【摘要】 节假日作为居民集中出行的高峰期，其客流特征直接关系到铁路运营的安全、运力配置效率和服务质量。节假日期间的铁路客流呈现出与日常显著不同的特殊性，主要表现为长距离出行需求剧增、旅游流与探亲流高度叠加，以及客流分布的时空不均衡性，为铁路运营管理带来了挑战。一是客流需求的突增，热门线路和高峰时段的运输能力趋于饱和，传统时间序列模型难以捕捉这种剧烈的非平稳波动；二是预售数据不完整性，旅客购票行为贯穿整个预售期，不同时间点获取的预售数据反映的未来客流信息是动态变化的；三是客流受时间、节假日效应、列车运行安排等多种因素共同影响，这些特征之间存在复杂的非线性耦合关系。为解决上述问题，提出一种基于Enhanced Transformer的铁路客运站节假日客流预测模型。在特征工程方面，主要从时间特征、节假日特征和运营特征3个维度构建了多源特征体系：时间特征包括预售提前量和小时周期编码，用于捕捉旅客出行决策行为和一天内客流的规律性波动；节假日特征涵盖周末指示、节假日标记、节前高峰和节假日周末叠加效应，用于精确捕捉节假日期间客流模式的突变特征；运营特征则提取了每小时上下行列车班次数，反映车站的实时运力供给情况。通过多头自注意力机制，模型能够在不同的表示子空间中并行学习这些多源特征间的复杂交互模式，实现对客流驱动因素的深度理解。创新性地将动态变化的预售数据作为关键输入特征，结合模型的时序信息处理能力，实现对未来客流的滚动预测，突破传统方法在处理预售期动态性上的局限，通过选取苏州地区4个核心铁路客站（苏州北站、苏州站、苏州新区站、苏州园区站）在2025年春节期间的客流数据进行案例分析。实验结果表明，Enhanced Transformer模型对于苏州北站和苏州站等客流规模大的枢纽站，预测准确率可达84.06%，证明了模型在处理高流量、高波动性时间序列数据时的有效性。与Transformer,XGBoost,LSTM,Bi-LSTM的4种基准模型的对比实验显示，Enhanced Transformer在MSE,RMSE,MAE和准确率等所有评估指标上均全面优于其他模型。相较于标准Transformer模型，其预测准确率提升了约6.29%～6.89%；相较于LSTM，准确率提升约3.4%。这些性能提升归因于模型在长序列依赖捕捉、非平稳数据适应和多源特征交互方面的结构优势，为铁路管理部门提供了有力的技术支持，有助于实现节假日期间运力的精准配置、提升旅客服务质量和保障运营安全。更多还原

【Abstract】 Holidays represent peak periods for concentrated travel, and the characteristics of passenger flow during these times are directly linked to the safety of railway operations, the efficiency of capacity allocation, and the service quality. Railway passenger flow during holidays exhibits distinct specificities compared to daily operations, primarily manifested in a surge in long-distance travel demand, a high overlap of tourism and family visitation flows, and a significant spatiotemporal imbalance, all of which pose challenges to railway operation management. These challenges include: first, a sudden surge in demand where transportation capacity on popular routes and during peak hours reaches saturation, making it difficult for traditional time-series models to capture such intense non-stationary fluctuations; second, the incompleteness of pre-sale data, as passenger ticketing behavior spans the entire presale period, meaning the future flow information reflected in pre-sale data changes dynamically across different time points; and third, the complex non-linear coupling relationship between passenger flow and multiple factors such as time, holiday effects, and train schedules. To address these issues, this paper proposes a holiday passenger flow prediction model for railway passenger stations based on an Enhanced Transformer. In terms of feature engineering, a multi-source feature system is constructed across three dimensions: temporal features, including advance booking time and hourly cycle encoding to capture passenger decision-making behavior and regular hourly fluctuations; holiday features, covering weekend indicators, holiday markers, pre-holiday peaks, and holiday-weekend overlap effects to precisely capture mutations in passenger patterns; and operational features, which extract the number of hourly inbound and outbound train services to reflect real-time capacity supply. Through a multi-head self-attention mechanism, the model can parallelly learn complex interaction patterns among these multi-source features in different representation subspaces, achieving a deep understanding of passenger flow drivers. Furthermore, this paper innovatively utilizes dynamic pre-sale data as a key input feature, combined with the model’s time-series processing capabilities, to achieve rolling forecasts of future passenger flow, breaking through the limitations of traditional methods in handling pre-sale dynamics. This study selects passenger flow data from four core railway stations in the Suzhou area(Suzhoubei Railway Station, Suzhou Railway Station,Suzhouxinqu Railway Station, Suzhou Industrial Park Railway Station) during the 2025 Spring Festival for case analysis. Experimental results show that for large hub stations like Suzhou North and Suzhou, the Enhanced Transformer model achieves a prediction accuracy of 84.06%, proving its effectiveness in processing high-volume and high-volatility time-series data. Comparative experiments with four benchmark models—Transformer, XGBoost, LSTM, and Bi-LSTM—demonstrate that the Enhanced Transformer consistently outperforms all others across evaluation metrics including MSE, RMSE, MAE, and the accuracy. Compared to the standard Transformer, the prediction accuracy improved by approximately 6.29% to 6.89%; compared to LSTM, the accuracy is increased by about 3.4%. These performance gains are attributed to the model’s structural advantages in capturing long-sequence dependencies, adapting to non-stationary data, and facilitating multi-source feature interaction. The research results provide robust technical support for railway management departments, aiding in the precise allocation of capacity during holidays, enhancing passenger service quality, and ensuring operational safety.更多还原