Anomaly detection in solar farms with a hybrid CNN-ViT model to enhance accuracy and efficiency

Solar cells are widely used renewable energy technologies for electricity generation in modern power systems. In networks with high adoption of solar farms, damage to solar panels can threaten the reliability of consistent power generation, underscoring the need for swift maintenance, which in turn demands significant financial investment. Therefore, effective and rapid anomaly detection is crucial for reducing maintenance costs. One challenge is the vast harvesting farms, which complicates monitoring and maintenance. To address this, employing modern technologies such as drones for live-streaming and DL-based methods for anomaly detection offers a promising approach. Real-time processing methods This paper proposes a hybrid Convolutional Neural Network and Vision Transformer (CNN-ViT) model that balances local feature extraction and global pattern recognition for detecting defects in solar arrays. Trained on 20,000 infrared images, it achieved 85.61% accuracy, surpassing standalone CNN (74.87%) and ViT (69.31%). The hybrid model demonstrates its suitability for real-time monitoring of large-scale solar farms with an average inference time of 0.763 ms per image, which can potentially reduce operational costs and improve maintenance efficiency.