Image-Based Power Quality Disturbance Classification Using a Parameter-Efficient Convolutional Neural Network
2nd International Symposium on AI-Driven Engineering Systems, ISADES 2026, Hybrid, Mbale, Uganda, 19 - 20 Haziran 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/isades69945.2026.11608112
- Basıldığı Şehir: Hybrid, Mbale
- Basıldığı Ülke: Uganda
- Anahtar Kelimeler: classification, convolutional neural network, deep learning, Power quality disturbances
- Bursa Uludağ Üniversitesi Adresli: Evet
Özet
Power quality disturbances (PQDs) represent a growing challenge to the reliability and efficiency of modern electrical power systems, driven by the increasing penetration of renewable energy sources, nonlinear loads, and power electronic devices. Accurate and automated classification of these disturbances particularly in their compound forms is essential for effective power quality monitoring and mitigation. This study proposes a lightweight convolutional neural network (CNN) for the comprehensive classification of 17 PQD classes, encompassing 7 single disturbances, 7 dual compound disturbances, and 3 multiple compound disturbances. PQ signals were synthetically generated in MATLAB/Simulink in compliance with the IEEE 1159 standard and represented as 224 × 224 × 1 grayscale images. The proposed CNN architecture employs three convolutional blocks with progressively decreasing filter counts (16, 8, and 4), followed by a fully connected layer, resulting in only approximately 38,741 trainable parameters. Despite its compact design, the model achieves a test accuracy of 91.76% on a balanced dataset of 8,500 samples with a macro-average F1 score of 0.9177. Benchmarking against four classical machine learning algorithms; Bagged Trees (84.31%), Decision Tree (73.18%), k-NN (71.53%), and Naive Bayes (6.75%) demonstrate the clear superiority of the proposed approach. With a total computation time of 4 minutes 17 seconds, the model also outperforms Bagged Trees and Decision Tree in efficiency. These results demonstrate that a lightweight CNN architecture, despite its minimal parameter count, can serve as a highly effective and computationally efficient tool for automated power quality disturbance classification in modern power systems.