Cross-Variable Spatiotemporal Graph Transformer via Data-Driven Interaction Patterns for Urban Multivariate Forecasting

Urban multivariate forecasting plays a critical role in smart city applications, including energy demand estimation, weather-aware mobility planning, and traffic flow prediction, where modeling cross-variable dependencies that evolve over time is essential. Although recent advances in multivariate modeling have succeeded, the majority of existing multivariate forecasting techniques still rely on channel-independent processing, shared cross-variable interaction modeling, or inflexible graph topologies, which significantly limit the models' ability to capture evolving interdependencies. To address this limitation, this paper proposes a cross-variable spatiotemporal graph transformer (CV-STGT) via data-driven interaction patterns for urban multivariate forecasting. The proposed framework consists of two main stages. First, the Temporal Interaction Pattern Discovery (TIPD) module, in combination with the Data-Driven Graph Generator (DGG), is introduced, capable of learning interaction prototypes from the input condition and producing a cross-variable dependency matrix. The second stage is the adoption of spatiotemporal architecture, where independent temporal representations are initially extracted via localized patching, followed by deep cross-variable reasoning implemented with stacked spatial attention blocks aligned with the generated cross-variable dependency matrix. The effectiveness of CV-STGT is verified through experiments across four public datasets, where the proposed model outperforms the state-of-the-art across all datasets, achieving Mean Absolute Error (MAE) improvements of 49.4%, 21.0%, 0.4%, and 0.7% on WTH, ECL, Traffic, and ETTh1 datasets, respectively, at a 168-step forecasting horizon.