基于综合相似日选取的SO-CNN-LSTM光伏功率预测模型研究  

RESEARCH ON SO-CNN-LSTM PHOTOVOLTAIC POWER PREDICTION MODEL BASED ON COMPREHENSIVE SIMILAR DAY SELECTION

作　　者：宋煜 许野 刘锋平 王旭 李薇 Song Yu;Xu Ye;Liu Fengping;Wang Xu;Li Wei(College of Environmental Science and Engineering,North China Electric Power University,Beijing 102206,China;Chinese Academy of Environmental Planning,Beijing 100041,China)

机构地区：[1]华北电力大学环境科学与工程学院,北京102206 [2]生态环境部环境规划院,北京100041

出　　处：《太阳能学报》2025年第4期301-312,共12页Acta Energiae Solaris Sinica

基　　金：国家自然科学基金面上项目(62073134)。

摘　　要：针对当前光伏发电功率预测的相似日选取标准单一、形状相似判定结果有误、组合预测模型的参数选取不合理导致的预测精度偏低问题,创新性地提出一种利用综合相似度选取相似日、蛇优化算法(SO)优化卷积神经网络-长短期记忆网络(CNN-LSTM)模型关键参数组合的日前光伏发电功率组合预测模型。首先使用皮尔逊相关系数法选取关键气象因素,然后使用欧式距离相似和孪生图形相似的综合相似日选取法选定待预测日的相似日和生成高质量的模型训练样本集,最终构建基于蛇优化算法的CNN-LSTM日前光伏出力组合预测模型。以春季为例,相较于单一的欧式距离相近和孪生形状相似的相似日选取方法,基于综合相似日选取的SO-CNN-LSTM预测模型的平均绝对误差(MAE)分别降低0.15和0.13;另外,与基于综合相似日选取的LSTM和CNN-LSTM两种模型相比,SO-CNN-LSTM模型在夏季、秋季和冬季的MAE值分别降低0.73和0.15、0.36和0.24,以及0.42和0.15。Aiming at the low prediction accuracy of photovoltaic power caused by the single selection standard of similar day,erroneous shape similarity measure results and irrational parameter identification of prediction model,a combined photovoltaic power day-ahead prediction model composed of synthesized similarity index,snake optimization(SO)algorithm and CNN-LSTM model was proposed innovatively.Firstly,the Pearson coefficient method was used to identify the critical meteorological factors.Then,a comprehensive similar day selection method combining the distance similarity and siamese graphic similarity was utilized to designate the similar day and generate the training sample set with the high quality.Finally,the combined photovoltaic output day-ahead forecasting model by aid of SO and CNN-LSTM was established and applied in the photovoltaic power station in Yunnan.Taking the spring season as an example,compared with two types of single similar day selection methods,the MAE value of SO-CNN-LSTM prediction model based on comprehensive similar day selection decreases by 0.15 and 0.13,respectively.In addition,compared with two models(i.e.LSTM and CNN-LSTM)based on comprehensive similar day selection,the MAE values of SO-CNN-LSTM model decreases by 0.73 and 0.15,0.36 and 0.24,and 0.42 and 0.15,respectively,in the summer,autumn and winter seasons.

关 键 词：综合相似度 蛇优化算法 卷积神经网络 长短期记忆网络 光伏发电 

分 类 号：TM615[电气工程—电力系统及自动化]