Artificial intelligence vision models have been widely applied in electric power scenarios such as transmission line inspection and operational safety management. However, existing models still face challenges such as limited accuracy and weak model generalization ability. This study aims to address these issues by constructing and pre-training a backbone for deep vision models in electric power scenarios. We collected a large number of images from the electric power domain and combined them with general-domain images for unsupervised pre-training, thereby improving the downstream task accuracy of the model. We propose several improvements to enhance the performance and training efficiency for visual backbones in power scenarios, including a multi-scale visual backbone combining convolutional and Transformer blocks, a rotational position embedding method for 2D images, a local self-attention operator based on the Flash Attention algorithm, an unsupervised pre-training algorithm which learns to reconstruct the features from a contrastive language-image (CLIP) model, and a grouped masking method for masked image modeling pre-training with both convolutional and Transformer architectures. We evaluate the performance of the proposed backbone model across multiple downstream scenarios and datasets.