Deep learning-based hybrid model for short-term subway passenger flow prediction using automatic fare collection data

Abstract

Short-term passenger flow prediction is one key prerequisite of decision making in daily operations. Subway operators concentrate on both the efficiency of the prediction model and the accuracy of the prediction results. In this study, the authors explore a deep learning-based hybrid model, which integrates a long-short-term memory neural network (LSTM NN) and stacked auto-encoders (SAEs), for predicting short-term passenger flows of each station in a subway network simultaneously. SAEs are employed to extract network passenger flow data features, which involves mapping high-dimensionality data to low-dimensionality data at the first stage. Then, LSTM NN is trained using low-dimensionality data. Finally, SAEs restore the predicted data output by LSTM NN to predict the network passenger flow data. By employing automatic fare collection data of the Guangzhou subway, the proposed hybrid model was evaluated. The experimental results show that the average mean relative error of the proposed hybrid model is 4.6%, which is much better than other current prediction models.