A General Spatio-Temporal Backbone with Scalable Contextual Pattern Bank for Urban Continual Forecasting

ICLR 2026 Conference Submission12674 Authors

Keywords: general backbone, contextual pattern bank, continual spatio-temporal forecasting

Abstract: With the explosive growth of spatio-temporal data driven by IoT deployments and urban infrastructure expansion, accurate and efficient continual forecasting remains a critical challenge. Recent Spatio-Temporal Graph Neural Networks assume static graph topologies and temporal scales, making them ill-suited for dynamic real-world data streams. Meanwhile, existing continual learning methods often adopt simple backbones, limiting their ability to capture evolving dependencies and adapt to distributional drift. To address these limitations, we propose STBP, a novel framework for Continual Spatio-Temporal Forecasting that bridges the gap between STGNNs and continual learning. STBP integrates a general-purpose spatio-temporal backbone with a scalable contextual pattern bank. The backbone extracts stable spatio-temporal representations in the frequency domain and models dynamic spatial correlations using linear graph attention. To support continual adaptation and alleviate catastrophic forgetting, the contextual pattern bank is incrementally updated via parameter expansion, capturing evolving node-level heterogeneous patterns. During incremental training, the backbone remains frozen to preserve general knowledge, while the contextual pattern bank adapts to new scenarios and distributions. Extensive experiments show that STBP surpasses state-of-the-art baselines in both accuracy and scalability, underscoring its effectiveness for continual spatio-temporal forecasting.

Primary Area: learning on time series and dynamical systems

Submission Number: 12674