A Variational Auto-Encoder Method in Early Fault Alarm of Smart Meters

Abstract

This study presents an effective solution for detecting faults early on in smart meters. The proposed method involves using a variational auto-encoder to evaluate the operational conditions of the meters by creating a variational feature state space matrix from the power consumption data. The use of a variational auto-encoder automates the generation of a state model, which allows for accurate detection of early faults. The method works by calculating a predefined state model and the state model, and comparing it to an self-generated threshold. This enables prompt alerts when issues are detected. The proposed approach has been experimentally validated and has shown better performance compared to a comparison method that relies on high-dimensional feature space. The accuracy rate of up to 91.67% has been achieved for alarming smart meter failures.