An Instrumental High-Frequency Smart Meter with Embedded Energy Disaggregation.

Area of Science:

• Electrical Engineering
• Computer Science
• Energy Systems

Background:

• Current smart meters often rely on low sampling rates and cloud-based processing for energy disaggregation.
• Transmitting high-frequency data from meters to the cloud presents challenges in bandwidth, latency, and privacy.

Purpose of the Study:

• To develop and evaluate a prototype smart meter capable of local, high-frequency energy disaggregation using embedded deep learning.
• To assess the impact of sampling frequency on model accuracy and edge device performance.
• To introduce novel metrics for quantifying non-intrusive load monitoring (NILM) efficiency on edge devices.

Main Methods:

• Designed a smart meter prototype with a custom signal conditioning circuit and an embedded board.
• Implemented a deep learning model for energy disaggregation directly on the edge device.
• Evaluated the prototype's accuracy, power consumption, throughput, and latency across six different embedded hardware platforms.
• Introduced and applied three hardware-aware performance metrics for NILM efficiency.

Main Results:

• The prototype successfully performed energy disaggregation locally at a high sampling frequency (15 kHz).
• Analysis revealed the trade-offs between sampling frequency, model accuracy, and edge device power consumption.
• Benchmarking across platforms provided insights into latency and throughput variations.
• Novel metrics offered a standardized way to evaluate NILM edge device performance.

Conclusions:

• Local, high-frequency energy disaggregation on smart meters is feasible and offers advantages over cloud-based approaches.
• The developed architecture enables compact and energy-efficient NILM-enabled edge meters.
• The hardware-aware metrics provide a valuable framework for future development and comparison of NILM edge devices.