Network-wide ride-sourcing passenger demand origin-destination matrix prediction with a generative adversarial network

Abstract

Accurate short-term passenger demand origin-destination (OD) matrix prediction contributes to the coordination of traffic supply and demand. This study proposes a novel generative adversarial network (GAN) named Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (CWGAN-GP) to predict the network-wide ride-sourcing passenger demand OD matrix. The proposed CWGAN-GP model can not only capture internal spatiotemporal features of OD matrices, but also characterise external dependencies of OD matrices on conditional information, such as the traffic zone-based average traffic speeds, the traffic zone area, and time variables. Based on the ride-sourcing GPS trajectories from Didi Chuxing, Chengdu, China, and ride-sourcing data from the New York City, numerical results illustrate that the predicted OD matrices are in good agreement with the actual ones, and CWGAN-GP has good convergence performance by analysing the discriminator loss and the Wasserstein distance with respect to training epochs. Comparison results also validate the outperformance of CWGAN-GP compared with the other counterpart prediction methods and the reasonability of specific structures of CWGAN-GP. Thus, CWGAN-GP is concluded to be promising to predict network-wide ride-sourcing passenger demand OD matrices.