1.College of Electrical Engineering, Sichuan University, Chengdu 610065, China;2.State Grid Fujian Ultra-high Voltage Company, Xiamen 361004, China

The operation mode of active distribution network is frequently switched and its measurement is affected by noise, so its topology identification is required to be real-time and robust. However, the topology identification under limited observation driven by traditional methods is a non-deterministic polynomial (NP)-hard problem. This paper proposes a robust identification method for distribution network topology based on generative adversarial network and partial current phasors. In order to solve the problem of poor generalization ability of traditional supervised learning in unknown topology identification tasks, the conditional generative adversarial network with gradient penalty optimization is used to learn the topology distribution mapped by line current amplitude, phase angle and node load pseudo-measurement. Meanwhile, by constructing generator with one-dimensional convolutional neural network, the sequential current data in continuous observation window are effectively used, which enhances the anti-noise and anti-data loss performance of the algorithm. In addition, the need for only partial current information greatly reduces the investment in observability retrofit of distribution network. Finally, the effectiveness of the proposed method is verified by arithmetic examples.