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OA北大核心CSTPCD

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 self-attention 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.