Temporal feature adaptive non-intrusive load monitoring via unsupervised probability density evolution

Abstract

Toward the smart power and energy consumption, non-intrusive load monitoring is emerging as the promising technical assistance of intelligent energy user. The load behaviors of individual power users are distinct, that is potential to enhance the monitoring performance if effectively addressed. In this paper, the customized temporal behaviors are thoroughly investigated and utilized for load disaggregation from the view of time characteristics. At the first stage, the temporal features of appliance usage are formularized via customized time of use probability, and the model is adaptive for the specific user habit via unsupervised probability density evolution method. Then, a generic two-stage load disaggregation framework is proposed, where the primary stage is formulized by dictionary learning and for basic load disaggregation, and the secondary stage is integrated with probabilistic temporal weights and for optimal disaggregation decision. Lastly, the sparse coding principle and risk analysis theory are employed for the robust problem solution. By comprehensive verifications on low voltage networks simulator, it is demonstrated that the proposed approach is effective in temporal load feature modeling, and thereby achieving the higher accuracy and flexibility for the non-intrusive load monitoring problem.

Keywords

• Dictionary learning
• Energy consumption disaggregation
• Risk analysis
• Time of use probability
• Unsupervised learning