基于异构数据特征的城市轨道交通OD客流短时预测方法

A Short-term Prediction for OD Passenger Flow in Urban Rail Transit Based on Heterogeneous Data Feature Extraction

城市轨道交通起讫点(origin-destination,OD)客流短时预测在智能交通系统中意义重大,它为交通管控策略实施以及出行者出行选择提供了重要的决策依据。卷积神经网络被广泛用于交通数据空间相关性提取,但其平移不变性与局部敏感性导致该方法更重视局部特征而忽视全局特征。本研究构建了基于注意力机制的异构数据特征提取机模型(heterogeneous data feature extraction machine,HDFEM)以实现OD矩阵空间相关性的全局感知。该模型从时空特征和用地属性特征出发,构造异构数据OD时空张量与地理信息张量,依托模型张量编码层对异构数据张量进行分割与编码,通过注意力机制连接各张量块特征,提取OD矩阵中各个部分间的空间相关性。该方法不仅实现了异构数据与OD客流数据的融合,还兼顾了卷积神经网络模型未能处理的OD矩阵远距离特征,进而帮助模型更全面地学习OD客流的空间特征。对于OD时序特性,该模型使用了长短时记忆网络来处理。在杭州地铁自动售检票系统(auto fare collection,AFC)数据集上的实验结果表明:HDFEM模型相对于基于卷积神经网络的预测模型,其均方误差、平均绝对误差与标准均方根误差分别下降了4.1%,2.5%,2%,验证了全局OD特征感知对于城市轨道交通OD客流预测的重要性。

As an important basis for rail transit operations and travel choices,prediction for origin-destination(OD)passenger flow in urban rail transit is of great significance in intelligent transportation systems.The conventional convolutional neural network(CNN)mostly focuses on local OD features due to their translation invariance and lo-cal sensitivity.To improve its global perception capacity in OD matrix modeling,a heterogeneous data feature ex-traction machine(HDFEM)model is proposed based on attention mechanism.The model constructs a heteroge-neous data OD spatio-temporal tensor and a geographic information tensor from the perspective of spatio-temporal characteristics and land use attributes.It segments and encodes heterogeneous data tensors via a tensor coding layer to obtain the features of tensor blocks in heterogeneous data tensors.Then,it connects the features of each tensor block through the attention mechanism to extract the spatial correlation among various OD matrix parts.This ap-proach not only realizes multi-source heterogeneous data fusion,but also extracts remote features of OD matrix.Meanwhile,the model uses long short-term memory(LSTM)network to deal with the OD temporal feature.Compared with the convolutional neural network-based prediction model,the results on the Hangzhou metro auto fare collection(AFC)dataset show that the mean square error,mean absolute error,and normalized root mean square er-ror of the HDFEM model decreases by 4.1%,2.5%,and 2%,respectively.The importance of extracting whole spa-tial features for OD passenger flow prediction of urban rail transit is verified.