With the acceleration of the transition to new energy systems, it is urgent to carry out in-depth research on the complex energy characteristics of multi-load users. A technology of constructing user energy characteristic label library and a user portrait method were proposed, which comprehensively considered the coupling characteristics of electric, cold and thermal multiple loads. Firstly, the high redundancy and low correlation features were eliminated by the fast correlation filtering algorithm, and the features with strong distinguishing ability were selected by the random forest and recursive feature elimination algorithm. In the clustering stage, the improved three-way adaptive density peak clustering (3W-ADPC) algorithm improved the load clustering effect by combining the adaptive neighbor search and the three-branch clustering algorithm. The empirical results show that the proposed method has dual advantages in computational efficiency and clustering accuracy, and can accurately reveal the comprehensive energy use characteristics and deep information of multi-load users, which confirms the practical value of the proposed method in the study of multi-load users’ behavior.