Electricity Theft Detection in Smart Grids Using a Hybrid BiGRU-BiLSTM Model with Feature Engineering-Based...

Area of Science:

• Machine Learning
• Data Science
• Cybersecurity

Background:

• Defused decision boundaries, caused by cross-pairs in data, lead to misclassification issues in machine learning models.
• Tackle the challenge of defused data samples that possess cumulative attributes of multiple classes, misleading classifiers.
• The presence of class imbalance and the need for robust feature engineering in classification tasks.

Purpose of the Study:

• To investigate and resolve misclassification problems arising from defused data and cross-pairs.
• To develop and validate a robust hybrid classification model for theft case scenarios.
• To enhance classifier detection capabilities and model integrity against adversarial attacks.

Main Methods:

• Tomek Links technique was employed to remove majority class cross-pairs, creating an affine-segregated decision boundary.
• Theft data was synthesized using six variants, and K-means minority oversampling addressed class imbalance.
• A hybrid model combining Bi-Directional Gated Recurrent Units (Bi-GRU) and Bi-Directional Long-Term Short-Term Memory (Bi-LSTM) was utilized for classification, with stochastic feature engineering.

Main Results:

• The proposed method successfully created an affine-segregated decision boundary, mitigating misclassification from defused data.
• The hybrid Bi-GRU and Bi-LSTM model demonstrated efficient classification of synthesized theft data, even with class imbalance.
• The model exhibited robustness against unseen attack vectors, maintaining its efficiency and integrity.

Conclusions:

• The integration of Tomek Links, data synthesis, and a hybrid Bi-GRU/Bi-LSTM model effectively handles defused data and class imbalance in theft detection.
• The proposed approach provides an efficient and robust solution for classifying complex datasets and resisting adversarial interventions.
• The study highlights the potential of advanced deep learning architectures and data preprocessing techniques for improving classification accuracy and security.