Short-term forecasting of rail transit passenger flow based on long short-term memory neural network

Abstract

Short-term forecasting of passenger flow in metro station is gaining increasingly popularity in the domain of rail transit, because this technique can provide reliable evidence for daily operation and management in rail transit system. Recently, artificial neural networks, especially Recurrent Neural Networks (RNNs) have been receiving more and more attention, due to their capability to capture the strong nonlinearity and randomness of short-term passenger flow. However, traditional recurrent neural networks are unable to learn and remember over long sequences due to the issue of back-propagated error decay. To address this problem, a novel neural network architecture, Long Short-term Memory Neural Network (LSTM NN) for short-term forecasting is proposed in the study. Root mean squared errors (RMSE), mean absolute percentage errors (MAPE) and variance of absolute percentage error (VAPE) are calculated as indicators to evaluate the prediction performance. Other topologies of recurrent neural networks, such as Simple Recurrent Neural Network (RNN) and Gated Recurrent Unit (GRU), are compared to validate the effectiveness of the proposed model. The empirical study with real datasets from Guangzhou Metro shows that LSTM NN outperforms other neural networks in terms of accuracy and stability for short-term forecasting with a 15 min interval.