This article focuses on the analysis of large-scale distribution network reconstruction fused with graph theory and graph partitioning algorithms. Graph theory and graph segmentation algorithms have been rushed by many researchers in the fields of medicine, drone, and neural network. It is a newcomer in the field of computer vision, which can not only realize the division in color but also divide it by image data. The distribution network is also indispensable for new energy, electric machines, but the traditional distribution network has many problems, such as not suitable for distributed power access and excessive network loss. To improve the performance of distribution networks and reduce network losses, this paper A multi-division model for distribution network construction and reconstruction is established, and a graph theory-based division algorithm method is proposed to effectively solve the problem of feeder-to-feeder reconstruction during large-scale distribution in distribution networks. Through its superconductivity phenomenon and the characteristics of clustering algorithm division, this paper uses formulas to show its division principle and gives examples of various distribution network reconstruction algorithms to explore which method of improvement can improve the performance of the distribution network and reduce network losses. The number of iterations is also strictly considered, and the value is taken after multiple iterations to reduce the error. Through the distribution network calculation example, the network loss reduction value is obtained, and the distribution network fault repair model is exemplified. The picture is used to briefly describe the process of distribution network reconstruction and find that the faults of the distribution network can be quickly located and isolated through the FTU, and quickly repaired. Finally, in order to reduce the network loss, reduce the load of power flow calculation, and solve the problem of local optimization, a JA-BE-JA optimization algorithm based on large-scale distribution network reconfiguration is proposed. The mixed sampling method is preferred to test the number of divisions in the four states, and the parameters are selected to test the performance of the improved annealing simulation algorithm, and the conclusion is drawn as follows: the improved graph segmentation algorithm has strong robustness, can avoid local optimization of graph data, and can reduce network loss. Compared with traditional distribution network reconstruction methods, the network loss can be reduced to 454.3 KW, which can be optimized by 10.68% compared with the initial network loss.