AI model detects small fires and smoke in PV plants
Researchers at China Huaneng Group have developed a deep-learning framework called PVFire-YOLO to detect early-stage fires and smoke in photovoltaic plants. Tested on a custom-built dataset, the system achieved an 87.4% detection accuracy while effectively reducing false alarms triggered by sunlight reflections from solar panels.

Original image temporarily unavailable
Background
While deep learning has been widely applied in photovoltaic (PV) inspections for crack detection and hot-spot localization, dedicated fire detection in PV environments has rarely been explored.
Details and Performance
Developed by China Huaneng Group, the system integrates a custom annotated corpus named PV-Fire, which contains 5,824 images gathered from operating utility-scale plants and controlled outdoor burns. Based on the YOLOv12 model, the framework incorporates a patch-aware attention block, a multi-level gather-and-distribute mechanism, and an adaptive fusion head to spot weak flames and tiny smoke plumes. In benchmarks, PVFire-YOLO reached an mAP@50 of 87.4% on its native dataset, lowered the reflection-related false-positive rate down to 3.1%, and ran at 79 frames per second on a desktop GPU.
Outlook
The researchers state that future work will expand the image corpus to cover diverse climates, introduce temporal filtering to minimize isolated false positives, and combine visible light detection with infrared thermal imaging.
View selection rationale
Importance 67/100
About this brief
- Content type
- Research summary
- Source type
- Publisher report
- Original source
- pv magazine Global ↗
- Source published
- Process
- AI-assisted discovery, translation, summarisation, classification and scoring; standards and corrections are maintained by the editorial desk.
- Verification
- The original URL and publication metadata are retained. This brief is not independently fact-checked unless explicitly stated.
- Published by
- Polyemergence Editorial Desk · PolyEmergence Pte Ltd.
- Conflict disclosure
- No financial or commercial conflict declared.