基于时间卷积-Transformer模型的多场景地铁短时进站客流预测

Multi-scenario Short Term Inbound Passenger Flow Prediction of Subway Based on TCN-Transformer Model

摘要:

为更好地预测不同场景下城市轨道交通短期进站客流，提出一种基于时间卷积-Transformer组合深度学习模型的多场景进站客流预测方法。该方法考虑时序特征等客流内部特征及日期属性等周期影响因素，通过特征嵌入层构造多因素客流特征输入矩阵，利用时序卷积网络TCN和因果注意力Transformer两个模块并行提取并学习客流数据的局部与全局信息，然后由全连接层构成的预测层输出预测结果。利用上海体育场站客流数据和相关信息验证模型的有效性，并与多个对比模型的预测结果进行比较。实验结果表明：TCN-Transformer模型能够更好地捕捉不同场景下的进站客流特征，具有更好的预测精度和泛化能力。与其他几种模型相比，本文模型的均方根误差（R_MSE）、平均绝对误差（M_AE）、平均绝对百分比误差（M_APE）至少分别降低8.42%、7.32%和6.18%。

Abstract:

In order to better predict the short-term inbound passenger flow of urban rail transit in different scenarios， a multi-scenario inbound passenger flow prediction method based on the combined TCN-Transformer deep learning model is proposed. The method takes into account the internal characteristics of passenger flow such as temporal features and periodic factors such as date attributes. A multifactor passenger flow feature input matrix is constructed through the feature embedding layer. The Temporal Convolutional Network （TCN） and causal attention Transformer modules are used in parallel to extract and learn local and global information from passenger flow data. Then the prediction layer composed of fully connected layers outputs the prediction results. The effectiveness of the model is validated by using the passenger flow data and related information of Shanghai Stadium Station， and compared with the prediction results of several comparative models. The experimental results show that the TCN-Transformer model can better capture the characteristics of inbound passenger flow in different scenarios， and has better prediction accuracy and generalization ability. Compared with several other models， the root mean square error （R_MSE）， mean absolute error （M_AE）， mean absolute percentage error （M_APE） of our model are reduced by at least 8.42%， 7.32%， and 6.18% respectively.