基于改进降噪自编码器和多元时序聚类的海上风电功率预测

OFFSHORE WIND POWER PREDICTION BASED ON IMPROVED DENOISING AUTO-ENCODER AND MULTIVARIATE TIME SERIES CLUSTERING

针对海上数值天气预报(NWP)精度低、气象因素复杂等特点,提出一种基于改进的双向降噪自编码器(BDAE)和多元时序聚类的短期海上风电功率预测方法。首先,利用Toeplitz矩阵逆协方差聚类(TICC)进行风况相似性分类,即根据30、70、100 m海上NWP风速进行多元序列实时分割与聚类;然后,针对不同风况类型分别建立可提取过去、未来双向有效信息的改进BDAE修正模型,以修正轮毂高度风速的预测误差;最后,基于修正后的轮毂高度风速以及其余NWP数据,利用TICC算法划分气象相似类型,并在此基础上建立对应的海上风电功率预测模型。采用国内某海上风电场数据进行实验验证,结果表明所提方法能提高海上风电功率预测精度,具有一定工程实用价值。

Aiming at the characteristics of low precision and complex meteorological factors of offshore numerical weather prediction(NWP),this paper proposes a short-term offshore wind power forecasting method based on improved bi-directional denoising auto-encoder(BDAE)and clustering of multivariate time series.Firstly,toeplitz inverse Covariance-based clustering(TICC)is used to classify the similarity of wind conditions,that is,the multivariate series are segmented and clustered in real time according to the wind speed of offshore NWP at 30,70,and 100 m.Secondly,according to different wind conditions,an improved BDAE correction model which can extract the past and future bidirectional effective information is established to correct the prediction error of hub height wind speed.Finally,based on the modified hub height wind speed and other NWP data,the TICC algorithm is used to classify meteorological similarity types,and the corresponding offshore wind power prediction model is established on this basis.The experimental verification is carried out with the data of an offshore wind farm in China,and the results show that the proposed method can improve the accuracy of offshore wind powerforecasting,which have certain engineering practical value.