Aiming at the large number of regional distribution transformer, a large number of new loads, distributed photovoltaics, etc., and the enhancement of the random voltage fluctuation of distribution transformer. The voltage quality of the substation users is facing challenge. In order to better analyze and predict the over-limit characteristics of regional distribution transformer voltage, a bilayer clustering regional distribution transformer voltage prediction method based on correlation feature screening is proposed. Firstly, the number of overrun days of regional distribution transformers are taken as the first layer clustering feature, and the distribution transformers with normal and over-limit voltage properties are obtained. Secondly, for the over-limit voltage distribution transformers, an optimal metric matrix combining Pearson′s correlation coefficient and Euclidean distance is proposed to extract the contained information of the original data as the input of K-means to realize the bilayer clustering of regional distribution transformer. On this basis, the representative distribution transformers in the cluster are selected to characterize the distribution transformers of this category, and the convolutional neural network-bidirectional long and short-term memory- attention（CNN-BiLSTM-Attention）model is used to predict the distribution transformer voltage, which can extract the bidirectional information features of the input data, weight the important features, and obtain the bidirectional feature information from multiple time scales for prediction. Finally, the effectiveness of the proposed method is verified in a certain area of Shanghai.