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 (4): 309-320.DOI: 10.16097/j.cnki.1009-6744.2026.04.027

• 智慧机场运营管理 • 上一篇    下一篇

基于大语言模型端到端的机场离港旅客流量预测方法

赵小奇1，唐铁桥*1, 2，李斌3，闫娜1   

  1. 1. 杭州市北京航空航天大学国际创新研究院，杭州 311115；2. 北京航空航天大学，交通科学与工程学院，北京 100191；3. 广东省机场管理集团有限公司工程建设指挥部，广州 510403
  • 收稿日期:2026-01-29 修回日期:2026-03-26 接受日期:2026-04-17 出版日期:2026-08-25 发布日期:2026-08-21
  • 作者简介:赵小奇（1999— ），男，山西太原人，博士生
  • 基金资助:
    国家自然科学基金 (72231001, 72288101)。

Airport Departing Passenger Flow Prediction Based on End-to-End Large Language Model

ZHAO Xiaoqi1, TANG Tieqiao*1, 2, LI Bin3, YAN Na1   

  1. 1. Hangzhou International Innovation Institute, Beihang University, Hangzhou 311115, China; 2. School of Transportation Science and Engineering, Beihang University, Beijing 100191, China; 3. Engineering Construction Command Department of Guangdong Airport Management Group Co Ltd, Guangdong 510403, China
  • Received:2026-01-29 Revised:2026-03-26 Accepted:2026-04-17 Online:2026-08-25 Published:2026-08-21
  • Supported by:
    National Natural Science Foundation of China (72231001, 72288101)

PDF

摘要： 为提升机场离港旅客流量在小时尺度下的短期预测精度与稳定性，本文提出一种基于大语言模型的机场离港旅客流量预测方法（ChatPAX）。该方法将离港客流预测统一表述为受航班计划约束的条件生成式序列预测问题，在端到端框架内联合刻画历史旅客量的需求侧惯性与未来航班供给对客流形成的约束作用。ChatPAX以因果语言模型为预测器，通过结构化提示将历史48 h离港旅客量序列与预测区间内的航班前视信息组织为条件输入，并以固定格式生成未来24 h离港旅客量序列；训练阶段引入答案掩码机制，将监督信号集中于预测目标片段，提升生成稳定性与收敛效率。同时，结合参数高效微调与低精度量化策略，实现单卡显存受限环境下的可训练与可部署预测流程。基于国内某机场多月真实运行数据开展实验，结果表明，ChatPAX在测试集上的全天加权平均绝对百分比误差（WMAPE）均值为13.003%，相较于差分整合移动平均自回归模型（Autoregressive Integrated Moving Average, ARIMA）、长短期记忆网络（Long Short-Term Memory, LSTM）、随机森林（Random Forest）和极端梯度提升算法（Extreme Gradient Boosting, XGBoost）等模型，预测误差分别降低约53.9%、29.5%、26.2%和28.4%，且在跨日期预测中表现更高的稳定性。消融实验进一步表明，引入航班计划这一供给侧约束能够带来稳定性能提升，而节假日与天气等外生因素未产生一致性的边际增益。

关键词: 航空运输, 客流预测, 大语言模型, 机场运行管理, 条件生成式建模

Abstract: To improve the accuracy and stability of short-term airport departing passenger flow prediction at an hourly scale, this paper proposes a large language model-based airport departing passenger flow forecasting method, ChatPAX, where Chat denotes a dialogue-/prompt-based large language model modeling approach, and PAX is a commonly used aviation abbreviation for passengers. The proposed method formulates departing passenger flow prediction as a conditional generative sequence forecasting problem constrained by flight schedules, and jointly characterizes demand-side inertia from historical passenger volumes and the constraint effect of future flight supply on passenger flow formation within an end-to-end framework. ChatPAX uses a causal language model as the predictor. Through structured prompts, the historical departing passenger volume sequence over the past 48 h and the forward-looking flight information within the prediction horizon are organized as conditional inputs, and the departing passenger volume sequence for the next 24 h is generated in a fixed format. During training, an answer-masking mechanism is introduced to concentrate supervision signals on the target prediction segment, thereby improving generation stability and convergence efficiency. Meanwhile, parameter-efficient fine-tuning and low-precision quantization are combined to enable a trainable and deployable forecasting workflow under single-GPU memory constraints. Experiments based on multi-month real operational data from a domestic airport show that ChatPAX achieves a mean all- day weighted mean absolute percentage error (WMAPE) of 13.003% on the test set. Compared with the autoregressive integrated moving average model (ARIMA), long short-term memory network (LSTM), Random Forest, and extreme gradient boosting algorithm (XGBoost), the prediction errors are reduced by approximately 53.9%, 29.5%, 26.2%, and 28.4%, respectively, and the proposed method exhibits higher stability in cross-date prediction. Ablation experiments further indicate that introducing flight schedules as a supply-side constraint can bring stable performance improvements, whereas exogenous factors such as holidays and weather do not produce consistent marginal gains.

Key words: air transportation, passenger flow forecasting, large language models, airport operations management, conditional generative modeling

中图分类号: