基于云图相位相关和低秩特征提取的超短期辐照度预测  

ULTRA-SHORT-TERM IRRADIANCE PREDICTION BASED ON CLOUD IMAGES PHASE CORRELATION AND LOW-RANK FEATURE EXTRACTION

作　　者：郭吉强 徐旭东 张帅 任密蜂 王芳 阎高伟 Guo Jiqiang;Xu Xudong;Zhang Shuai;Ren Mifeng;Wang Fang;Yan Gaowei(College of Electrical and Power Engineering,Taiyuan University of Technology,Taiyuan 030024,China;Shanxi Gemeng Sino-US Clean Energy R&D Center Co.,Ltd.,Taiyuan 030031,China)

机构地区：[1]太原理工大学电气与动力工程学院,太原030024 [2]山西格盟中美清洁能源研发中心有限公司,太原030031

出　　处：《太阳能学报》2025年第3期494-503,共10页Acta Energiae Solaris Sinica

基　　金：国家自然科学基金(61973226);格盟集团科技创新基金(2022-05);山西省自然科学基金(20210302123189);新型电力系统运行与控制全国重点实验室开放基金(SKLD22KM22)。

摘　　要：针对云的动态变化对辐照度的直接影响,基于傅里叶相位相关(PC)理论从连续地基云图获取云的动态信息矩阵(DIM),为了减小噪声及冗余信息对模型的影响,构建低秩矩阵特征提取(LMFE)模型作为提取模块(EM),实现DIM有效信息的提取,EM通过迭代训练得到最优特征提取矩阵W和偏置项b。同时,以双向门控循环单元(Bi-GRU)为基础搭建回归模块(RM),RM共享EM输出W和b,从云图提取的有效信息结合历史辐照度序列作为输入,实现辐照度的多步回归预测。对真实采集的数据进行实验仿真,结果表明该预测方法可有效提高辐照度预测精度。Aiming at the direct impact of dynamic changes of the clouds on irradiance,this paper obtains the cloud dynamic information matrix(DIM)from continuous ground-based cloud images based on the Fourier phase correlation(PC)theory.To minimize the effects on the model led by noise and redundant information,a low-rank matrix feature extraction(LMFE)model is constructed as an extraction module(EM)to extract valid information from the DIM productively.EM obtains the optimal feature extraction matrix W and bias terms b through iteration.Simultaneously,the regression module(RM)is constructed based on the bidirectional gated recurrent unit(Bi-GRU),where RM shares EM outputs W and b,and the valid information extracted from the cloud images is combined with the historical irradiance sequence as input to achieve multi-step regression prediction of irradiance.The experimental simulation of the actual collected data demonstrates that the prediction method can effectively improve the irradiance prediction accuracy.

关 键 词：太阳辐射 预测 特征提取 图像相关 神经网络 

分 类 号：TM615[电气工程—电力系统及自动化]