Original research

ANOMALY DETECTION IN SMART GRID STABILITY USING CONVOLUTIONAL VARIATIONAL AUTOENCODER

Pages 9-16

10.61552/JEMIT.2027.01.002

Benjamin Essilfie-Nyame, Robert Agyare Ofosu, Philip Blewushie

• View Article
• PDF 1085 K

Abstract Ensuring a continuous supply of electricity, at the expense of system failures, requires diligent maintenance of electrical grid stability. Although traditional rule-based and Machine Learning (ML) methods seem promising, they often struggle when relatively high-dimensional data is involved. Moreover, supervised Deep Learning methods are often limited by the lack of sufficient labeled anomaly samples. This study presents an unsupervised anomaly detection framework that uses a Convolutional Variational Autoencoder (CNN-VAE) to learn the distribution of normal grid operating conditions. In an experiment conducted on an electrical grid dataset, the CNN-VAE achieves better results than other unsupervised baselines, with F1-score (0.897) and AUC (0.920). These results suggest that this method can be considered for real-time electrical grid stability monitoring in smart grid systems.

Keywords: Anomaly detection, Smart grid, Unsupervised learning, Machine learning, Variational Autoencoder.

Received: 27.07.2026 Revised: 03.09.2026. Accepted: 11.10.2026.