CDGAT: A conditioned diffusion-based graph attention network for spatio-temporal prediction

Spatio-temporal series prediction is critical for informed decision-making in domains such as environmental monitoring, traffic, and energy. Although deep learning has significantly improved prediction accuracy, most existing architectures remain deterministic, failing to provide the uncertainty quantification essential for risk assessment. To address this gap, this paper introduces CDGAT, a Conditioned Diffusion-based Graph Attention Network for probabilistic spatio-temporal series prediction. CDGAT integrates Graph Attention Networks with Denoising Diffusion Implicit Models to effectively capture complex spatial and temporal dynamics. Key innovations include a -means Correlation Dimension method for semi-supervised time-series model order estimation, integrated alongside exogenous variables via a cross-attention mechanism. The predictive performance of CDGAT was comprehensively evaluated across diverse real-world benchmarks. Specifically, CDGAT compares favourably against state-of-the-art opponents on the Elergone energy dataset and consistently ranks among the top two models across the Metr-LA and air pollutant benchmarks, while maintaining robust uncertainty intervals.