Low-Complexity Non-Intrusive Load Monitoring Using Unsupervised Learning and Generalized Appliance Models

Qi Liu, Kondwani Michael Kamoto, Xiaodong Liu, Mingxu Sun, Nigel Linge

IEEE Transactions on Consumer Electronics · 2019 · 136 citations · 35 references

DOIFull text

Open access

Concepts

TL;DR

Awareness of electric energy usage yields societal and economic benefits, and recent interest in load monitoring—particularly unsupervised non‑intrusive load monitoring—has grown as a key area for practical energy‑saving solutions. The paper proposes a low‑complexity unsupervised NILM algorithm aimed at practical deployment. The algorithm builds on a fuzzy clustering method called entropy‑index‑constraints competitive agglomeration, refined to generate generalized appliance models for household appliance detection. Experimental results on the reference energy disaggregation dataset show the method outperforms recent state‑of‑the‑art unsupervised NILM approaches in accuracy, precision, recall, F‑measure, and total energy correctly assigned.

Abstract

Awareness of electric energy usage has both societal and economic benefits, which include reduced energy bills and stress on non-renewable energy sources. In recent years, there has been a surge in interest in the field of load monitoring, also referred to as energy disaggregation, which involves methods and techniques for monitoring electric energy usage and providing appropriate feedback on usage patterns to homeowners. The use of unsupervised learning in non-intrusive load monitoring (NILM) is a key area of study, with practical solutions having wide implications for energy monitoring. In this paper, a lowcomplexity unsupervised NILM algorithm is presented, which is designed toward practical implementation. The algorithm is inspired by a fuzzy clustering algorithm called entropy index constraints competitive agglomeration, but facilitated and improved in a practical load monitoring environment to produce a set of generalized appliance models for the detection of appliance usage within a household. Experimental evaluation conducted using energy data from the reference energy data disaggregation dataset indicates that the algorithm has out-performance for event detection compared with recent state-of-the-art work for unsupervised NILM when considering common NILM metrics, such as accuracy, precision, recall, F-measure, and total energy correctly assigned.

References

35