Multi-Layer Model Classifier for Cyberattack Detection in Smart Electric Grid

In the Smart Grid, communication lines and physical open access points are always prone to cyber-attacks, and electric theft is the most common one. To detect electricity theft, researchers have developed several advanced machine learning models. However, existing work has not explored the problem of data imbalance properly, which is one of the significant challenges in electricity consumption data. This paper aims to compare various data balancing techniques and present an integrated theft detection model. This paper presents a multi-layer model for detecting fraudulent consumers in the smart grid. The detection process starts with data preparation steps, which include data interpolation, outlier handling, and data standardization. The next crucial step is handling data imbalance. Various techniques are tested, and AdaSys performs better than others. The model is being trained on a balanced dataset and validated on a real imbalanced dataset for realistic results. For higher performance, a two-layer model is chosen for electricity theft detection. The first layer consists of three heterogeneous machine learning models, and an Artificial Neural Network (ANN) model is used for the second layer. The first layer's probabilistic prediction serves as input to the second layer, which makes the final prediction. Experimental results confirm that multilayer model classifiers perform better than individual classifiers for detecting cyber-attacks on real consumption datasets.