Against the backdrop of global energy transition and large-scale grid integration of wind power, ultra-short-term wind power prediction is crucial for ensuring power grid security and the efficient operation of wind farms. However, the inherent nonstationarity of wind power sequences stems from the complex coupling of time-domain transient disturbances and frequency-domain periodic fluctuations, posing two core challenges for high-precision prediction: modelling the time-frequency coupling effect and capturing multi-scale dynamic features. Existing deep learning methods often struggle to characterise this cross-scale dynamic interaction mechanism simultaneously and effectively. To address this, this study proposes a wind power prediction method based on Dual-Stream(DS) time-frequency feature decoupling modelling and hierarchical dynamic fusion. First, the technique explicitly decouples and collaboratively learns time-domain dynamic and frequency-domain periodic features through a dual-stream architecture. Furthermore, it introduces a Multi-Scale Convolutional-enhanced Transformer(MS-Convformer) that integrates local convolution and global self-attention to capture second- to minute-level turbulent fluctuations and long-period trends. Finally, through a hierarchical dynamic fusion strategy guided by a gating mechanism and cross-attention, it achieves adaptive weighting and structured integration of multi-scale features. Experimental results on real wind farm datasets show that, compared with mainstream deep learning methods, the DS-MS-Convformer method proposed in this paper achieves significantly higher prediction accuracy, providing adequate technical support for grid integration and renewable energy consumption.