Semi-supervised fault detection and classification in photovoltaic arrays using a modified generative adversarial networks architecture

Photovoltaic (PV) arrays installed in outdoor environments are prone to various faults and defects that can significantly compromise their reliability. Traditional fault diagnosis methods heavily depend on extensive labeled datasets, which are often expensive and time-intensive to generate. This study introduces a novel semi-supervised learning framework for fault detection and classification in PV arrays based on a modified Generative Adversarial Network (GAN) architecture. The proposed approach leverages a discriminator with a dual functionality design, incorporating both supervised and unsupervised discriminators to leverage labeled and unlabeled data effectively. This design minimizes reliance on labeled datasets while achieving high fault detection accuracy. With only 2% labeled data, the model achieves 96.857% accuracy, and with 10% labeled data, it reaches 99.143%. Simulation results also reveal that the proposed framework outperforms the semi-supervised autoencoder and traditional self-training method, particularly in Gaussian noise-affected data scenarios, underscoring its robustness and practicality for real-world applications. By significantly reducing the need for labeled data, this semi-supervised approach offers a cost-effective and scalable solution for monitoring and ensuring the optimal performance of PV arrays.