基于CEEMDAN-SE-GWO-LSTM模型的短期风速预测  

Short-term Wind Speed Prediction Based on the CEEMDAN-SE-GWO-LSTM Model

作　　者：王胜研 王娟娟[1] WANG Shengyan;WANG Juanjuan(Dalian Jiaotong University,Dalian 116028,China)

机构地区：[1]大连交通大学,辽宁大连116028

出　　处：《电工技术》2024年第4期74-78,81,共6页Electric Engineering

摘　　要：为了降低风速具有的非线性和随机性带来的预测难度,提高预测准确性,提出一种融合完全自适应噪声集合经验模态分解(CEEMDAN)、样本熵(SE)、灰狼优化算法(GWO)和长短期记忆神经网络(LSTM)的组合预测模型来预测短期风速。首先利用CEEMDAN将风速数据分解为若干模态分量,再通过样本熵对各分量进行筛选,将样本熵值相近的模态分量进行叠加,形成新的若干个子序列,然后对各子序列采用GWO-LSTM模型进行训练与预测,最后叠加子序列的预测结果。实验结果表明,所提CEEMDAN-SE-GWO-LSTM模型相对于单一的LSTM模型在均方根误差、平均绝对误差和平均相对误差这3个误差指标上分别降低了21.7%、44.5%和40.9%,因此该模型具有较好的预测精度与稳定性,可有效预测短期风速。In order to reduce difficulty of wind speed prediction due to its nonlinearity and randomness and improve prediction accuracy,the present work proposed a combined prediction model integrating fully adaptive noise ensemble empirical mode decomposition(CEEMDAN),sample entropy(SE),gray wolf optimization algorithm(GWO),and long short-term memory neural network(LSTM)to predict short-term wind speed.The method works by using first CEEMDAN to decompose wind speed data into several modal components,second SE to screen the components and to superimpose modal components with similar SE values which form new subsequences,and third GWO-LSTM model for training and prediction of each subseries whose results of are finally superimposed.The proposed CEEMDAN-SE-GWO-LSTM model was proved by simulative experiment capable of achieving reduction in root mean square error,mean absolute error,and average relative error by 21.7%,44.5%,and 40.9%,respectively,compared with isolated LSTM model,and therefore satisfactorily accurate,stable,and effective in predicting short-term wind speed.

关 键 词：风速预测 CEEMDAN SE GWO 长短期记忆神经网络 

分 类 号：TM614[电气工程—电力系统及自动化]