基于CEEMDAN-精细复合多尺度熵和Stacking集成学习的短期风电功率预测

CEEMDAN-refined composite multiscale entropy and stacking ensemble learning-based short-term wind power prediction

为了解决风电功率的间歇性与非平稳性带来的功率预测难度,提出了一种基于CEEMDAN-精细复合多尺度熵和Stacking集成学习的短期风电功率预测方法。在对风电功率进行预测之前,对风电功率数据进行预处理。首先引入自适应噪声完备集合经验模态分解(CEEMDAN)方法分解风电功率原始序列,并计算各分解分量的精细复合多尺度熵(RCMSE)。然后,将熵值相近的分量序列重组成新序列,以降低模型复杂度和提高计算效率。在预测阶段,对重组之后的序列分别建立Stacking集成学习模型进行风电功率短期预测,最后对预测结果进行重组。通过新疆某风电场实测数据证明:结合各单一预测模型优点的Stacking集成学习模型方法与其4种基学习器KNN、RF、SVR和ANN相比,Stacking模型具有更高的风电预测准确度。在同等条件下,CEEMDAN-RCMSE-Stacking模型均方根误差相比单一的Stacking模型及EMD-RCMSE-Stacking模型分别减少了20.34%和9.74%,平均绝对误差分别减少了24.55%和6.35%,而拟合优度系数分别提高了4.09%和1.62%,即CEEMDAN-RCMSE-Stacking模型拥有更高的预测性能。

In order to solve the difficulty of power prediction caused by the intermittence and non-stationarity of wind power, a CEEMDAN-refined composite multiscale entropy and Stacking ensemble learning-based short-term wind power prediction method is proposed herein. Before the prediction of wind power, the wind power data is preprocessed. First, the complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN) method is introduced to decompose the original wind power sequence, and then the refined composite multiscale entropy(RCMSE) of each decomposition component is calculated. Afterwards, the component sequences with similar entropy values are reconstituted into new sequences to reduce the complexity of the model and improve the computational efficiency. In the prediction stage, the Stacking ensemble learning models are respectively established for the reorganized sequences to predict the short-term wind power, and finally the prediction results are reorganized. Through the measured data of a wind farm in Xinjiang, it is proved that the Stacking ensemble learning model method combined with the advantage of each single prediction model has higher wind power prediction accuracy than its four base learners, i.e. KNN, RF, SVR and ANN. Under the same conditions, the root mean square error of CEEMDAN-RCMSE-Stacking model is reduced by 20.34% and 9.74% with the reductions of mean absolute error of 24.55% and 6.35% respectively if compared with the single Stacking model and EMD-RCMSE-Stacking model, while the fitting goodness coefficients are increased by 4.09% and 1.62% respectively as well. That is to say, CEEMDAN-RCMSE-Stacking model has higher prediction performance.