Aiming at the interaction between supply and demand of power grid, based on fine grained electricity consumption behavior measurement data collected by non household terminals and network behavior statistics of marketing system, the research on resident user portrait method is carried out. A user multi?source feature label system is established from three dimensions: user behavior, power consumption characteristics and consumption habits,and extraction methods of each feature label is proposed. Based on Euclidean distance and Manhattan distance, an improved k means clustering algorithm is proposed, and the improved k means clustering algorithm is used to divide the overall control clusters of power customers as the basis for precise positioning of target users.The system of feature labels and the results of the overall control cluster partition are used to synthetically portray and visualize the users. At last, 1 500 residential users in Jinji Lake demonstration area of Suzhou were used to analyze their portraits and applications.