A Deep Convolutional Neural Network Based Metro Passenger Flow Forecasting System Using a Fusion of Time and Space

Urban rail transit systems produce big automatic fare collection (AFC) data. With the support of big data analytics, more accurate forecasting of passenger flow in the urban rail transit system can optimize the train operation timetable and alleviate urban traffic congestion. Lacking in adequate data, most of the existing works study the urban rail passenger forecasting using time series analysis. The historical passenger data at one station is used to predict its future passenger flow, where the passenger flow correlation between time and station is largely ignored. In this paper, we use the data fusion of time and space to predict the passenger flow. We obtain the historical pristine passenger flow data from the real automatic fare collection system. Teradata big data platform is used to process the data and obtain the spatiotemporal fusion data. Spatiotemporal passenger flow dynamics is converted to a two-dimensional time-space matrix describing the time and space relations of passenger flow. A convolutional neural network model is established under the two-dimensional time-space matrix. We obtain the optimal hyperparameter combinations of CNN models using the grid search algorithm. The performance of the CNN model is evaluated using real metro data. The simulation results demonstrate the high efficiency and accuracy of the proposed method.