Abstract: Due to the lack of micro-granularity travel behavior prediction model in metro system, a deep learning prediction framework based on attention mechanism LSTM was proposed to predict the travel behavior of passengers in metro system. As the travel information of the passengers in metro system had the characteristics of high latitude and multi-forms, two feature extraction modules were proposed to capture the temporal and spatial features respectively. By use of the attention mechanism and the characteristics of LSTM neural network, the sequential characteristics of long travel sequence were modeled, so as to predict the temporal and spatial information of passengers next trip. AFC data of Guangzhou Metro was used to verify the proposed model. The results show that the proposed time information feature extraction method can more accurately characterize the time feature. Compared with the conventional statistic learning models, the prediction accuracy for inbound station, outbound station points and inbound time of the proposed prediction model has reached 74.9%, 61.6% and 44.8% respectively, which has been greatly improved compared with the existing algorithms.

Key words: traffic and transportation engineering; metro system; travel behavior; feature extraction; LSTM; attention mechanism

摘要： 针对地铁系统缺乏微观粒度出行行为预测模型的问题，提出了一种基于注意力机制LSTM的深度学习预测框架对地铁系统乘客出行行为进行预测。针对乘客地铁出行信息高纬度、多形式的特点，提出了两种特征提取模块分别对时间和空间信息进行特征提取；利用注意力机制与LSTM神经网络的特点，对长出行序列的时序特征进行建模从而预测乘客下一次出行的时空信息。利用广州地铁AFC数据对所提出模型进行验证，结果表明：所提出时间信息特征提取方法可以更加准确的表征时间特征；相较于传统统计学习模型，提出的预测模型对于进、出站站点及进站时间的预测准确率分别达到了74.9%、61.6%、44.8%，相对于现有算法取得了较大提升。

关键词: 交通运输工程；地铁系统；出行行为；特征提取；LSTM；注意力机制