<p>Taxi dispatching involves dynamically allocating vehicle resources, aiming to respond to residents travel demands and enhance operational efficiency. Perceiving residents’ travel patterns is a necessary prerequisite for achieving precise scheduling. However, the complex dependencies between dimensions in traffic multidimensional data make feature extraction and data modeling extraordinarily difficult, posing challenges for accurately mining and analyzing patterns. Additionally, traffic prediction is the core of taxi dispatching, and prediction models based on deep learning need to consider both temporal correlations and spatial dependencies, presenting challenges for real-time traffic information acquisition. To address these issues, this paper proposes a taxi dispatching visual analytics framework, which mainly comprises two modules for exploring travel patterns and predicting demand. We propose a method to explore the travel patterns of residents, which integrates hierarchical progressive aggregation with non-negative tensor decomposition to extract potential patterns from multidimensional spatiotemporal urban data. Meanwhile, we propose R-GCN, a deep learning model that fully considers spatiotemporal dependencies, for real-time prediction of taxi demand in various regions. Experimental results based on different datasets demonstrate that the method improves prediction accuracy. Furthermore, we develop TPMVis to support users in intuitively perceiving and deeply exploring resident travel patterns and taxi demand trends. Through the exploration of real-world urban traffic data and case studies conducted with domain experts, the effectiveness of TPMVis in taxi dispatching tasks is validated.</p> Graphical abstract <p></p>