Short-term passenger flow prediction plays a pivotal role in metro operations. Based on the predicted results, metro operators can develop feasible strategies for dynamic train scheduling and passenger flow controlling in peak periods. To gain precise prediction results, a considerable number of deep learning models have been applied. However, the prediction has its intrinsic difficulty for high volatility in small time granularity and the prediction accuracies and stabilities of these models are limited. In this paper, a Long short-term memory neural network (LSTM) based architecture is proposed with the purpose of volatility reduction and precision improvement in short-term passenger inbound flow prediction. The framework combined with high correlated stations data and temporal characteristics to ease data volatility. The case study in Guangzhou implicit that the proposed LSTM-based architecture outperforms the simple LSTM model and achieves a high prediction accuracy of 92.15% . It is a reasonably feasible method to address the data volatility problem that exist in the domain of short-term passenger flow prediction based on deep learning. Metro operators can thus effectively allocate resources to areas with unbalanced proportion about transportation resource for service improvement.

CICTP