Abstract: Despite notable advancements in wind speed forecasting through deep learning techniques in recent years, the adoption of globally averaged optimal loss functions often results in smoothed forecasts that overlook small-scale details. This paper introduces a wind speed forecasting approach based on deep generative models. By integrating historical data with deep generative models, this approach samples and simulates the spatial distribution of the target area, thereby enhancing the precision of forecast outputs. Experimental results from short-term hourly wind field forecasts in the Hebei region demonstrate that augmenting the U-Net forecasting model with SDEdit significantly enhances forecast accuracy and reliability compared with using the U-Net model alone. This combined model achieves lower root mean square error and higher anomaly correlation coefficient scores, particularly excelling in capturing small-scale information. These findings improve the accuracy and reliability of wind speed forecasts and provide valuable insights for the future development and refinement of similar forecasting models.