融合长短期记忆网络和图卷积网络的轨道交通短时客流起讫点预测

轨道交通客流起讫点(origin-destination,OD)矩阵存在时间相关性和空间相关性。根据客流OD的时空特征,提出长短期记忆(long short-term memory,LSTM)网络和图卷积网络(graph convolutional networks,GCN)的短时组合预测方法。预测方法主要利用LSTM网络来获取客流的时间相关性,利用GCN来获取客流的空间相关性,基于出站口建立客流OD矩阵,对整个路网的客流OD进行训练预测。实验表明:融合LSTM神经网络和GCN神经网络的短时预测模型能有效预测轨道交通客流OD。相较于单独的LSTM神经网络,组合模型在预测误差方面有所改善,更适用于短时客流OD的预测。

Urban rail transit short-term passenger flow origin-destination forecast based on LSTM and GCN

The urban rail transit passenger flow origin-destination (OD) matrix has temporal correlation, spatial correlation. According to the spatio-temporal characteristics of passenger flow OD, a short-term prediction method based on long short-term memory (LSTM) neural network and graph convolution network (GCN) is proposed. The proposed prediction method uses the LSTM neural network to capture the temporal correlation, employs the GCN to capture the spatial correlation of the passenger flow, and builds the passenger flow OD matrix based on exit stations to train and test the passenger flow of the whole road network. The experiment shows that the short-term prediction modelled by combing LSTM neural network and GCN can predict the urban rail transit passenger flow OD more effectively. Compared with the single LSTM neural network, the proposed method reduces the prediction error, and is more suitable for short-term passenger flow OD prediction.