SolarSynthNet (SSN): A deep learning framework for binary and multiclass classification of damaged or obstructed solar panels using images

The rapid rise in solar photovoltaic (PV) installations globally drives demand for tools to optimize system operations and maintenance. Machine learning models can identify and classify faults (such as cracks or physical damage) or obstructions (such as dust or snow) that reduce performance. However, existing models lack the scalability, accuracy, and generalizability required to handle large, noisy, and diverse datasets of real-world solar panel images. This paper presents SolarSynthNet (SSN), a novel deep learning framework for classifying faulty or obstructed solar panels. SSN is built upon a base feature extraction block supported by an Enhanced Feature Mixing Block and Contextual Focus module. These improve feature extraction by combining information from multiple layers to capture complex patterns, and increase classification accuracy by prioritizing the most relevant regions to focus on discriminative features. Experiments across multiple datasets demonstrate SSN's accuracy, correctly identifying 91.48 % of solar panels in 6-class, 95.83 % in 3-class, and 97.42 % in binary classification. SSN outperforms all existing models in both binary and 6-class tasks. It offers strong potential for optimizing operational performance and maintenance strategies in real-world solar energy systems, thus improving project economics. Future work will improve sample balancing, image quality, and expand SSN to real-time applications.