Triplet Network for Topology Identification of Distribution Network

Abstract

The distribution network has the problems of inaccurate topology and incomplete measurement configuration. In this paper, a method of distribution network topology identification based on the triplet network is proposed. In order to improve the generalization ability of the model, the Latin Hypercube Sampling (LHS) method considering the source and load correlation was used to generate PV and load data. A hybrid feature selection algorithm combining MLP and PSO is proposed to reduce the number of input measurements. Sequence-to-image conversion using Gramian Angular Field (GAF) is implemented to improve model training efficiency. We introduce a momentum encoder to select hard triplet samples, which solves the problem of easy gradient dissipation when triplet samples are selected randomly. The IEEE33 node system is used to verify the accuracy and superiority of the proposed algorithm, especially in small sample and weak loop network scenarios, the identification accuracy can reach 92% and 89%.