Daily OD Demand Prediction in Urban Metro Transit System: A Convolutional LSTM Neural Network with Multi-factor Fusion Channel-wise Attention

Abstract

The distribution of Metro passenger flow is an important indicator of the operation status of urban rail systems. The accurate short-term passenger flow prediction at the network level plays an important role in the management of smart urban rail transit. Compared with road systems, urban rail transit OD (Origin-Destination) distribution has more obvious characteristics such as high dimension and sparseness, which brings great challenges to Data-driven prediction models. Although the existing research has made some progress in mining spatial-temporal features, the model lacks objective consideration in terms of the rationality of prediction framework embedding, multi-source data fusion, and spatial-temporal feature mining integration. To solve the above problems, we propose a deep learning framework - Convolutional LSTM (Long Short-Term Memory) Prediction Model (STMACL) based on multi-factor fusion channel-wise attention mechanism. First, we design a multi-factor spatial-temporal feature parallel structure encoder to capture the intrinsic relationship between inter-station, date attributes, external factors and high-level spatial correlation of passenger flow distribution, and use the Convolutional LSTM to further abstract spatial-temporal features to output the prediction result. The proposed framework was tested on two large-scale real datasets in Shanghai Metro, and its performance outperformed other benchmark methods, enabling higher prediction accuracy at the network-level, with high flexibility, scalability and robustness.