Spatio‐Temporal Fusion Conditional Diffusion Network for Multivariate Time Series Prediction

ABSTRACT Multivariate time series forecasting is crucial in intelligent systems and widely applied to tasks like traffic flow prediction. Existing spatio‐temporal forecasting models struggle to capture interactions between temporal dynamics and spatial dependencies, limiting their representation of complex spatio‐temporal patterns. To address this, we propose a multivariate time series forecasting method based on a spatio‐temporal fusion conditional diffusion model. A spatio‐temporal fusion conditional generation module is proposed that integrates dynamic graph construction, timestamp encoding, and spatio‐temporal embedding fusion to overcome static graph limitations. Furthermore, the graph structure and timestamp‐guided reverse denoising process is introduced, injecting spatio‐temporally oriented constraints and prior knowledge into the conditional diffusion mechanism to effectively improve modelling of data uncertainty and complex distributions. Finally, a spatio‐temporal feature‐decoupled denoising network is designed to strengthen the precise separation and joint representation of spatio‐temporally coupled dependencies. Extensive experiments demonstrate that the method significantly enhances spatial modelling accuracy in spatio‐temporal forecasting.