اميرحسين كوهستاني

The identification and classification of Power Quality Disturbances (PQDs) is crucial for maintaining network efficiency in modern electrical systems. While deep learning approaches have proven highly successful in this domain, a significant limitation persists: most models are developed using labeled synthetic data, which fail to capture the complexity of real-world power quality signals that are often unlabeled. This discrepancy leads to a phenomenon known as "domain shift" that challenges the performance of deep learning models in real-world power networks. This study introduces a novel approach to PQD classification using Convolutional Domain-Adversarial Neural Networks. The proposed method addresses the domain shift challenge through learning domain-invariant features. The proposed model is trained on labeled synthetic PQD data while utilizing unlabeled real-world signals during the adaptation process. This unsupervised "domain adaptation" method enables learning the complexities present in real power quality data without requiring labeling, which represents a significant advantage over conventional methods. To eva‎luate the model, two datasets of synthetic and quasi-real power quality data were used, and its performance was compared with traditional convolutional neural networks. Results demonstrate that the domain-adversarial model significantly outperforms traditional convolutional models, particularly when classifying real-world power quality disturbances (classification accuracy of 82.6% for the domain-adversarial model versus 52.2% for traditional convolutional models). The performance of the domain-adversarial approach demonstrates the importance of domain adaptation in monitoring power quality data in real-world scenarios. Furthermore, this approach can be extended to other power system monitoring problems where domain shift occurs, defining new and innovative applications for deep learning models in power systems.

اختلالات كيفيت توان , يادگيري عميق , جابجايي دامنه , شبكه‌هاي عصبي دامنه-متخاصم , شبكه‌هاي عصبي كانولوشني

Power Quality Disturbances , Deep Learning , Domain Shift , Domain-Adversarial Neural Networks , Convolutional Neural Networks

Amir Hossein Koohestani

Dr. Alireza Jalilian