A Low-Resource Approach to GAN Images Generation and Evaluation in Photovoltaic Electroluminescence Dataset

The increasing demand for synthetic training data to enhance machine learning algorithms in photovoltaic operations and maintenance applications has driven the need for efficient image generation techniques that can operate under resource-constrained conditions. In this study a Generative Adversarial Network framework, aimed at generating electroluminescence images of damaged PV cells, albeit with limited hardware resources is proposed. By focusing on minimum computational requirements, the proposed methodology exploits optimized deep learning architectures to generate images with a comparable resolution to real applications in the photovoltaic sector. A fundamental element to ensure the significance of the study is the assessment of the generated images of damaged cells through consolidated and recognized metrics, such as the Fréchet Inception Distance. The analysis highlights how the variability of the dataset is crucial, as well as the resolution of the images, in making generative models effective, recommending a cautious attitude in the interpretation of the evaluation metrics in scenarios with limited variability such as the one considered. This study highlights the influence of dataset properties on the perceived image quality but also allows considering a resource-efficient and practically usable image synthesis in the photovoltaic sector.