<p>Wind power generation datasets are widely encountered in renewable energy systems because of their inherent periodicity, non-stationarity, and substantial noise. Recent advancements in graph neural networks have improved wind power time series forecasting via multi-scale feature extraction. However, a key challenge is effectively capturing multi-scale temporal trends, suppressing intrinsic noise, and leveraging correlations among wind power variables. We propose a novel architecture: Multi-scale coupling graph neural network with enhanced simple attention (MultiGSA). In contrast to traditional approaches that fail to address the instability, high sampling frequency, and intrinsic noise of wind power data, MultiGSA extends feature learning across multi-scale temporal dynamics and cross-variable correlations. It uses a cross-scale Co-GNN module to capture multi-scale trends and suppress time-domain noise, leveraging inter-scale complementary information. The cross-variable Co-GNN capitalizes on variable homogeneity and heterogeneity to strengthen correlation detection, whereas the enhanced simple attention mechanism alleviates frequency-domain noise. We evaluate MultiGSA on four real-world wind farm datasets, and the results show that it consistently outperforms state-of-the-art methods. Specifically, MultiGSA achieves up to 4.8% MAE and 3.2% RMSE improvements over the baselines, demonstrating its effectiveness and robustness in managing wind power data complexity.</p>