Detecting Electricity Theft in Smart Grids: A Hybrid CNN-RF Approach

Abstract

Electricity theft poses a significant economic and operational challenge to modern power grids, resulting in substantial financial losses and compromising grid reliability. Traditional detection methods, such as manual inspection and basic statistical analysis, are insufficient for addressing the complexity and scale of theft in smart grids. This study proposes a hybrid machine learning model, CNN-RF, combining convolutional neural networks (CNNs) for feature extraction and random forests (RF) for classification. Using a publicly available dataset spanning 12 months of hourly energy consumption data across 16 consumer categories, the model achieves superior accuracy of 88.84%, precision of 86.55%, recall of 88.84%, and an F1-score of 86.60% compared to standalone CNN, RF, and support vector machine (SVM) models. Specifically, CNN-RF improves accuracy by 1.6% over RF and 0.8% over CNN, demonstrating the advantage of integrating deep feature extraction with robust classification.