Study on Low Power On-Device AI-Based Fire Prevention Monitoring System for Photovoltaic Systems

This study proposes a low-power on-device artificial intelligence (AI)-based monitoring and cleaning system for the early recognition of contamination states associated with fire risks in photovoltaic installations. Photovoltaic panels are continuously exposed to outdoor environments, where dust, soil, leaves, bird droppings, and other contaminants keep accumulating over time. Such contamination may cause partial shading and localized heating (hot spots), which can be associated with abnormal thermal and electrical conditions in photovoltaic systems. To address this issue, an on-device image-based classification system was implemented using a microcontroller-based camera module and a lightweight AI model to classify the surface contamination state of photovoltaic panels. The proposed system classifies panel conditions into two states, namely, normal and contaminated, and performs inference entirely on the device without using external servers. The model was trained using 1,285 images, and the test results showed a precision of 0.94327, recall of 0.94307, F1-score of 0.94299, and classification accuracy of 94.05%. The average inference latency of the ESP32-S3 microcontroller-based device was 8,607 ms. Experimental results demonstrated stable classification performance under various contamination conditions, indicating that the proposed system can be used as a first-stage monitoring tool for maintenance and surveillance of fire-risk-related conditions in photovoltaic installations. Future work will focus on expanding the long-term real-world datasets and integrating thermal and electrical features for a more quantitative risk assessment.