New Diffusion Model Enhances Probabilistic Time Series Forecasting

Probabilistic multivariate time series (MTS) forecasting is essential for understanding and predicting complex dynamic systems. Current diffusion-based methods for this task often rely on rigid, task-specific conditional paradigms, which struggle with the varying noise levels and evolutionary patterns across different variables—a problem termed "information heterogeneity." To overcome these limitations, researchers introduce DynG-Diff, a variable-sensitive dynamic guidance diffusion framework. This framework employs a two-stage training strategy, utilizing an unconditional diffusion backbone to model the joint distribution of MTS data. Crucially, DynG-Diff incorporates a lightweight state-aware policy network. This policy network adaptively assesses the reliability of each variable based on real-time noisy states and one-step denoising estimates. It then generates a dynamic guidance strength matrix, which is mathematically formulated as the local precision of the observation distribution. This allows for precise guidance for high-confidence variables during inference while effectively filtering out interference from anomalous noise. Extensive experiments demonstrate that DynG-Diff achieves competitive probabilistic forecasting performance against state-of-the-art conditional diffusion models and exhibits improved robustness, especially under severe observation corruption.