基于残差的分布式光伏发电功率组合预测方法

Skip-based combined prediction method for distributed photovoltaic power generation

分布式光伏发电功率预测在保障电网运行安全和就近消纳方面发挥着重要作用,为提升分布式光伏发电功率预测精度,提出一种基于多元气象的特征提取方法和基于残差连接的多模型融合的光伏发电功率预测模型.在特征提取时,引入统计、交叉、周期信息、近似熵和光伏板温度等特征提取方法,实现对时间、气象和发电功率的深层特征提取,丰富模型的输入.在模型构建时,建立基于残差连接的多层模型融合方法,首先提出基于k最近邻(k-nearest neighbor,kNN)的softmax回归预测模型,其次设计3层模型整体结构,并通过残差连接和多层堆叠的方式融合多个预测模型,持续提升光伏发电功率预测精度.基于电力公司真实数据,采用本研究方法与随机森林(random forest,RF)、TabNet和极端梯度提升(extreme gradient boosting,XGBoost)等模型,对光伏发电功率进行预测.结果表明,所提模型在均方根误差、平均绝对误差、均方误差和平均绝对百分比误差等方面可分别降低0.1097、0.0591、0.0507和0.0368,拟合优度可提升0.0804.基于多元气象的特征提取方法和基于残差连接的多模型融合的光伏发电功率预测模型能有效提升分布式光伏发电功率预测的精度和稳定性.

Distributed photovoltaics(PV)power generation forecasting plays an important role in ensuring the safety of power grid operation and nearby consumption.In order to enhance the accuracy of distributed PV power generation forecasting,we propose a meteorological feature extraction method and futher design a PV power generation forecasting model based on skip-connect models fusion.In the feature extraction,we use statistical analysis,features cross-correlation,periodicity information,approximate entropy,and the temperature of PV panels to achieve deep feature extraction of time,weather,and power generation data,enriching the model inputs.In model construction,we propose a multi-layer model fusion method based on residual connections.Firstly,we introduce a knearest neighbor(kNN)-based softmax regression prediction model.Secondly,we design a three-layer model structure with multiple prediction models fused through residual connections and multi-layer stacking,continuously impg the prediction accuracy of PV power generation forecasting.Based on the real data of electric power companies,we compare the proposed method with others,such as random forest(RF),TabNet and extreme gradient boosting(XGBoost)for photovoltaic power generation prediction.The results show that the proposed model can reduce the root mean square error,mean absolute error,mean squared error,and mean absolute percentage error by 0.1097,0.0591,0.0507,and 0.0368 respectively,and improve the goodness of fit by 0.0804.The feature extraction method based on multi-meteorological factors and the photovoltaic power generation prediction model based on residual connections for multi-model fusion effectively improve the accuracy and stability of distributed PV power generation forecasting.