I built an AI system that detects solar panel faults from a single image — and deployed it with a live demo.
Here's what it does:
𝟲 𝗳𝗮𝘂𝗹𝘁 𝗰𝗹𝗮𝘀𝘀𝗲𝘀: Bird-drop, Dusty, Snow-Covered, Electrical Damage, Physical Damage, Clean
• ResNet101 fine-tuned on solar panel imagery
• 96.3% training accuracy with strong per-class F1
• REST API deployed on Hugging Face Spaces (Docker + Flask)
• React frontend with real-time confidence scores
𝗪𝗵𝘆 𝘁𝗵𝗶𝘀 𝗺𝗮𝘁𝘁𝗲𝗿𝘀:
According to Raptor Maps' 2025 Global Solar Report, equipment-driven underperformance has increased 214% since 2019 — resulting in $10 billion in lost energy value in 2024 alone. The average solar site loses 5.77% of its power output annually, costing ~$5,720 per MWdc.
The use cases that excite me most:
• Drone inspection over utility-scale farms
• Ground robotics for close-up panel analysis
• Direct integration into SCADA systems via REST API
𝗕𝘂𝗶𝗹𝘁 𝘄𝗶𝘁𝗵: PyTorch · ResNet101 · Flask · Docker · HuggingFace Spaces · React + Vite