Joint prediction of zone-based and OD-based passenger demands with a novel generative adversarial network

Abstract

Online ride-hailing plays an important role in modern urban transportation systems, and accurate short-term passenger demand prediction contributes to improving ride-hailing services. Many existing studies have made great achievements in zone-based demand prediction. However, origin–destination (OD)-based demand prediction has attracted little attention, even though it provides abundant information and is of high practical importance, e.g., it facilitates the routing and matching of ride-hailing services. Considering the importance of both types of demands, a conditional generative adversarial network (CGAN) with Wasserstein divergence (CWGAN-div) is proposed to jointly predict zone-based and OD-based demands. Residual blocks are utilized to capture internal spatiotemporal features, which benefit the training and prediction processes of the CWGAN-div model. Conditional information is also incorporated to characterize the external dependencies of the demands. Numerical experiments are performed by using GPS trajectory data from Didi Chuxing, Chengdu, China. The results show that the proposed CWGAN-div model yields good joint prediction performance and outperforms both classic models (historical average (HA) and convolutional neural network (CNN) models) and other prevailing GAN models (i.e., GAN, CGAN, Wasserstein GAN with a gradient penalty (WGAN-GP), and GAN with Wasserstein divergence (WGAN-div)). The proposed CWGAN-div model displays promise for the joint prediction of zone-based and OD-based passenger demands.