A spatial-temporal dynamic attention based Mamba model for multi-type passenger demand prediction in multimodal public transit systems

Abstract

Predicting multi-type passenger demand, such as for adults, seniors, pensioners, and students, is essential for improving the operational efficiency, equity, and sustainability of multimodal public transit (PT) systems. However, traditional demand prediction models often struggle to capture the complex spatial-temporal variability inherent in diverse socio-demographic groups. To address this gap, we propose a novel spatial-temporal dynamic attention-based state-space model, i.e., STDAtt-Mamba, tailored for multi-type passenger demand prediction in multimodal PT systems. The STDAtt-Mamba model comprises three key components: an adaptive embedding layer that integrates station-level, passenger-type-specific, and temporal embeddings into a unified representation for efficient data processing; a spatial-temporal dynamic attention (STDAtt) module that employs sparse attention mechanisms to selectively capture crucial global spatial-temporal dynamics; and a spatial-temporal dynamic Mamba (STDMamba) module that extends state-space modeling to fuse spatial and temporal dependencies dynamically. We reformulate STDAtt-Mamba as a spatial-temporal dual-path attention mechanism and theoretically prove the complementarity of STDMamba and STDAtt in capturing local and global dependencies, thereby improving the interpretability of the STDAtt-Mamba. We conduct extensive experiments on a large-scale multimodal dataset of over 1.58 million smart card users of 9 passenger types from Queensland, Australia, from 01/2021 to 01/2023. Experimental results demonstrate that STDAtt-Mamba outperforms state-of-the-art baseline models regarding the prediction accuracy across all passenger types and travel modes. By addressing the challenges of heterogeneity in spatial-temporal travel patterns and socio-demographic groups, this study offers an adaptive, robust, scalable, and data-driven tool for managing the heterogeneous passenger demand in multimodal PT systems.