<p>Efficient wind power forecasting is essential for the stability of renewable-dominated power systems and grid scheduling. However, the inherent stochasticity of wind speeds and the complex influence of multiple covariates pose significant challenges in simultaneously capturing multi-scale temporal dynamics and nonlinear variable interactions. To address these issues, we propose W-MixerTA, a wavelet-enhanced mixer with twin-stage attention. The model first integrates a wavelet-based multi-scale convolution (WMC) module, which hierarchically decomposes wind power sequences into multi-resolution sub-bands to enrich the representation of both fine-grained fluctuations and global trends. A time-channel decoupled Mixer is then introduced to independently model long-range temporal dependencies and multivariate correlations, thereby preventing information dilution through explicit feature disentanglement. Building on this, a twin-stage attention mechanism is designed to prioritize target-specific dynamics via self-attention while dynamically fusing covariate influences through target-guided cross-attention. Extensive evaluations on the SDWPF dataset show that W-MixerTA reduces average MSE by 2.5% over the strongest MLP-based baseline and by 5.6% over the competitive Transformer variant Informer. In long-horizon forecasting, the model exhibits a more gradual error accumulation than Transformer architectures, maintaining lower RMSE at 288 steps in all tested turbines.</p>