<p>Accurate wind turbine power prediction (WPP) is especially important for grid dispatching and grid integration. However, the high randomness and intermittency of wind resources make wind turbine (WT) have complex spatial–temporal dynamics and hierarchical characteristics. Most existing studies rely on combinatorial methods, which cannot accurately predict WT power. To realize accurate WPP, this paper proposes a novel model called multi-scale spatial–temporal interaction network (MSTINet). The model constructs a hierarchical spatial–temporal feature extraction framework to process subsequences with different scales to fully extract the hierarchical characteristics of WT. In each layer, MSTINet utilizes the interactive learning strategy that allows each subsequence to have both local and global views. The interactive learning strategy, temporal and spatial feature extraction module are combined to fully extract local and global information and complex spatial–temporal dynamics. Subsequently, we use the slight convolutional block attention module to highlight dynamic changes in the sequence and capture deeper spatial–temporal dynamics. The experiment results show that MSTINet achieves a maximum <i>R</i>² of 0.9799, a 15.74% performance improvement compared to baseline methods. Additionally, the RMSE is as low as 0.0298, indicating a significant enhancement in performance. The highest prediction accuracy was achieved in prediction tasks across multiple time steps.</p>