预警覆盖率提升。本方法实现预警阈值与客流模式的动态绑定，融合LLM的自适应优化能力与自然语言交互优势，有效解决传统预警的滞后性与适配性问题。

关键词: 铁路运输, 大客流预警, 大语言模型, 聚类分析, 城市轨道交通, 智能体

Abstract: To address the issues of fixed thresholds, response lag, and poor adaptability in early warning of inbound large passenger flow in urban rail transit, this study proposes an adaptive early warning agent method integrated with a Large Language Model (LLM). A technical framework is proposed: "offline clustering-real-time early warning-LLM-based adaptive correction". A two stage algorithm of "hierarchical clustering preprocessing + k-means optimization" is adopted to analyze historical passenger flow patterns, and generate a typical passenger flow pattern library which includes information such as date types and peak characteristics. Then, three types of abnormal indicators (passenger flow extremum, change rate, and cumulative volume) are calculated based on pattern matching. The interquartile range method is used to dynamically determine early warning thresholds under each pattern and defines three early warning levels. The LLM is then introduced to build an adaptive early warning agent, which realizes natural language interactive optimization and dynamic correction of early warning parameters based on the Retrieval-augmented Generation (RAG). The results of empirical study using 373-day inbound passenger flow data from Wusi Square Station of Qingdao Metro show that the system successfully identifies 5 types of typical passenger flow patterns. In a high passenger flow scenario on a day during the May Day holiday, the number of early warnings that might have been triggered was reduced from 24 to 7 valid ones, significantly reducing the false alarm rate. Meanwhile, the case supports parameter adjustment and changes in early warning sensitivity through language instructions, achieving a 28.6% improvement in early warning coverage. This method realizes the dynamic binding of early warning thresholds and passenger flow patterns, and integrates the adaptive optimization capability and natural language interaction advantages of LLMs, effectively solving the lag and adaptability problems of traditional early warning systems.

Key words: railway transportation, large passenger flow warning, large language mode, cluster analysis, urban rail transit, agent

中图分类号: