Abstract: High-fidelity digital modeling and visualization of power grid facilities is an important step to construct the digital twin power grid and promote the construction of the new power system. To solve the disadvantages of the manual modeling and meet the requirements of 3D visualization of the digital twin power grid facilities, a method of digital modeling and visualization of the power grid facilities based on the multi-view foreground segmentation is proposed. Based on the combination of deep learning network with the foreground segmentation algorithm, a model of detection, location and foreground segmentation of the power grid facilities is constructed. The original multi-view of the camera is converted into the foreground image of the facility to be built, which reduces the influence of the mismatching of features, the increase of the external noise of the model and other problems caused by the complex image backgrounds, and improves the visual effect of reconstruction. Taking the structure from motion as the fundamental principle, a high-quality camera pose estimation method based on the improved SURF-RANSAC algorithm and a sparse point cloud densification scheme are proposed. Then a unified structured static model reconstruction method for the distributed grid facilities is finally formed. The data set for the facility reconstruction is constructed by the actual images. The power transformer, the transformer winding and the high voltage vacuum AC circuit breaker are simulated and reconstructed respectively, and good reconstruction visual effects and low reprojection errors are obtained, which verifies the universality and effectiveness of the method. The method proposed in this paper not only provides the 3D model support for the unified representation and condition monitoring of the distributed facilities in the virtual space, but also offers the possibility to realize the interaction across space-time and supply-demand balance adjustment of the power grid facilities on the supply-demand side.