基于电压特征谱聚类和EMA季节趋势分解网络的配变电压预测OA

Aimed at the problems such as a large number of distribution transformers in a region,complex voltage char-acteristics of distribution transformers and a low prediction accuracy of distribution transformer voltage,a distribution transformer voltage prediction framework based on voltage characteristic spectrum clustering and exponential moving av-erage(EMA)seasonal trend decomposition network is proposed.First,the voltage characteristics on the daily time scale of a distribution transformer are extracted to construct a voltage characteristic matrix,and the clustering of distribu-tion transformers with different voltage characteristics is realized using the spectrum clustering algorithm.Second,an EMA seasonal trend decomposition network(EMA-STDNet)is constructed,which uses the EMA method to smooth the voltage sequence to achieve seasonal trend decomposition.This model constructs a Transformer encoder improved by sparse cross-attention mechanism to extract long-term trend features,as well as a combination module of depthwise sepa-rable convolution and convolution block attention module based on patch embedding to extract short-term seasonal fea-tures,and it realizes prediction through the full-connection layer.Finally,this method is applied to the distribution trans-former voltage data set in a certain area.Experimental results show that the proposed method can effectively character-ize the daily voltage operation of various types of distribution transformers.Compared with those of other deep learning models,the prediction accuracy of typical distribution transformer voltage is improved by more than 5%,indicating that this method can effectively distinguish various types of distribution transformers and improve the prediction accuracy of distribution transformer voltage.