Advances in artificial intelligence and intelligent algorithms have driven the evolution of wind power forecasting toward spatio-temporal collaborative modeling and distributed learning. This paper proposes a wind farm cluster power forecasting model that integrates ST-GCN and FedFormer. Based on wind farm cluster graph modeling, ST-GCN is used to extract spatially coupled and local temporal features, while FedFormer is employed to enhance the representation of long-term trends and frequency-domain information. Collaborative training is performed within a federated framework.Experimental results show that on the validation set, the model achieves MAE, RMSE, and MAPE of 16.87%, 23.14%, and 6.38%, respectively, outperforming FedFormer’s 18.21%, 24.97%, and 6.95% and ST-GCN’s 18.74%, 25.86%, and 7.21%, demonstrating higher accuracy and stability.
