Eliciting Frequency-Conditioned Spatial Dynamics for Long-Term Spatio-Temporal Forecasting

Accurately forecasting the long-term spatio-temporal dynamics is fundamental to public service management and urban renewal. However, existing approaches often struggle to capture spatial dynamics within entangled long-term temporal patterns, inherently constrained by two fundamental gaps: (i) the absence of a unified temporal decomposition scheme that generalizes across diverse urban scenarios, and (ii) insufficient capability to learn spatial correlations that evolve across co-existing temporal patterns over long horizons. To tackle these limitations, we propose FCoSD, an end-to-end framework that elicits frequency-conditioned spatial dynamics for long-term spatio-temporal forecasting. It aims to reorganize the raw spatio-temporal signal into frequency-aware components via adaptive frequency decomposition and learning spatial dynamics tailored to each component, thereby achieving accurate and efficient forecasting for various urban scenarios. Concretely, FCoSD introduces a learnable Gaussian soft frequency-band mask mechanism to decouple long-term temporal patterns into components dominated by low, mid, and high-frequency in a data-adaptive manner. It further equips node variables with a frequency-conditioned spatial Mamba module to capture dynamic spatial dependencies within each frequency band. Finally, a multi-band gated fusion module integrates the decomposed representations into a coherent long-term spatio-temporal embedding for forecasting. Extensive experiments on multiple real-world spatio-temporal datasets demonstrate that FCoSD outperforms state-of-the-art methods.