A deep learning and ensemble learning based architecture for metro passenger flow forecast

Abstract

Accurate short-term forecast of metro outbound passenger flow is of great significance for real-time traffic control and guidance. A good forecast method should have high accuracy, timeliness and practicality. Based on deep learning and ensemble learning technology, this study proposes an end-to-end forecast hybrid architecture for metro outbound passenger flow that integrates multiple passenger flow features. The architecture innovatively integrates bagging ensemble learning strategy and transfer learning with deep learning, and includes multiple extensible passenger flow feature processing components. In addition, this study presents a new coding method to incorporate the operating characteristics of the metro into the forecasting architecture. Use the automatic fare collection (AFC) data of Chengdu Metro Tianfu Square Station for training and verify the forecast architecture on workdays, weekends and holidays. The results reveal that compared with other widely used passenger flow forecast models, the architecture proposed in this study has achieved the highest forecast accuracy in the above-mentioned different time types. Furthermore, the fusion of transfer learning improves the accuracy of forecast model while significantly speeding up the convergence of training, increasing its timeliness and practicability.