<p>The prediction of wind speed and direction is crucial in fields such as weather forecasting, wind energy resource development, and aviation safety. Due to the complex terrain and unique meteorological conditions in southwest China, the changes in wind speed and direction in the region are difficult to predict. With its powerful ability to represent spatiotemporal data, spatiotemporal graph neural networks have become a powerful tool to solve this problem. However, the current spatiotemporal graph neural networks still have the following shortcomings: 1. Complex terrain conditions and meteorological phenomena lead to drastic wind changes, and models that rely on static graphs struggle to capture these dynamic characteristics. 2. Relying solely on historical data for prediction overlooks the potential value of future meteorological forecast data, making it difficult to fully utilize forward-looking information to improve prediction accuracy. 3. Using a single-scale model, it is difficult to flexibly capture meteorological data with multi-time scale characteristics. To address these issues, a multi-scale spatiotemporal model is proposed. First: Combine the dynamic wind field changes and the static site spatial distance to fully capture the dynamic characteristics and local spatial characteristics of meteorological data that change over time. Additionally, by introducing future meteorological forecast data and combining historical meteorological data, the training sample data is further enhanced. Next, by introducing a dynamic multi-scale sampling method, meteorological data at different time scales are sampled. Finally, multiple experiments verified that the new model has significantly improved the prediction accuracy and stability.</p>