Anomaly detection and topology identification of distribution network based on conditional variational autoencoder

Area of Science:

• Electrical Engineering
• Power Systems Analysis
• Data Science

Background:

• Modern power systems face increasing complexity due to the integration of intermittent distributed energy sources and new loads.
• Manual updates of distribution network topology are prone to errors, risking grid stability.
• Accurate and up-to-date topology information is crucial for reliable grid operation.

Purpose of the Study:

• To develop an automated method for detecting anomalous data in distribution networks.
• To propose an accurate topology identification model for complex power grids.
• To enhance the stability and operational efficiency of power distribution systems.

Main Methods:

• Utilized a label-conditioned conditional variational autoencoder (CVAE) for anomaly detection in distribution network data.
• Developed a modified CVAE with an improved task formulation for topology identification.
• Implemented knowledge transfer by initializing the topology identification model with parameters from the anomaly detection stage.

Main Results:

• The proposed CVAE-based anomaly detection model achieved high performance with an AUC of 0.9873 and F1-score of 0.9857.
• The topology identification model demonstrated superior accuracy, achieving an AUC of 0.9865 and an F1-score of 0.9623 on the test dataset.
• Validation on a practical dataset confirmed the effectiveness of the strategy in detecting anomalies and identifying network topology.

Conclusions:

• The developed strategy effectively detects anomalous data and accurately identifies distribution network topology.
• Knowledge transfer between anomaly detection and topology identification enhances model performance.
• The approach offers a robust solution for managing complex power distribution networks.