The analysis of behavior feature in smart grid users plays an important role in power marketing strategy. Based
on KL_divergence and KL transform,this paper completed feature extraction and classification in smart grid data.It achieved grid data division of different types.Moreover, by composite analysis of the daily load profile in all users,it is extracted that the typical daily load profile in different types of users. The results prove that it is realized that the compression of raw data and the retention of main feature by means of KL transform, it is greatly reduced that calculation of extraction and classification in smart grid users,so it increases the time efficiency;it is optimized that the selection of initial clustering center and cluster number by means of KL-divergence, the accuracy of clustering is improved; the normal data of grid users in this instance is 38 groups, the users is divided into 3 categories: peak electricity type，peak avoidance electricity type and part meeting peak electricity type.The results can be used to classify behavior feature of smart grid users more effectively, it will provide a technical basis for business expansion of electric power company.