Abstract
Effective diagnosis of distribution network faults is crucial to ensuring the reliability of power systems. However, the bidirectional power flow caused by the integration of new energy limits the effectiveness of traditional detection methods. Although data-driven approaches are not restricted by power flow direction, their performance is heavily dependent on the quantity and quality of training samples. In addition, factors such as measurement noise, variable fault impedance, and volatile photovoltaic output complicate fault information. To address this, we present a new fault diagnosis model named the dynamic, adaptive, and coupled dual-field-encoding graph neural network (DACDFE-GNN), which introduces a dynamic aggregation module to assign different weights to reduce noise interference and fully integrates information from observable nodes. On this basis, the coupled dual-field-encoding module is proposed, which encodes topological information and physical-electrical domain information as part of the initial features, thereby capturing fault features and learning the law of feature propagation. The experimental results for the IEEE 34- and IEEE 123-node feeder systems indicate that the proposed model surpasses recent fault diagnosis methods in detection performance, particularly regarding its low training sample rate.