摘要
为提高风速序列预测精度,提出一种基于两步分解的短期风速组合预测模型,首先使用鲁棒经验模态分解(REMD)将风速数据分解为不同频率的子序列,然后将REMD分解得到的高频模态分量使用小波包分解(WPD)进行第二步分解,降低风速序列不稳定性,提高其可预测性。其次对分解得到的高频子序列建立长短期记忆神经网络(LSTM)预测模型,低频子序列建立差分自回归移动平均模型(ARIMA)预测模型。最后叠加子序列预测结果得到风速预测结果。通过两组不同风速数据集的实验对该模型的性能进行科学评估,模型预测结果的平均绝对误差分别为0.3026、0.1255;均方根误差分别为0.498、0.1607。与其他几种对比预测模型相比,验证该模型具有一定的优越性。
To improve the accuracy of wind speed series prediction,a combined short-term wind speed prediction model based on twostep decomposition is proposed.The wind speed data are first decomposed into subseries of different frequencies using robust empirical modal decomposition(REMD),and then the high-frequency modal components obtain from REMD decomposition are decomposed in a second step using wavelet packet decomposition(WPD)to reduce the wind speed series instability and improve its predictability.Next,a long short-term memory neural network(LSTM)prediction model is built for the decomposed high-frequency subsequences,and a differential autoregressive moving average model(ARIMA)prediction model is built for the low-frequency subsequences.Finally,the wind speed prediction results are obtained by superimposing the subseries prediction results.The performance of the model is scientifically evaluated through experiments with two different wind speed datasets,and the mean absolute errors of the model predictions are 0.3026 and 0.1255.The root mean square errors are 0.498 and 0.1607,respectively.Compared with other comparative prediction models,it is proved that this model has certain advantages.
作者
陈蕻峰
王贺
李岩
熊敏
Chen Hongfeng;Wang He;Li Yan;Xiong Min(College of Technology,Beijing Forestry University,Beijing 100083,China;Energy Development Research Institute of China Southern Power Grid,Guangzhou 510623,China;Department of Electrical Engineering and Computer Science,University of Tennessee,Knoxville 37996,USA)
出处
《太阳能学报》
EI
CAS
CSCD
北大核心
2024年第2期164-171,共8页
Acta Energiae Solaris Sinica
基金
中央高校基本业务费(BLX2014-05)。
关键词
风速
神经网络
统计方法
两步分解
鲁棒经验模态分解
组合预测
wind speed
neural network
statistical method
two-step decomposition
REMD
combination prediction
