Spatiotemporal Graph Attention Network modeling for multi-step passenger demand prediction at multi-zone level

Abstract

Multi-step passenger demand prediction is a crucial task in on-demand vehicle sharing services. However, predicting passenger demand over multiple time horizons is challenging due to nonlinear and dynamic demand patterns. In this work, with the consideration of citywide spatiotemporal dependencies, we propose Graph Attention Networks with Convolution Gated Recurrent Units (GAT-ConvGRU) to predict the multi-step passenger demand by constructing citywide passenger demands into a demand graph. Specifically, GAT blocks are applied to capture spatial dependencies from multiple geographical zones with the help of attention mechanism. Then, based on the ConvGRU, an encoder network is used to capture temporal dependencies. Finally, a ConvGRU-based decoder network is utilized to carry out the multi-step passenger demand prediction at a multi-zone level. Experiments on two real-world datasets validate that the proposed model consistently outperforms state-of-the-art models.