Urban rail transit (URT) has been widely constructed in Chinese large cities and its increasing complexity of network structure results to a higher requirement in operation mode. To ensure the efficiency of railway operation and management, it is crucial to forecast short-term passenger flow for optimal scheduling and organization. This study aimed to develop a dynamic demand prediction model for urban rail transit passengers using the deep learning method based on the rail transit data from Suzhou, China. Firstly, the spatio-temporal passenger distribution was analyzed after processing the Automatic Fare Collection (AFC) data to observe the passenger flow of different sites and different dates (workday, weekend, and holidays) respectively. Considering the land-use and the station location influence the generation and attraction of people of URT, these two features along with the history passenger flow were introduced into the clustering model and seven types of stations were classified by fuzzy C-means clustering method. Then, the Long Short-Term Memory (LSTM) was applied to predicate short-term passenger. The LSTM network improved from RNN become a popular choice to forecast the dynamic traffic demand, and the data of URT for different stations were chose according to the categories. In order to evaluate the performance the proposed method, the traditional method, ARIMA model, was used as a comparison. The experimental results showed that the LSTM can provide a higher accurate prediction for different types of dates and different stations.

CICTP