Abstract: To enhance the quality of Light Detection and Ranging (LiDAR) point clouds in adverse weather and provide precise detection information for the navigation of photovoltaic (PV) panel cleaning robots, this paper proposes a framework for PV panel contour detection based on a multi-stage data enhancement pipeline. The framework first employs a hybrid algorithm that fuses dynamic statistical filtering and dynamic radius filtering to adaptively handle non-uniform noise. Subsequently, it leverages temporal information for multi-frame matching to complete occluded data. Finally, it achieves precise segmentation and contour detection of PV panels through a cascade of Random Sample Consensus (RANSAC) plane segmentation, Density-Based Spatial Clustering of Applications with Noise (DBSCAN) clustering refinement, and Oriented Bounding Box (OBB) fitting. Validated on 197, 952 simulated point clouds (≈50, 000 points/frame) generated from Velodyne VLP-16 LiDAR data—covering three adverse weather levels and typical PV plant scenarios—the framework demonstrates strong adaptability to man-made regular environments. Under mild (20%-30% data loss) and severe (40%-60% data loss) occlusion, the temporal completion module achieves 93.21% and 95.37% completeness, respectively. The final contour detection algorithm estimates PV module dimensions with width error ≤4.2 cm and thickness error as low as 1.5 mm (millimeter-level) against ground truth. The framework provides a robust, ready-to-deploy solution for overcoming perceptual fading in solar farm robotic navigation under extreme weather, as validated by experimental results.