基于CEEMDAN-GRA-PCC-ATCN的短期风电功率预测

为提高风电功率的预测精度, 提出基于数据分解和输入变量选择的短期风电功率预测方法。利用自适应噪声完备集成经验模态分解(complete ensemble empirical mode decomposition with adaptive noise, CEEMDAN)对原始风电功率和风速数据进行分解, 平缓数据波动以提取内部隐藏信息。通过排列熵算法(permutation entropy, PE)将风电功率分量简化重构以降低模型复杂度。为提升输入变量与风电功率之间的关联程度, 剔除冗杂信息, 降低输入数据维度, 结合Pearson相关系数(Pearson correlation coefficient, PCC)和灰色关联分析(grey relation analysis, GRA)对各风电重构功率分量的输入变量进行选择。最后利用基于注意力的时序卷积网络(attention-based temporal convolutional network, ATCN)对各重构功率分量进行预测, 将各预测值叠加得到最终结果。试验结果表明, 基于CEEMDAN-GRA-PCC-ATCN的短期风电功率预测方法能够提取更多风电数据内部的关键信息, 降低输入数据的维度, 强化输入变量与风电功率之间的关联性, 有效提高预测精度。

Short-term wind power prediction based on CEEMDAN-GRA-PCC-ATCN

To improve the accuracy of wind power prediction, a short-term wind power prediction method based on data decomposition and input variable selection was proposed. The complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) was used to decompose the original wind power and wind speed data, and smooth data fluctuation to extract internal hidden information. The wind power components were simplified and reconstructed by permutation entropy (PE) algorithm to reduce the model complexity. To enhance the correlation between the input variables and wind power, eliminate redundant information and reduce the dimensionality of the input data, the Pearson correlation coefficient (PCC) and gray relation analysis (GRA) were combined to select the input variables for each reconstructed wind power component. The attention-based temporal convolutional network was used to predict the reconstructed power components, and the predicted values were superimposed to obtain the final result. The experimental results showed that the short-term wind power prediction method based on CEEMDAN-GRA-PCC-ATCN could extract more internal key information of wind power data, reduce the dimension of input data, strengthen the correlation between input variables and wind power, and effectively improve the prediction accuracy.