Metro Passenger Flow Prediction Model Using Attention-Based Neural Network

Abstract

Metro passenger flow prediction plays an essential role in metro operation system. Due to characteristics of metro operation system, the station operation state is difficult to be described by the passenger flow at a single station. Thus, a novel attention mechanism based end-to-end neural network is presented to predict the inbound and outbound passenger flow to improve predictive effect. The novel model explores the latent dependency between flow of forecast target station and historical flows from surrounding stations by attention mechanism. The relation between variable length flow lists with respect to target station is represented as a fix length vector by the attention mechanism. Furthermore, a deep and wide structure is presented to deal with the inherent information of each station, which are discretized into high dimensional categorical features. Experiments on Beijing Subway line 5 with 1.8 million samples demonstrate the effectiveness of presented approach, which shown the performance on capturing latent dependency.