Wildfire risk plays a critical role in natural disaster emergency response and climate change mitigation. Existing data-driven methods mainly focus on short-term wildfire forecasting (days to months), often relying on numerous driving factors and binary fire/no-fire representations. In this paper, we redefine wildfire risk from the perspective of next-year wildfire state transitions and propose a novel Physical-Aware Wildfire State Transition Discrete Diffusion Model (PA-WildfireSTDDM) that directly learns the high-dimensional distribution of wildfire risk using only historical wildfire records, with the following contributions: (1) We construct a 25-year (2000–2024) daily wildfire dataset for British Columbia (BC), Canada, derived from FIRMS at 10 km spatial resolution through spatial aggregation. Four wildfire state transition categories are defined: <em>Persistent no-fire</em>, <em>New ignition</em>, <em>Fire cessation</em>, and <em>Persistent fire</em>. (2) We propose a physically aware diffusion framework, where the prediction process is regularized using a physical and empirical wildfire transition kernel, improving physical consistency and long-term prediction reliability. (3) The proposed end-to-end model predicts the categorical distribution of wildfire state transitions conditioned on historical wildfire events and preceding wildfire states, reducing cumulative errors compared with iterative forecasting approaches. (4) Our framework generates high-confidence maps of next-year wildfire states using only long-term historical wildfire records, without requiring additional environmental or climate-driving factors, while effectively capturing complex and stochastic wildfire patterns. (5) The model formulates wildfire evolution as a discrete-time inhomogeneous stochastic process, enabling uncertainty quantification for next-year wildfire projections through Monte Carlo posterior sampling.