Deep learning-based electricity theft prediction in non-smart grid environments.

Area of Science:

• Electrical Engineering
• Data Science
• Artificial Intelligence

Background:

• Developing countries often lack smart grids, leading to significant power supply issues due to electricity theft.
• Existing electricity theft detection models struggle with accuracy for the theft class, despite high performance for non-theft instances.

Purpose of the Study:

• To propose a lightweight deep learning model for effective electricity theft detection in non-smart grid environments.
• To enhance the detection accuracy for the electricity theft class, which is often underrepresented and poorly detected.

Main Methods:

• Utilized monthly customer electricity readings as input for a lightweight deep learning model.
• Employed advanced feature engineering techniques: Principal Component Analysis (PCA), t-distributed Stochastic Neighbor Embedding (t-SNE), and Uniform Manifold Approximation and Projection (UMAP).
• Applied resampling techniques including Random-Under-Sampler (RUS), Synthetic Minority Over-sampling Technique (SMOTE), and Random-Over-Sampler (ROS), with a focus on ROS for imbalanced data.

Main Results:

• Achieved significant improvements in detecting the electricity theft class (class 1).
• Reported high performance metrics for the theft class: 89% precision, 94% recall, and 91% F1 score after parameter tuning and using Random-Over-Sampler (ROS).
• The proposed model demonstrated superior performance compared to existing methods in electricity theft detection.

Conclusions:

• The developed lightweight deep learning model is effective for detecting electricity theft in regions lacking smart grid infrastructure.
• Feature engineering and advanced resampling techniques, particularly ROS, are crucial for improving the detection of the minority theft class.
• This research offers a viable solution to mitigate power supply losses caused by electricity theft in developing countries.