Conservative Power Theory Approach for Feature Extraction in Non-Intrusive Load Monitoring

Non intrusive load monitoring offers a cost-effective alternative traditional energy monitoring by disaggregating a household's total energy consumption into individual appliance loads from a single measurement point. This technology enhances energy conservation, demand-side management, and smart grid development However, there are many electric appliances exhibit similar operational characteristics because they share similar front end or have similar core technology. This is one of the most challenging and important task in NILM because there is an increasing trend in electric appliances to use active front ends for better energy efficiency. In this paper, we demonstrate that feature extraction based on conservative power theory (CPT) performs better, compared to Fryze's method in discerning loads with similar core technology or sharing the same front ends. Incorporating a feature extraction layer based on CPT into a deep learning neutral network can achieve an F1 score of 0.946, compared to those based on Fryze's method and the Vanilla case (i.e. no feature extraction layer), which correspond to 0.875 and 0.685, respectively to identify the turn-on and turn-off status of a total of 1023 load combinations.