Real-Time Drone-Based Snow Detection for PV Systems Using Robust Lightweight Deep Learning Models

Abstract

The widespread deployment of solar photovoltaic (PV) technology across diverse geographical regions, including cold and snowy climates, introduces distinct operational challenges. One of the most critical among these challenges is snow accumulation on PV panels, significantly hindering energy production efficiency [1]. Snow acts as a physical barrier, obstructing sunlight and preventing photons from reaching the active surface of the solar cells, thereby curtailing or completely halting electricity production. Therefore, timely snow shedding is crucial to restoring optimal energy generation and minimizing energy losses. Studies have shown that delayed snow removal can lead to energy losses of up to 34% [2]. Conventional approaches, such as manual visual inspections and fixed-camera systems [ 3 - 5 ], are often impractical, particularly for large-scale solar farms spanning vast areas. Manual inspections are inefficient and risky, often delaying snow-related energy loss detection, emphasizing the need for automated, data-driven monitoring systems. In our previous work [6] , we introduced a lightweight snow detection algorithm based on YOLOv11n, which analyzed drone imagery and achieved a precision of 0.93 and a recall of 0.75. This algorithm enabled the estimation of Snow Coverage Percentage (SCP), a critical metric for optimizing snow shedding strategies, reducing operational downtime, and enhancing overall energy yield. However, a key limitation of this real-time method lies in its reliance on a static pixel intensity threshold, making it vulnerable to inaccuracies under low-light or uneven lighting conditions. As a result, SCP predictions may be unreliable during dawn, dusk, or overcast weather, leading to potential misclassifications.