为提高不同天气类型下光伏输出功率的预测精度，提出了一种基于注意力机制的超短期光伏预测组合模型。首先，通过皮尔逊相关系数分析，选取与光伏发电功率密切相关的关键气象因子，并对其进行逐月标准化，然后加权求和计算得到分类指标天气条件因子(Sky Condition Factor， SCF)，以降低输入变量的维度，并消除季节对天气分类的干扰和众多气象因子之间的耦合关系。其次，通过自组织映射神经网络(SelfOrganizing Map， SOM)对SCF进行无监督聚类，划分出3种天气类型。然后，在3种天气类型下分别构建卷积神经网络(Convolutional Neural Network， CNN)预测模型，并引入高效通道注意力模块（Efficient Channel Attention， ECA），自适应地为特征信息的多重通道分配相应的权重，使模型集中于重要的特征信息，提高模型的预测精度。采用历史实测数据进行仿真，结果表明：与〖JP2〗未引入ECA模块的预测模型相比，所提预测模型在3种天气类型下的预测准确度分别提高了1.006 1%，〖JP〗1.626 1%和1.610 4%，验证了该模型的有效性。

In order to improve the prediction accuracy of PV output power under different weather types, a combined model of ultrashortterm PV prediction based on attention mechanism was proposed. First, the key meteorological factors closely related to PV power generation were selected by Pearson correlation coefficient analysis and standardized month by month, and then weighted summation was calculated to obtain the classification index sky condition factor (SCF) in order to reduce the dimensionality of input variables and eliminate the interference of season on weather classification and the coupling relationship between numerous meteorological factors. Second, three types of weather types were classified by unsupervised clustering of SCFs through selforganizing map (SOM) neural network. Then, convolutional neural network (CNN) prediction models for each of the three weather types were constructed, and the efficient channel attention (ECA) module was introduced to adaptively assign relevant weights to each of the multiple channels of feature information extracted by the CNN, so that the model focused on important feature information and the prediction accuracy of the model was improved. Finally, simulations were performed using historical measured data. The results show that the prediction accuracies of the proposed prediction model under three different weather types are improved by 1.006 1%, 1.626 1%, 1.610 4%, respectively, compared with the prediction model without the introduction of ECA module, which verifies its effectiveness.

TP183；TM615

国家自然科学基金（62222307）