Effective feature representation is crucial for improving short-term wind power prediction accuracy. While previous studies commonly utilize a point-wise attention mechanism, this approach treats each time step of the wind power time series independently, neglecting semantic information that connects different time steps. To address this limitation, we propose a short-term wind power prediction strategy based on a patch attention mechanism. This approach segments historical time series data, encodes positions, and utilizes the Transformer's multi-head attention mechanism to calculate position weights. High-dimensional feature vectors are obtained through residual connection networks and feedforward neural networks, and future wind power is predicted using a linear layer. Case study results show that wind power prediction with the patch attention mechanism achieves higher accuracy compared to Transformer and Informer models.