Abstract:

An accurate prediction of short-term subway passenger flowscan effectively alleviate traffic congestion and improve the quality of travel services for urban residents. Herein, we propose a multitask learning-based model for the prediction of short-term subway passenger flows, which uses a residual convolutional neural network (NN) and a nested long short-term memory NN to extract the spatio-temporal correlation of traffic patterns, and introduces an attention mechanism to enhance the feature extraction performance of the NNs. Considering the characteristics of subway operations, the model selects train operation features, bus stops around subway stations, and point of interest data as external features to improve the accuracy of the prediction. Based on the historical data of the Beijing Subway, experiments were conducted in multiple time granularity scenarios, such as 10, 30, and 60 min. The results showed that the methodsuccessfully modeled and analyzed the inflow-outflow interaction through multitask learning, improved the prediction performance and generalization ability of the model, and providednovel approaches for the prediction of short-term subway passenger flows.

Key words: subway, passenger flow prediction, multitask learning, attention mechanism, deep neural network