A method for intelligent recognition and multi-parameter inversion of photovoltaic panels based on UAV-LiDAR point clouds and DGRNet

Abstract

Addressing the shortcomings of existing monitoring methods for mountainous photovoltaic power plants—including insufficient 3D data acquisition, low levels of automation, and point cloud recognition methods that are poorly adapted to mountainous scenarios, lack systematic comparisons of mainstream models, and rely on prior models for parameter inversion—this paper proposes a method for intelligent recognition of photovoltaic panels and multi-parameter inversion based on UAV LiDAR point clouds and DGRNet. Leveraging DGRNet’s strengths in multiscale feature extraction and global modeling under non-uniform point cloud conditions to adapt to complex mountainous scenarios, this study uses a DJI Matrice 4E equipped with a Zenmuse L2 to collect point clouds. After preprocessing, the vertical and horizontal accuracies reach RMSE ≤ 5 cm and RMSE ≤ 8 cm, respectively. Compared with four mainstream deep learning models, DGRNet demonstrates the best performance, with a core PV panel category union-set ratio of 0.624 and an F1 score of 0.768; a workflow for building a model without prior knowledge was established, enabling multi-parameter extraction of PV panels, capacity verification, and soiling detection. The results demonstrate that this method achieves excellent segmentation and inversion performance in complex mountainous terrain, showing high consistency with design parameters and field data, and can provide technical support for intelligent inspection and O&M optimization of mountainous PV power plants.