Photovoltaic systems are crucial for sustainable energy, yet manufacturing defects and environmental degradations compromise their efficiency. While deep learning approaches show promise, existing methods face limitations in precision, computational efficiency, and generalization across diverse defect categories. We introduce Alpha, an attention-enhanced YOLO framework integrating Feature Cross-Attention, Channel Attention, and Efficient Multi-scale Attention to improve spatial and channel-wise feature representations. Comprehensive experiments conducted on electroluminescence image datasets demonstrate that Alpha outperforms state-of-the-art models in terms of accuracy, robustness, and inference speed.