PvolaSolar & Storage← All stories

Technology

China Huaneng PVFire-YOLO Detects Solar Plant Fires at 87.4% Precision

3 min read
China Huaneng PVFire-YOLO Detects Solar Plant Fires at 87.4% Precision

The short answer

China Huaneng Group built PVFire-YOLO, a real-time fire detection system for solar plants that achieved 87.4% mean average precision on a dedicated dataset. The system reduces false alarms caused by sunlight reflection to 3.1%, enabling reliable edge deployment for utility-scale monitoring.

Key facts

Technology
PVFire-YOLO, a deep-learning framework based on YOLOv12 with attention blocks for flame and smoke detection.
Status
Research study published in Scientific Reports; not yet commercialized or deployed in mass production.
Partner
China Huaneng Group, a Chinese state-owned utility, developed the system.

A research team at China Huaneng Group , a Chinese state-owned utility, has built a deep-learning framework aimed at spotting fires and smoke in solar photovoltaic plants in real time, according to pv magazine . The system, called PVFire-YOLO , is intended to catch small flames and smoke plumes while cutting down on false alarms triggered by sunlight bouncing off module surfaces.

The researchers noted that deep-learning work in PV inspection has largely focused on three areas: electroluminescence crack detection, infrared hot-spot localization, and panel-level visible-light anomaly detection covering surface defects and soiling. Fire detection in PV settings, they said, has drawn almost no dedicated attention.

According to the team, it compiled the first annotated flame-and-smoke dataset built specifically for PV environments. The corpus, named PV-Fire, draws on surveillance and drone imagery from operating utility-scale plants along with controlled outdoor burns staged among real PV modules.

The method builds on the YOLOv12 real-time object detection model and adds three components. A patch-aware attention block strengthens weak flame and smoke features, a gather-and-distribute mechanism merges information across four feature levels, and an adaptive fusion head introduces a high-resolution detection branch for small targets.

The PV-Fire dataset holds 5,824 images with 9,107 annotated flame and smoke objects, combining surveillance and drone imagery, controlled outdoor burns, and synthetic images. Of these, 4,070 images were set aside for training, 870 for validation, and 884 for testing. Scenes were kept separate across the three subsets, and synthetic images were confined to the training set.

The team tested PVFire-YOLO against 14 existing detection models on the PV-Fire dataset, including Faster R-CNN, transformer-based models, and several YOLO variants. It also compared the system with 13 alternative models on the public D-Fire dataset. The evaluations covered detection accuracy, small-target performance, reflection-driven false alarms, cross-dataset generalization, and processing speed on desktop and embedded hardware, with results averaged over three training runs.

PVFire-YOLO reached a mean average precision of 87.4% at an intersection-over-union threshold of 0.5 on the PV-Fire dataset. That figure came to 85.0% when testing was limited to footage from operating PV plants and 84.4% on the D-Fire dataset.

Against the baseline YOLOv12 model, the proposed method raised small-target average precision by 5.2 percentage points. It also brought the reflection-related false-positive rate down to 3.1% while identifying a larger share of actual flame and smoke targets.

The researchers reported that the model runs at 79 frames per second on an RTX 3090 desktop graphics card and 21 frames per second on a Jetson Orin Nano compact AI computer, which they said fits the latency budget for real-time edge deployment in distributed plant monitoring. Planned future work includes extending the corpus to plants in varied climates, adding temporal filtering to suppress isolated false positives, and combining visible-light detection with infrared thermal imaging to improve robustness under poor illumination.

The findings appear in the study titled Real time small target fire detection in photovoltaic plants using attention enhanced multiscale fusion, published in Scientific Reports.

Interactive table based on the Store Companies dataset for this report.

Pvola take

This work addresses a critical gap in PV safety, as fire detection has received little dedicated attention compared to other inspection tasks. By achieving high precision on a purpose-built dataset and running at 21 fps on edge hardware, the system offers a viable path for real-time monitoring that balances accuracy with the latency constraints of distributed plant operations.

Frequently asked questions

What is the mean average precision of PVFire-YOLO on the PV-Fire dataset?

PVFire-YOLO achieved a mean average precision of 87.4% at an intersection-over-union threshold of 0.5. This performance was measured on the first annotated flame-and-smoke dataset built specifically for PV environments.

How does PVFire-YOLO improve small-target detection compared to the baseline YOLOv12 model?

The proposed method raised small-target average precision by 5.2 percentage points against the baseline YOLOv12 model. It incorporates a patch-aware attention block to strengthen weak flame and smoke features.

Why is dedicated fire detection technology important for utility-scale solar plant operators?

Fire detection in PV settings has drawn almost no dedicated attention in deep-learning research. The system aims to catch small flames and smoke plumes while cutting down on false alarms triggered by sunlight bouncing off module surfaces.