DynG-Diff: A State-Aware Dynamic Guidance Diffusion Framework for Probabilistic Time Series Forecasting

Time series forecasting is crucial in fields like energy, transportation, finance, and healthcare. Traditional methods, such as deterministic models, struggle to capture the stochasticity and multimodal distributions in complex dynamic systems, driving the development of generative probabilistic models. Diffusion models have become the preferred framework for generative modeling due to their progressive noise removal mechanisms and exceptional distribution-fitting capabilities.

Existing diffusion models rely on task-specific conditional paradigms, struggling to address the challenge of information heterogeneity in multivariate systems. Different variables exhibit significantly varying noise levels and evolutionary patterns, leading to decreased forecasting accuracy.

DynG-Diff combines an unconditional diffusion model with a state-aware policy network to implement dynamic guidance for multivariate time series. This framework adapts to different task scenarios without retraining the model, addressing the issue of information heterogeneity.

• �� Pre-train an unconditional diffusion model to capture the joint distribution of multivariate time series.
• �� Use a state-aware policy network to infer variable reliability based on real-time noisy states and one-step denoising estimates.
• �� Generate dynamic guidance strength matrices for precise guidance of high-confidence variables, filtering out anomalous noise.

Experiments were conducted using multiple real-world datasets, including transportation, finance, and healthcare domains. Benchmarks were compared against existing models like CSDI, using forecasting accuracy and probabilistic metrics as evaluation standards. Ablation studies validated the effectiveness of the policy network.

DynG-Diff achieved significant improvements in forecasting accuracy across multiple datasets, particularly under severe noise conditions. It demonstrated stronger robustness and probabilistic forecasting performance compared to existing conditional diffusion models.

The framework can be directly applied to scenarios such as energy management, traffic flow forecasting, and financial market analysis, significantly improving prediction accuracy and decision quality.

In extreme noise conditions, the policy network may fail to accurately estimate variable reliability, leading to prediction bias. Handling ultra-long sequences incurs significant computational overhead, necessitating network design optimization.