<p>Accurate wind speed forecasting is essential for optimizing wind energy utilization and maintaining power grid stability. However, wind speed time series exhibit pronounced multi-scale fluctuations and complex time–frequency characteristics, which remain challenging for existing forecasting models. To address these challenges, this paper proposes a Multi-scale Temporal–Frequency Network (MFTNet), a deep learning framework that integrates multi-scale temporal analysis, frequency-domain feature extraction, and adaptive attention. MFTNet consists of three components: a Multi-scale Temporal Feature Extractor (MTFE), which captures wind speed dynamics at multiple temporal resolutions via hierarchical sliding-window operations; a Frequency-domain Feature Reconstruction Module (FFRM), which performs time–frequency transformation and representation learning to recover periodic patterns and trend information that are often under-exploited in purely time-domain modeling; and an Adaptive Attention Mechanism (AAM), which selectively emphasizes salient predictive cues in long sequences based on fused temporal–frequency features rather than isolated single-domain representations. Experiments on three real-world wind speed datasets show that MFTNet consistently outperforms state-of-the-art baselines, including PatchTST and Informer, yielding average improvements of approximately 5.5% in mean squared error (MSE) and 8.0% in mean absolute error (MAE), along with consistently higher coefficient of determination (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation>) scores that validate superior goodness of fit. These results demonstrate the effectiveness of MFTNet and highlight its potential for improving wind power forecasting accuracy in practical deployments.</p>