Abstract: To cope with the problems of traditional ultra-short-term wind power prediction methods in terms of data potential relationship mining and model convergence speed, an ultra-short-term wind power prediction method based on transform domain analysis and extreme gradient boosting (XGBoost) is proposed. First, time-sliding windows and wind power indicators are utilized to perform data construction and low-level feature extraction. Then, the wind power data are decomposed by the multilevel transform domain analysis method which is composed of fast Fourier transform (FFT) and Haal wavelet transform, and the importance of frequency domain information in feature learning is fully considered. Finally, an ultra-short-term wind power prediction model is established with FFT, Haal wavelet transform, and the XGBoost algorithm. The experimental results show that the multilevel transform domain analysis method can be adopted to fully explore the potential relationship between the original features and capture the temporal correlation of the data in depth. The XGBoost algorithm can be adopted to effectively improve the prediction performance of the model. The proposed method shows high prediction accuracy and excellent feature extraction ability on different data sets compared to other prediction models.