Short-Term Wind Power Prediction Based on Data Reconstruction and Improved Extreme Learning Machine

With the purpose of improving the accuracy of the wind power short-term forecasting in an effective way, improved wavelet threshold denoising and principal component analysis (PCA) are applied to denoise and reduce the dimension of the original wind power data, the wind power data are reconstructed, and the quality of the information is effectively improved. Next, with the aim of decreasing the nonstationarity of the wind power signal, the reconstructed signal is decomposed in the frequency domain by ensemble empirical mode decomposition (EEMD). In terms of the problem that the extreme learning machine (ELM) algorithm has low stability and accuracy when the quantity of hidden layer neurons is randomly determined, particle swarm optimization algorithm improved (IPSO) by differential evolution (DE) algorithm is introduced to improve the ELM, which is then used to build prediction models for wind power signals in different frequency domains, respectively, and a lower prediction error is achieved. Then, the Markov chain is employed to construct the dynamic combination prediction model, and the weights of different subsignal prediction models are adaptively determined. Compared with adding the prediction values of different subsignals directly, the higher prediction accuracy is demonstrated by the prediction scheme proposed in this paper. At last, actual wind farm operation data are applied to conduct a simulation experiment, and the superiority of the introduced prediction scheme is verified.