Abstract:

A scalable deep learning framework is proposed to leverage multi-source data in bus passenger flow prediction. First， on the basis of four external factors， three internal factors are introduced as explanatory variables of passenger flow of a bus route. The variance reduction method is used to verify the linkage relationship between internal factors and external factors， and the necessity of capturing the linkage relationship of multi-source data. Then， by utilizing the advantages of convolutional operations of convolutional neural networks in handling two-dimensional data， the influential factors of passenger flow are visualized and an hourly passenger flow subdivision matrix is constructed to adapt to convolutional operations， such that the linkage between multi-source data is captured. To further improve the prediction performance， the matrix structure optimization problem is transformed into a Traveling Salesman Problem， and the surrogate-based optimization technique is used to efficiently optimize the structure of hourly passenger flow subdivision matrix. Finally， the data from Route-281 bus in Guangzhou， China was used as an example for validation. The results show that， through optimizing the structure of hourly passenger flow subdivision matrix， the prediction accuracy can be effectively improved， and optimal usage of data resources can be achieved. The independent effect of internal factors was not significant， whereas the combined effect of external and internal factors is significant. In terms of prediction accuracy，the proposed method has certain advantages compared to the results only considering external factors and other deep learning models.

Key words: engineering of communication and transportation system, passenger flow prediction, deep learning, bus passenger flow, multi-source data, surrogate-based optimization