The recognition of three-dimensional semantic information in power scenes is the foundation and key to its subsequent fine-grained management. However, due to the complexity of ground structure information and diverse textures in power scenes, it brings certain difficulties and challenges to its fine-grained understanding and recognition. In this paper, a semantic segmentation method based on improved RandLA Net for power scene point clouds is proposed, which improves the performance of the model by introducing feature expansion and separation pooling operations. And the actual performance of this method is tested on a power dataset and compared with existing semantic segmentation methods. The results show that this method has strong advantages in accuracy and efficiency. and in a comprehensive comparison, it improves the overall accuracy and the average intersection and merger ratio values by 2.64 and 2.9 over the cutting-edge RandLA-Net (Random Sampling and Local Feature Aggregator Network), which verifies the effectiveness of the method.