A Fault Detection Method for Multi-microgrid Considering Topological Variations Based on Message Passing Neural Network

With the increase of distributed generators in multi-microgrid, multi-microgrids face a large number of problems in fault detection, for example, traditional in-place detection methods are difficult to adapt to topology changes, high data quality and network communication are required for multi-site detection methods, and samples of data-driven methods are massively needed. To solve these problems, a fault detection method based on message passing neural network (MPNN) is proposed, in which the fault detection is modeled as a fault probability prediction problem for multi-microgrid nodes, and a mapping from the multi-microgrid fault waveform is established to the fault type of each node. In the model, convolutional neural networks are adopted to extract features from each node fault waveform. The resulting local node features are propagated among neighboring nodes via an MPNN to update node features for fault localization and fault type classification. For complex fault scenarios with topology changes, topology features are designed to change with the operating state of the multi-microgrid. An enhanced sample dataset is constructed to resist data loss and reduce the total sample requirements. Finally, a multi-microgrid simulation model is built in CloudPSS. The test results show that the proposed method has high fault location accuracy, high fault type classification accuracy, strong resistance to data loss, low sample requirement, and can adapt to the changing operating state and topology of multi-microgrid. © 2023 Science Press. All rights reserved.