Abstract: A key point completion method based on the fusion of deep learning and Markov random fields is proposed to address the texture loss and key point loss in the visualization and structural monitoring of 3D equipment models in the power grid. Using multi-source data of grid and texture as inputs, a joint graph neural network and convolutional neural network are used to extract local global features. Markov random fields are used to characterize spatial dependencies and missing keypoints are generated through Markov chain Monte Carlo sampling. Experimental verification shows that compared to interpolation and single convolutional neural networks, the proposed method significantly improves peak signal-to-noise ratio and structural similarity index, increases completion efficiency by 38%, and provides better rendering realism. It can effectively support 3D visualization and operation of power grid equipment.