A Multistate Load State Identification Model Based on Time Convolutional Networks and Conditional Random Fields

The increasing concern for carbon emissions requires smarter monitoring solutions for industrial load operating conditions. Nonintrusive load monitoring is one possible monitoring solution that can provide users with detailed information on the operating status of the load, thus improving energy management capabilities. We note that industrial loads have two outstanding characteristics compared to domestic loads: 1) long single-state operation times; and 2) highly directional state transfer probabilities. Considering these two characteristics, we propose the temporal convolutional network-conditional random field architecture to construct a deep learning architecture with long-term dependencies capability and probabilistic transfer modeling. The results show that the proposed architecture can achieve over 97% recognition accuracy on the industrial load dataset used, and can suppress the problem of frequent switching of recognition state results.