Hybrid CNN-Transformer Network for Electricity Theft Detection in Smart Grids

Area of Science:

• Electrical Engineering
• Data Science
• Artificial Intelligence

Background:

• Electricity theft is a major cause of power loss globally.
• Existing neural network models for electricity theft detection (ETD) struggle with deep feature extraction.
• Accurate anomaly detection in power consumption data remains a challenge.

Purpose of the Study:

• To develop a novel hybrid model combining Convolutional Neural Network (CNN) and Transformer networks for improved electricity theft detection.
• To enhance the capability of extracting profound features from power consumption data for more reliable ETD.
• To address the limitations of existing methods in capturing multi-scale and temporal dependencies.

Main Methods:

• A hybrid model integrating a CNN with a dual-scale dual-branch (DSDB) structure for shallow feature extraction.
• Utilizing a Transformer module with Gaussian weighting (GWT) for deep-level temporal dependency extraction.
• Applying the model to analyze power consumption data for anomaly detection.

Main Results:

• The proposed hybrid model demonstrates superior feature extraction capabilities compared to existing methods.
• The model achieves high F1 scores and Area Under the Curve (AUC) values, indicating effective electricity theft detection.
• The method exhibits significant robustness in identifying anomalies in power consumption patterns.

Conclusions:

• The hybrid CNN-GWT model offers a powerful and robust solution for electricity theft detection.
• This approach significantly advances the field of anomaly detection in energy consumption data.
• The findings suggest a promising direction for reducing power losses due to electricity theft.