Point cloud data enhancement and automatic topology structure extraction technology for complex distribution network scenes based on deep learning

Point cloud data in complex distribution network scenarios find it difficult to accurately extract topological structures due to noise interference, severe occlusion, and blurred features. To address this, this paper proposes a method that integrates an improved PointNet++ with a Generative Adversarial Network (GAN). This method uses the GAN to perturb and reconstruct point cloud space to expand sample distribution and improve data diversity. It also leverages PointNet++’s layered sampling and local feature aggregation mechanisms, combined with a residual structure, to enhance the capture of spatial detail features. A graph structure is then constructed based on the enhanced point cloud, and a graph convolutional network is applied for iterative updates, enabling automated reasoning and relational modeling of power grid topology. Experimental results show that the average point density in the point cloud enhancement interval of 0.2–0.3 increases from 890 to 1280 points. The improved PointNet++ achieves an accuracy of 96.4% on the transmission line classification task, and the topology recognition rate of the fusion method stabilizes at 96% after 200 iterations. This method achieves an organic fusion of data enhancement and automatic topology extraction, providing effective support for the digitalization and intelligentization of distribution networks.