Subway passenger flow prediction based on optimized PSO-BP algorithm with coupled spatial-temporal characteristics

To improve the accuracy of subway passenger flow prediction, by considering the Xi'an Metro Line 1 as an example, five main factors affecting subway passenger flow variations, such as festival, non-festival, time period, station, and weather, were extracted to analyze the coupled spatial-temporal characteristics of subway passenger flow. A back propagation (BP) neural network was constructed to predict the subway passenger flow. The proposed BP neural network was further optimized by using a particle swarm optimization (PSO) algorithm that introduced adaptive mutation and balanced inertia weights to form a subway passenger flow prediction system that could consider complex influence factors. Transfer stations and non-transfer stations including a first and an intermediate station were selected, the weather, festival, and non-festival factors were considered, and the BP neural network models for different time periods were compared. Then, the prediction errors of the PSO-BP neural network model were optimized. Research results show that by considering the weather, festival and non-festival factors, the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) of the optimized PSO-BP neural network model predictions at transfer stations within the optimized time periods decrease by 40.13%, 31.46% and 23.89%, respectively, compared with the optimized PSO-BP neural network models prediction errors without the time periods, decrease by 17.50%, 17.86% and 17.32% compared with the BP neural network models prediction errors within the optimized time periods. The MAE, RMSE, and MAPE of the optimized PSO-BP neural network model predictions in the non-transfer stations within the optimized time periods decrease by 16.50%, 20.99% and 32.59%, respectively, compared with the optimized PSO-BP neural network model prediction errors without time periods, and decrease by 11.48%, 12.10% and 17.73%, respectively, compared with the BP neural network model prediction errors within the optimized time periods. The MAE, RMSE, and MAPE of the optimized PSO-BP neural network model predictions at each station within the optimized time periods decrease by 24.37%, 24.48% and 29.69%, respectively, compared with the optimized PSO-BP neural network model prediction errors without time periods, and decrease by 13.49%, 14.02% and 17.59%, respectively, compared with the BP neural network model prediction errors within the given time periods. Therefore, using the optimized PSO-BP neural network model and considering the influencing factors can improve the accuracy of subway passenger flow prediction. © 2021, Editorial Department of Journal of Traffic and Transportation Engineering. All right reserved.