摘要
为提高光伏功率预测的准确性,提出一种基于混合蛙跳算法(SFLA)相似日选取和改进Transformer模型的短期光伏功率预测模型——PVformer模型。基于灰色关联度计算相似因子,基于SFLA优化综合相似因子,实现对光伏相似日的选取;PVformer模型基于卷积神经网络(CNN)对光伏数据进行特征获取并降低数据维度,基于双向门控循环单元(BiGRU)提取光伏数据时序特征并对数据进行位置嵌入,基于多头自相关注意力机制寻找序列间关系,并打破信息利用瓶颈。综合相似因子最大的历史日作为预测日的相似日,选择相关性较高的特征作为模型输入,构建历史特征向量和未来气象向量输入到PVformer模型中。对比实验结果显示,PVformer模型可提高日前光伏功率预测的精度,E_(MAPE)、E_(MAE)、E_(MSE)分别达到1.526%、0.274 MW、0.134 MW^(2)。最后通过消融实验证明模型改进的有效性和必要性,具有一定的实用价值。
In order to improve the accuracy of PV power prediction,this paper proposes a short-term PV power prediction model PVformer model based on SFLA similar day selection and improved Transformer model.This article calculates similarity factors based on grey correlation degree,optimizes comprehensive similarity factors based on SFLA,and realizes the selection of photovoltaic similarity days.The PVformer model is based on CNN networks to obtain features from photovoltaic data and reduce data dimensions.The BiGRU network is used to extract temporal features from photovoltaic data and embed data positions.The multi-head autocorrelation attention mechanism is used to find inter sequence relationships and break the bottleneck of information utilization.This article selects the historical day with the highest comprehensive similarity factor as the similarity day for the predicted day,selects features with high correlation as the model input,constructs historical feature vectors and future meteorological vectors to input into the PVformer model.The results of the control experiments show that the PVformer model improves the accuracy of day-ahead PV power prediction,E_(MAPE),E_(MAE),and E_(MSE) reach 1.526%,0.274 MW,and 0.134 MW^(2),respectively.Finally,the effectiveness and necessity of model improvement are demonstrated through ablation experiments,which have certain practical value.
作者
李练兵
代亮亮
李新达
杨鹏伟
杨少波
高国强
Li Lianbing;Dai Liangliang;Li Xinda;Yang Pengwei;Yang Shaobo;Gao Guoqiang(State Key Laboratory of Reliability and Intelligence of Electrotechnical Equipment,Hebei University of Technology,Tianjin 300400,China;Zhangjiakou Power Supply Company,State Grid Jibei Electric Power Co.,Ltd.,Zhangjiakou 075000,China;Research Institute of Electric Power Science,State Grid Hebei Electric Power Co.,Ltd.,Shijiazhuang 051334,China)
出处
《太阳能学报》
北大核心
2025年第8期341-351,共11页
Acta Energiae Solaris Sinica
基金
河北省省级科技计划(20312102D)。
关键词
光伏发电
相似日选取
功率预测
PVformer模型
多头自相关注意力机制
photovoltaic power generation
similar day selection
power prediction
PVformer model
multi-head autocorrelated attention mechanism
作者简介
通信作者:李练兵(1972-),男,博士、教授,主要从事新能源发电与主动配电网技术、先进电源技术与装备、智能化生产管理平台等方面的研究。lilianbing@hebut.edu.cn。
