Abstract: Wind power forecasting faced the dual challenges of unpredictability and randomness in wind conditions. Existing forecasting models based on graph neural networks were usually constructed from only a single perspective to build graph structures， which made it difficult to comprehensively and accurately characterize the complex spatial correlations among wind farms. To address this issue， a Dual-MSGNN model was proposed， which integrated frequency domain analysis and adaptive graph convolution to capture the dynamic correlations of wind power series across multiple temporal and spatial scales. In this method， graph structures were constructed from both temporal and spatial dimensions： on the one hand， the time series were decomposed into different scales， and a self-attention mechanism was introduced to extract intra‑sequence temporal dependencies； on the other hand， an adaptive graph convolution layer was adopted to automatically learn inter‑sequence spatial correlation patterns under each time scale. Experimental results on real‑world wind power datasets demonstrated that the Dual‑MSGNN model achieved significant advantages in prediction accuracy and robustness， thus validating its effectiveness.