The causes of uncertainty in wind farm power generation are not yet fully understood. A method for the scale division of wind power based on the Hilbert-Huang transform (HHT) and Hurst analysis is proposed in this paper, which allows the various multi-scale chaotic characteristics of wind power to be investigated to reveal further information about the dynamic behavior of wind power. First, the time-frequency characteristics of wind power are analyzed using the HHT, and then Hurst analysis is applied to analyze the stochastic/persistent characteristics of the different time-frequency components. Second, based on their fractal structures, the components are superposed and reconstructed into three series, which are defined as the Micro-, Meso- and Macro-scale subsequences. Finally, indices related to the statistical and behavioral characteristics of the subsequences are calculated and used to analyze their nonlinear dynamic behavior. The data collected from a wind farm of Hebei Province, China, are selected for case studies. The simulation results reveal that (1) although the time-frequency components can be decomposed, the different fractal structures of the signal are also derived from the original series; (2) the three scale subsequences all present chaotic characteristics and each of them exhibits its own unique properties. The Micro-scale subsequence shows strong randomness and contributes the least to the overall fluctuations; the Macro-scale subsequence is the steadiest and exhibits the most significant tendency; the Meso-scale subsequence which possesses the greatest variance contribution rate and the maximum largest Lyapunov exponent, is the dominant factor driving the fluctuation and dynamic behavior of wind power; (3) the short-term predictions of these three subsequences based on extreme learning machine (ELM) and least-squares support vector machine (LSSVM) models have validated the above analysis results, which show that the number of steps of look-ahead predictability have pursued an ordinal trend in term of the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) and the prediction error contribution rate of the Meso-scale subsequence is the maximum. Furthermore, the short-term wind power forecasting of 6-step-ahead based on the multi-scale analysis is performed by EMD-LSSVM + ELM and the normalized Mean Absolute Error (nMAE) and normalized Root Mean Square Error (nRMSE) have been decreased by 49.45% and 44.30% compared with those of LSSVM, and 37.96% and 27.12% compared with those of EMD-LSSVM, respectively. (C) 2015 Elsevier Ltd. All rights reserved.