<p>In response to the insufficient feature mining of the data of wind power in current wind power prediction methods, this paper proposes a zoning collaborative power prediction method for large-scale wind farm cluster based on multi-scale component reconstruction and fusion technology. Firstly, by improving the deep embedded graph attention clustering network, a topology map is constructed based on the geographical location information of the wind farm, guiding the numerical weather prediction (NWP) wind speed to divide the large-scale wind farm cluster into several sub-clusters. Then, the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) algorithm is used to decompose the wind power and NWP of each sub-cluster into several sub-components and the permutation entropy of each sub-component is calculated. The permutation entropy is then clustered by the Kmeans++ clustering algorithm to further reconstruct the sub-components of wind power and NWP wind speed into low-frequency and high-frequency components. Next, the TimesNet-Informer model is used to predict the low-frequency and high-frequency components of each sub-cluster, and finally the wind power prediction result of large-scale wind farm cluster is obtained. Through experimental verification, the method proposed in this paper achieved an average reduction of 0.02268 in RMSE and 0.02327 in MAE compared to the comparative algorithm as well as R2 and AR increased by an average of 0.0367 and 2.27%, respectively. which is superior to comparative algorithms and it reduces the scheduling difficulty of large-scale wind farm cluster.</p>