The rapid and accurate forecast of wind power is crucial for power grid dispatching, because wind power time series are frequently nonlinear and noisy. Currently, typical decomposition and ensemble approaches can enhance prediction precision, but at the expense of computing efficiency. In addition, existing approaches rely primarily on a single model to forecast the deconstructed sequences, which may not correspond to the physical significance of each sequence. This research presents a wind power forecasting paradigm based on distributed decomposition-reconstruction-ensemble learning to address the aforementioned issues. Initially, the wind power data are decomposed using variational mode decomposition (VMD) to reduce data noise. Second, frequency, grey correlation degree, and Pearson coefficient are employed to cluster the decomposed data into high-frequency and low-frequency sequences in order to decrease the complexity of the model. Then, the reconstructed sequences are equally divided into 10 parts and distributed model training is performed using federated learning (FL), federated averaging (FedAvg) algorithm, and differential privacy Laplace (DPLA) method in a prediction pool composed of different machine learning models. For each sequence, an optimal machine learning model is selected and combined to form the optimal hybrid model. Finally, the final wind power forecast results are obtained by adding and summing the prediction results of each sequence. The experimental results show that the root mean square error (RMSE) of the proposed method is 10.230, mean absolute error (MAE) is 5.175, R-square (R2) is 0.990, and accuracy rate (ACC) is 0.644, which is better than all the 28 compared models. Moreover, the technique protects data privacy, solves the problem of data silos, provides a new concept for the conventional decomposition-ensemble prediction method, and is more in accordance with the physical relevance of wind power prediction.