Low-Cost Hardware Platform for Testing ML-Based Edge Power Grid Oscillation Detectors

Abstract

This paper introduces a low-cost hardware testing platform designed to investigate the performance of a Machine Learning (ML)-based edge application developed to detect forced oscillations in power grids. The core of the ML application lies in a Convolutional Neural Network (CNN) model deployed on an ML edge device (NVIDIA Jetson TX2). The proposed platform consists of a method for real-time signal emulation using the WaveForms Software Development Kit (SDK) that defines low-voltage signals generated by Digilent’s Analog Discovery Board. The output of the signal generator is read by the Jetson board using an Analog-to-Digital Converter (ADC). Our experiments compare the performance of different ADCs when performing inference with the same CNN model. Additionally, we give an overview of the communication scheme that allows experiment automation, which is particularly useful when experiment design is time-consuming and laborious.