Energy consumption feature adaptive non-intrusive energy disaggregation via weighted sparse coding

Abstract

Toward the future smart energy consumption, non-intrusive load monitoring is a technology with great potential by realizing the energy disaggregation in a user-friendly way. However, as smart energy users, the individual energy use patterns are customized, leading to the challenges and opportunities in implementations of non-intrusive energy monitoring. In this paper, a comprehensive study focusing on handling and utilizing the customized consumption features of diverse energy users is investigated from the view of temporal characteristics. At the first stage, the temporal energy use patterns are modelled in a probabilistic way, i.e. time of use probability. Then the individual features are adaptively evolved following the unsupervised probability density evolution method, to avoid the inaccurate model descriptions for the personalized characteristics. Furthermore, the probabilistic temporal feature model is embedded in dictionary learning formulation via weighted sparse coding scheme, achieving the direct load disaggregation with temporal characteristics into consideration. By verifications on both the low voltage network simulator platform and public field measurement dataset, the proposed study is demonstrated to be effective in handling and utilizing the customized energy features. In addition to the enhanced accuracy, the proposed approach provides a feasible solution to distinguish the appliances with similar electrical signatures, which is a contribution to the smart energy consumption. • Customized energy use features are handled and utilized in NILM problem. • An adaptive and personalized learning for the individual energy use patterns. • Unsupervised temporal characteristics modelling by probability density evolution. • Usage pattern adaptive disaggregation formulation via weighted sparse coding. • Verifications on both simulation platform and field measurements.