Graph Neural Networks for Fault Diagnosis in Photovoltaic-Integrated Distribution Networks with Weak Features.

Area of Science:

• Electrical Engineering
• Power Systems Analysis
• Artificial Intelligence in Energy

Background:

• Effective power system fault diagnosis is vital for reliability.
• New energy integration causes bidirectional power flow, challenging traditional methods.
• Data-driven methods require substantial, high-quality training data and struggle with noise and variable conditions.

Purpose of the Study:

• To develop an advanced fault diagnosis model for power distribution networks.
• To overcome limitations of traditional and existing data-driven fault detection methods.
• To improve fault detection performance with reduced training samples and increased robustness.

Main Methods:

• Introduction of a dynamic aggregation module for noise reduction and information integration.
• Proposal of a coupled dual-field-encoding module to encode topological and physical-electrical domain information.
• Utilizing graph neural networks (GNNs) for feature extraction and propagation learning.

Main Results:

• The proposed DACDFE-GNN model demonstrates superior fault detection performance.
• The model shows significant effectiveness even with a low rate of training samples.
• Experimental validation on IEEE 34- and IEEE 123-node feeder systems confirms performance gains.

Conclusions:

• The DACDFE-GNN model offers a robust solution for power distribution network fault diagnosis.
• The model's ability to handle noise and limited data makes it practical for real-world applications.
• This approach enhances power system reliability by improving fault detection accuracy and efficiency.