A Novel Feature-Engineered-NGBoost Machine-Learning Framework for Fraud Detection in Electric Power Consumption Data.

Area of Science:

• Electrical Engineering
• Data Science
• Machine Learning

Background:

• Power consumption fraud poses significant economic challenges.
• Accurate detection of electricity theft is crucial for utility companies.
• Existing methods often struggle with complex patterns in smart meter data.

Purpose of the Study:

• To develop and validate a novel machine-learning framework for detecting power consumption fraud.
• To enhance the accuracy and efficiency of electricity theft detection using advanced algorithms.
• To identify key features influencing fraud detection in smart meter data.

Main Methods:

• A three-stage framework involving data pre-processing, feature engineering, and model evaluation.
• Utilized Random Forest for missing data imputation and Majority Weighted Minority Oversampling Technique (MWMOTE) for class balancing.
• Employed time-series feature extraction, Whale Optimization Algorithm for feature selection, and NGBoost for classification.
• Tree SHAP was used for feature impact analysis.

Main Results:

• The NGBoost framework achieved 93% accuracy, 91% recall, and 95% precision.
• Demonstrated superior performance compared to existing competing models.
• Successfully classified consumers into 'Healthy' and 'Theft' categories.

Conclusions:

• The proposed feature-engineered NGBoost framework is highly effective for power consumption fraud detection.
• The methodology offers a significant advancement in identifying electricity theft.
• The model's high performance validates its efficacy and importance in the field.