<p>Short-term wind speed forecasting remains challenging because atmospheric wind fields exhibit strong non-stationarity and stochastic variability, which limit the effectiveness of conventional deterministic forecasting models, particularly under complex or data-constrained conditions. This study proposes a Diffusion-Encoder framework that reformulates forecasting as a conditional diffusion-based generative process and serves as a model-agnostic enhancement layer for existing forecasting architectures. The framework was integrated with five representative deep learning architectures and evaluated using four datasets spanning mountainous and offshore environments under both warm- and cold-season conditions. Performance was assessed across 1-, 3-, 6-, and 9-step forecasting horizons and under progressively reduced training data availability (100%, 70%, 50%, and 30%). Across most forecasting scenarios, the diffusion-enhanced variants achieved lower forecasting errors than their deterministic counterparts, with the largest gains observed under longer forecasting horizons, limited training data, and more complex wind regimes. Statistical evaluation further showed that 82.5% of model–horizon comparisons yielded significant forecasting improvements. Under severe data scarcity, diffusion-enhanced models consistently exhibited slower error growth and maintained more stable predictive performance than the corresponding baseline models. Overall, the current study introduced the development of a model-agnostic diffusion enhancement framework that can be integrated with existing forecasting architectures, thus providing a practical approach for improving forecasting robustness and data efficiency under complex and uncertain wind conditions.</p>