The itemized calculation of electricity is an important part of the smart meter, which is to test the consumption of electricity of each electrical equipment in the entry line. It is of great significance to accurately predict the power companies, improve the stability and reliability of the system, formulate scheduling plans, design the rate structure of "off-peak electricity", and discover the aging and failure of equipment. A non-invasive load decomposition method based on deep cyclic convolutional neural network is proposed in this paper. Different power states of target electrical appliances are coded and the spatial and temporal characteristics of the total power of input load are extracted by circular convolutional neural network. The normalization of input data improves the speed of model training, drouput technology is used to reduce model fitting, and transfer learning technology is used to realize the power state prediction modeling of different target electrical appliances. And compared with the traditional hidden markov model. The results show that the model proposed in this paper can predict the power state of the target electrical appliance well.