Short-term Wind Power Forecasting Based on CEEMDAN-FE-KELM Method

Based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN)-fuzzy entropy (FE) and an extreme learning machine with kernels (KELM), we propose a combined forecasting method for short-term wind power forecasting. The CEEMDAN method adds a particular white noise at each stage of the decomposition and computes a unique residue to obtain each stage's intrinsic model function (IMF). Compared with the EEMD method, the decomposition process of the CEEMDAN is complete. In order to weaken the influence of the signal's non-stationary effects on the prediction accuracy and to reduce the computational scale, we use the CEEMDAN-FE method to decompose the original signal into a series of subsequences with obvious differences in their degree of complexity. Then, we build the corresponding KELM forecasting model. Finally, we combine these forecasting results to output the final forecasting result. We applied the proposed CEEMDAN-FE-KELM method to a short-term wind power forecasting situation in one area. Under the same conditions, a comparison of the results using the single KELM method with those from the combined KELM-based forecast model shows the proposed method to be more effective.