Internal information of users can be mined by non-intrusive load decomposition (NILD) to obtain load information of various electrical equipments, which enables the smart grid to have a closer connection with daily life and provides effective data for the establishment of perception layer of ubiquitous power Internet of things (UPIoT). However, there exist some problems regarding traditional NILD, for example, the input data is complicated and a lot of factors need to be considered. Besides, sampling hardware is highly demanding and identification accuracy is relatively low. In order to solve these problems, the operating state of electrical equipment is extracted by using improved iterative K-means clustering to establish characteristic data set firstly. And then, constructed by inputting data set, sequence-to-sequence one-dimensional deep convolutional neural network (1-D-DCNN) and sequence-to-sequence LSTM and Bi-LSTM network model are decomposed to mine user information. Finally, verified by the REFITPowerData, the identification accuracy of 1-D-DCNN is over 93% though it is quite time-consuming. Compared with other deep learning model and artificial neural network methods, NILD based on characteristic data set and sequence-to-sequence 1-D-DCNN show more significant information extraction and identification capabilities.