作者：李忠红, 何乐生, 汪静, 李路迟, 杨航

作者单位：云南大学信息学院，云南 昆明 650500

关键词：光伏功率预测;关联分析法;因子分析;双向长短时记忆神经网络

摘要：

针对光伏功率时间序列具有波动性及预测易受环境因素影响的问题，提出一种充分挖掘气象特征的多尺度光伏功率预测模型。首先利用灰色关联分析法选取预测日的相似日，通过因子分析法对存在近似共线性的输入气象特征进行特性因子提取，然后对特性因子与光伏功率序列分别进行自适应噪声完备集成经验模态分解（complete ensemble empirical mode decomposition with adaptive noise, CEEMDAN）；将分解后的模态分量分别按排列熵值重构为高、中、低频分量矩阵，最后对不同分量矩阵分别建立双向长短时记忆神经网络预测模型（bidirectional long short-term memory, BILSTM），再将各分量矩阵预测结果叠加得到光伏功率预测值。通过实验表明，该预测模型在不同天气类型下皆有较高预测精度，在阴天条件下，平均相对误差（mean relative error, MRE）分别比BILSTM、BP、LSTM模型降低6.0380%、16.9413%、20.1712%。

Research on multi-scale photovoltaic power prediction based on full mining of meteorological information
LI Zhonghong, HE Lesheng, WANG Jing, LI Luchi, YANG Hang
Information Institute, Yunnan University, Kunming 650500, China
Abstract: Aiming at the problems that photovoltaic power time series is non-stationary and the prediction is easily affected by environmental factors, a multi-scale photovoltaic power prediction model that fully exploits meteorological characteristics is proposed.Firstly, the similar days of the forecast days are selected by the gray correlation analysis method, and the characteristic factors of the input meteorological features with approximate collinearity are extracted by the factor analysis method. Then, the characteristic factor and photovoltaic power sequence are respectively subjected to daptive noise complete integrated empirical mode decomposition (CEEMDAN),and decomposed modal components are reconstructed into high, medium and low frequency component matrices according to their arrangement entropy values. Finally, a bidirectional long-short term memory neural network prediction model (BILSTM) is established for different component matrices, and then the prediction results of each component matrix are superimposed to get the PV power prediction value.Experiments show that the prediction model has high prediction accuracy under different weather types. Under cloudy conditions, the mean relative error (MRE) is 6.0380%, 16.9413%, and 20.1712% lower than that of BILSTM, BP, and LSTM models.
Keywords: photovoltaic power prediction;gray correlation analysis;factor analysis;bidirectional long-short term memory
2022, 48(12):111-117  收稿日期: 2022-05-01;收到修改稿日期: 2022-06-18
基金项目: 国家自然科学基金资助项目(U1631121)
作者简介: 李忠红(1995-),男,云南曲靖市人,硕士研究生,专业方向为能源物联网
参考文献
[1] PILLOT B, MUSELLI M, POGGI P, et al. Historical trends in global energy policy and renewable power system issues in Sub-Saharan Africa: the case of solar PV[J]. Energy Policy, 2019, 127: 113-124
[2] 谢晓頔, 潘霄峰, 高云峰, 等. GSIG分布式光伏云缓存调度研究[J]. 中国测试, 2020, 46(6): 13-17
[3] 田壁源, 李永新, 文玉玲, 等. 光伏功率预测技术与方法[J]. 电工技术, 2019(15): 25-27
[4] FAN C,LU Y,XIAN X, et al. Fine planning method of power grid based on principal component analysis and index correlation degree[C]//2020 IEEE 4th Conference on Energy Internet and Energy System Integration, 2020.
[5] 简献忠, 顾洪志, 王如志. 一种基于双通道CNN和LSTM的短期光伏功率预测方法[J]. 电力科学与工程, 2019, 35(5): 7-11
[6] 徐先峰, 刘阿慧, 陈雨露, 等. 基于气象因素充分挖掘的BiLSTM光伏发电短期功率预测[J]. 计算机系统用, 2020, 29(7): 205-211
[7] 张雲钦, 程起泽, 蒋文杰, 等. 基于EMD-PCA-LSTM的光伏功率预测模型[J]. 太阳能学报, 2021, 42(9): 62-69
[8] 朱瑞金, 龚雪娇, 张娟娟. 基于EEMD-MPE-LSSVM的光伏发电功率预测[J]. 中国测试, 2021, 47(9): 158-162
[9] 邹文进, 郝少飞, 马刚, 等. 基于CEEMD-GA-BP神经网络的风光发电功率预测[J]. 电网与清洁能源, 2022, 38(3): 111-118
[10] 余向阳, 赵怡茗, 杨宁宁, 等. 基于VMD-SE-LSSVM和迭代误差修正的光伏发电功率预测[J]. 太阳能学报, 2020, 41(2): 310-318
[11] 王英立, 陶帅, 候晓晓, 等. 基于MIV分析的GA-BP神经网络光伏短期发电预测[J]. 太阳能学报, 2020, 41(8): 236-242
[12] 卢忠山, 袁建华. 基于EEMD-LSTM方法的光伏发电系统超短期功率预测[J/OL]. 中国测试: 1-8[2022-05-28]. http://kns.cnki.net/kcms/detail/51.1714.TB.20220228.1116.002.html.
[13] 陈弘川, 蔡旭, 孙国歧, 等. 基于智能优化方法的相似日短期负荷预测[J]. 电力系统保护与控制, 2021, 49(13): 121-127