Autonomous UAV Inspection Powered by Generative AI for Photovoltaic Power Loss Analysis

Abstract

Photovoltaic power generation in large-scale solar farms is strongly affected by localized factors such as soiling, partial shading, and component degradation, which are often difficult to quantify accurately using conventional monitoring systems. While UAV-based inspection techniques have shown promise for visual anomaly detection, they typically fail to translate observed defects into measurable impacts on energy production. In this work, we propose a UAV-assisted photovoltaic power loss estimation framework based on large language models, designed to bridge the gap between visual observations and quantitative energy assessment. By integrating UAV imagery, environmental context, and system-level metadata within a unified multimodal reasoning pipeline, the proposed approach estimates panel- and farm-level power losses directly in terms of energy output. Extensive experiments conducted on photovoltaic datasets combining UAV-acquired images and SCADA-based ground truth demonstrate the effectiveness of the proposed framework compared to vision-only and deep learning baselines. The results show a consistent reduction in power loss estimation error, achieving a relative improvement of 18.7% in MAE and 15.2% in RMSE over the strongest baseline, along with a higher correlation with actual power production across varying irradiance conditions. Additional analyses confirm the robustness of the proposed method under partial observations and heterogeneous environmental settings. These findings highlight the potential of LLM-driven aerial monitoring as a practical and energy-aware solution for photovoltaic performance assessment and operational decision-making.