Topology Identification of Distribution Networks Using Multi-Object Binary Classification Graph Transformer

With the continuous increase of renewable energy, the distribution network's structure tends to be more complex. The real-time topology is difficult to obtain timely and the topol-ogy categories increases exponentially, which poses challenges to further analysis and the network's computation. Traditional distribution network topology identification methods rely on high computational complexity and require a large number of fault samples. To address these problems, this paper proposed a Multi-Object(MO) binary classification graph transformer method for topology identification. Through mapping the switch status and voltage measurement and branch conductance of the distribution network, a MO classification mechanism is introduced, performing binary classification for each edge with undetermined connection status, and physically map the output of the identification model to the switches status in the distribution network. The method utilized a Graph Transformer-based convolutional neural network which used to extract the nodes and the edges features. Combing the proposed architecture with an online hard example mining mechanism and maximum spanning tree filter strategy, the model collapse caused by imbalanced data is dealt with and the accurate identification is realized. The proposed method is validated by the IEEE 33-bus distribution network system, and the results show the high accuracy for practical topology identification.