This paper proposes a load identification approach that combines Weighted Recursive Graph (WRG) and Convolutional Neural Networks (CNN) to meet the higher identification demands of residential users. Firstly, the single-period steady-state current signal is extracted from the aggregate signal; the current is decomposed into two forms: active current and reactive current by the Fryze power theory, and then the recursive matrix is used to convert the decomposed reactive current into a two-dimensional image data, and the CNN multi-label classifier is used to extract and learn WRG image features automatically. Finally, the convolutional neural network is used to extract and learn the features of WRG images to complete the task of multi-label load identification.