Deep learning framework for fault detection and diagnosis in grid-connected PV systems using GAN-based data augmentation

The evolution of photovoltaic (PV) systems has been instrumental in the transition towards sustainable and renewable energy. However, in the real world, in PV monitoring, there are issues that are yet to be resolved. There is a dearth of failure-labelled data, and the present models do not generalise well across many different PV systems and edge deployment settings. This paper suggests a robust deep learning model to detect and classify photovoltaic system (GPVS) faults on solar panels by exploiting frontier technologies, such as the utilisation of Generative Adversarial Networks (GAN) for data augmentation and carbon-efficient deep learning for classification. This framework applies high reliability, improved classification performance to seven different fault types under both limited and maximum power conditions. The conditional GAN achieves data enhancement in the form of generating highly representative edge-case synthetic samples, while class imbalance is resolved through SMOTE. The various GANs give synthetically augmented images to increase the augmented images per class, with SMOTE also used to over-sample the minority class in the feature space. The authors chose to use these particular augmentation techniques, which is due to the fact that conditional GANs (cGANs) put out very good representations of rare patterns, which traditional oversampling does not produce, and SMOTE, which does a great job at balanced feature space interpolation; in total, they address issues of under-representation and structural diversity of PV faults. A Hybrid feature selection technique that integrates advanced techniques like Boruta, Lasso, XGBoost, decision trees, and recursive feature elimination (RFE) is used to achieve better efficiency by filtering out the least influential features with respect to the target feature. Four lightweight carbon-efficient deep learning models – Quantised CNN (QCNN), Tiny-LSTM, FBNet, and Carbon-MLP – are utilised for enhancing the fault classification. The proposed framework greatly improves the reliability and sustainability of PV systems. These architectures are chosen for their carbon-efficient design, small parameter set size, and also do well in low-power field-deployed PV monitoring units, which in turn gives fast inference speed without a trade-off of accuracy, which in turn also improves on the issue of very heavy computation of traditional deep models in edge settings. The results of this study underline the potential of using advanced deep learning methods for enhancing the efficiency of solar systems. Optimising fault detection and diagnosis helps to significantly improve the overall resilience of renewable energy systems, thereby facilitating a smoother transition to a more sustainable energy future. Overall, the paper presents a unified framework that puts forward data-centred augmentation, hybrid feature selection, and edge-optimised deep learning to put together a scalable and sustainable PV fault diagnosis pipeline. • Unified PV fault model handles data scarcity, imbalance, and edge limits. • Train-only cGAN+SMOTE improves rare faults, keeps test data unbiased. • Hybrid feature selection via Boruta, LASSO, XGB, DT, and RFE union. • QCNN, TinyLSTM, FBNet, CarbonMLP show strong accuracy-efficiency trade-offs. • Achieves strong PV fault detection, suited for real-time edge systems.