As the monitoring field of substation is wide and the high-voltage equipment in the station is uneven and scattered, the 3D-BoNet example segmentation model has a large deviation for the point cloud segmentation of high-voltage electrical equipment in the station. In this paper, a 3D target instance segmentation model 3DPowerSegNet for substation scene based on improved 3D-BONET is proposed. Firstly, the Groupsift feature extraction module is proposed to carry out convolution from multiple directions to capture more key features and insert them into the backbone network to enhance the ability to extract local features. Secondly, in the sampling stage under point cloud, the set abstraction module is improved to achieve normalization by expanding the query radius to increase the sensitivity field. Then, the inverted residual feature module (IRF) is proposed to obtain richer feature details by expanding the channel, and alleviate the problems of gradient disappearance and model overfitting. Finally, the feature propagation module is improved in the up-sampling stage to reduce information loss during data processing. The comprehensive experimental results on the self-built substation scene data set and the public data set show that the 3DPowerSegNet model can accurately extract the point cloud features of the target object in sparse point cloud environment, and the point cloud segmentation accuracy reaches 63.87%, 71.80% and 52.20%, compared with the original 3D-BoNet model. mAP increases by 2.80%, 4.51% and 6.97%, respectively.