基于谱聚类和优化极端学习机的超短期风速预测

Ultra-Short-Term Wind Speed Prediction Based on Spectral Clustering and Optimized Extreme Learning Machine

较高精度的超短期风速预测是并网运行风电场风电功率预测预报系统建立和运行的必要前提及保证。由于风速影响因素众多,具有较大的波动性和随机性,并具有高度的自相关性,给传统的风速预测方法带来了极大的挑战。提出一种基于谱聚类和极端学习机的超短期风速预测方法。该方法首先利用小波变换和主成分分析对风速数据进行去噪和降维处理,剔除数据的不规则波动,有效降低数据维度;然后分别应用谱聚类对小波变换后的各分解序列进行聚类分析,减少训练样本空间,提高样本有效性,降低计算复杂度;再应用极端学习机对各分解序列分别进行训练,同时通过遗传算法对极端学习机输入权值、偏置等参数进行优化,确保各分解序列输出最佳预测模型;最后将各分解序列预测结果相加得到最终预测结果。以某风电场实际数据进行的建模结果表明该模型有效实现了对风速的超短期、多步预测,采用的方法合理有效。

The ultra-short-term wind speed prediction with higher accuracy is the necessary premise and assurance for the establishment and operation of wind power forecasting system for grid-connected wind farm. Numerous factors impacting on wind speed, which possess evident fluctuation property and randomness as well as high degree of self-correlation, bring huge challenge to traditional wind speed prediction method. A spectral clustering （SC） and extreme learning machine （ELM） based ultra-short-term wind speed prediction method is proposed. Firstly, the wavelet transform and principal component analysis are utilized for the denoising and dimensionality reduction of original wind speed data to get rid of irregular fluctuation of original data to effectively reduce the dimensionality of original data; secondly, the SC is applied to each wavelet-transformed decomposed sequence respectively to perform clustering analysis to decrease the space of training samples and improve the effectiveness of samples, thus the calculation complexity is reduced; thirdly, the ELM is applied to each decomposed sequence to carry out the training respectively, meanwhile~ by means of genetic algorithm the parameters input to ELM such as weights and bias of the hidden layer are optimized to ensure that each decomposed sequence can output optimal predicted model respectively; finally, the predicted results of each decomposed sequences are added to obtain final prediction results. The modeling result based on actual data of a certain wind farm shows that the ultra-short-term and multi-step prediction of wind speed can be effectively realized by this model, and the proposed method is reasonable and effective.