Utilizing Unlabeled Data to Detect Electricity Fraud in AMI: A Semisupervised Deep Learning Approach

Area of Science:

• Electrical Engineering
• Data Science
• Cybersecurity

Background:

• Nontechnical losses in power systems are a growing global concern.
• Smart meters enable electricity fraud detection via consumption patterns.
• Existing fraud detection models struggle with high-dimensional data and insufficient labeled data.

Purpose of the Study:

• To develop a deep-learning model for effective electricity fraud detection in advanced metering infrastructure.
• To address limitations of existing models in handling high-dimensional data and data scarcity.

Main Methods:

• Developed a deep-learning model named the multitask feature extracting fraud detector (MFEFD).
• Employed a semisupervised learning approach with multitask training to leverage both labeled and unlabeled data.
• Utilized deep architecture for effective extraction of consumption patterns from high-dimensional load profiles.

Main Results:

• MFEFD demonstrates high detection performance in real-world case studies.
• The model shows robustness and effective privacy preservation.
• The semisupervised, multitask approach enhances generalization despite data limitations.

Conclusions:

• The MFEFD model offers a practical and effective solution for electricity fraud detection.
• Deep learning and semisupervised learning are crucial for improving fraud detection accuracy and generalization.
• The developed model addresses key challenges in current electricity fraud detection systems.