基于双重注意力变换模型的分布式屋顶光伏变电站级日前功率预测

Substation-level Distributed Rooftop Photovoltaic Power Day-ahead Prediction Based on Double Attention Mechanism Transformer Model

分布式屋顶光伏地理位置分散,受地理环境遮挡和多种气象因素影响,导致光伏出力特性存在差异,给变电站级分布式屋顶光伏日前功率预测造成挑战。针对上述问题,提出了一种基于双重注意力变换模型的分布式屋顶光伏变电站级日前功率预测方法。首先,基于动态时间规整算法计算分布式光伏用户出力特性间的相似度,并基于凝聚层次聚类法将其划分成若干类;然后,利用自主注意力网络学习各时间步间的时序关联特性,通道卷积注意力机制学习多特征变量间的相关性,构建日前功率预测模型;最后,将每一类日前预测结果相加,实现变电站级日前功率预测。算例结果表明所提方法在多种天气状况下,较Transformer、长短期记忆神经网络和时序卷积网络,预测精度显著提升。

Distributed rooftop photovoltaic(PV)is spread geographically and affected by geographic shading and weather factors.It causes differences in distributed PV power output characteristics,making it challenging to predict distributed rooftop PV power at the substation level accurately.This paper proposes a day-ahead power prediction method of distributed rooftop PV based on a double attention mechanism-transformer model.Firstly,the similarity between the output characteristics of distributed PV users is determined using the dynamic time warping method and classified using the agglomerative hierarchical clustering approach.Secondly,the selfattention mechanism is used to learn the temporal correlation characteristics between each time step,and the channel convolution attention mechanism learns the correlation between multiple feature variables,and a day-ahead power prediction model is constructed.Finally,the day-ahead prediction results of each class are summed up to achieve the day-ahead power prediction at the substation level.The example results show that the proposed method in this paper significantly improves the prediction accuracy compared with Transformer,long short-term memory neural network,and time series convolution network under various weather conditions.