A Feature Reduction Strategy for Enabling Lightweight Non-Intrusive Load Monitoring on Edge Devices

Non-Intrusive Load Monitoring (NILM) implies disaggregating the power consumption of individual appliances from a single power measurement point. Recent approaches use a mix of low and high-frequency features, but real-time NILM on low-cost and resource-constrained smart meters is still challenging due to the computing effort needed for feature extraction and classification. In this paper, we present a thorough survey on low, mid, and high-frequency features for enabling the deployment of NILM algorithms on edge-devices. We compare four different supervised learning techniques on different use-cases. Moreover, we developed a novel Microcontroller (MCU) based Smart Measurement Node for collecting measurements, providing computational capabilities to perform NILM on-the-edge. Experimental results demonstrate that by selecting the proper features, a robust disaggregation model for real-time load monitoring is feasible on our MCU-based meter with an accuracy of 95.99%, relying on merely 9.4kB of memory requirements and 16K MACs operation.