Hybrid CNN-Autoencoder model for accurate and efficient fault diagnosis in grid-connected photovoltaic systems

The reliability of grid-connected photovoltaic (PV) systems is critical for optimizing energy output and minimizing maintenance costs, yet traditional fault diagnosis methods often lack accuracy, scalability, and real-time capabilities. This paper proposes a hybrid CNN-Autoencoder model to address these limitations by combining the feature extraction prowess of Convolutional Neural Networks (CNNs) with the dimensionality reduction efficiency of Autoencoders. The CNN captures high-dimensional spatiotemporal patterns from time-series photovoltaic data, while the Autoencoder compresses these features into a discriminative latent representation, enhancing computational efficiency and generalization. Evaluated on the GPVS-Faults dataset, comprising various fault scenarios such as inverter failures, grid anomalies, and sensor faults, the model achieves a state-of-the-art test accuracy of 98.84%, outperforming existing methods like MobileNet+ESC (97.36%) and traditional 1D-CNNs (90.24%). In particular, it demonstrates 100% precision for critical faults (e.g. MPPT controller faults) and robust performance under noisy conditions. Comparative analysis with 15 benchmark models confirms its superiority in balancing precision, recall, and computational complexity. The proposed framework offers a scalable, AI-driven solution for real-time PV fault monitoring, paving the way for improved grid stability and reduced operational disruptions in large-scale solar installations. This research underscores the potential of hybrid deep learning architectures to advance renewable energy system diagnostics. • A hybrid CNN-Autoencoder model is proposed for PV fault classification. • The method combines feature extraction and reconstruction for better representation. • The hybrid model improves accuracy from 91.00% to 98.94%. • False alarm rate is reduced from 10.5% to 1.8% with the hybrid approach.