Time series forecasting faces multiple uncertainties in complex dynamic scenarios, such as sudden disturbances and volatile data distributions. Traditional methods typically assume that data exhibits stable regularity, and ignoring such non-stationary characteristics makes it difficult to quantify prediction risks. While existing diffusion models generate probabilistic forecasts, their single-granularity modeling approach results in the loss of non-stationary information. Therefore, this paper proposes a granularity-aware diffusion model that achieves precise modeling of uncertainty through a multi-granularity collaborative learning framework. Firstly, a Dynamic Granularity Decomposition (DGD) is designed, utilizing differentiable probabilistic modeling to determine granularity size. Secondly, the Perturbation Restoration Conditional Encoder (PRCE) reconstructs non-stationary features via Gated Linear Units, incorporating prior knowledge from historical time series. Furthermore, a Multi-granularity Guided Diffusion Generator (MGDG) is designed to align coarse-grained trends with fine-grained fluctuations across different denoising stages, achieving gradual distribution optimization through conditionally constrained Markov chains. Experimental results demonstrate that the proposed framework outperforms existing methods on multi-granularity tasks, particularly exhibiting significant advantages in prediction accuracy and uncertainty forecasting.