What question did this study set out to answer?

The research aims to develop a robust framework for detecting energy theft in smart distribution systems using machine learning techniques.

February 16, 2026Open Access

Robust Energy Theft Detection in Smart Distribution Method Using a Data-driven Method

Key Points

The research aims to develop a robust framework for detecting energy theft in smart distribution systems using machine learning techniques.
Utilized a dataset of recorded energy consumption values for analysis.
Employed numerical features for predicting theft conditions: energy used by theft and normal energy consumption.
Implemented four classifiers: Decision Tree, SVM with ECOC, Random Forest, and kNN.
Trained and evaluated models using an 80/20 hold-out validation approach in MATLAB.
Performance assessed through accuracy metrics and confusion matrix analysis.
SVM achieved the highest accuracy at 86.67%, followed by Random Forest at 83.33% and kNN at 82.33%.
Decision Tree had the lowest accuracy at 73.33%.
All classifiers effectively detected theft cases, with errors mostly due to ambiguities between normal and elevated energy usage.

Abstract

Energy theft poses a significant challenge to modern power systems, leading to economic losses, reduced efficiency, and compromised reliability in smart grids. Detecting such anomalies requires robust, scalable analytical frameworks that can accurately distinguish normal consumption, marginally increased usage, and patterns indicative of electricity theft across diverse operating conditions. This study investigates the application of machine learning techniques for energy theft detection using a dataset of recorded consumption values. Two numerical features, energy used by theft in per unit and normal energy, were employed as predictors. At the same time, the target variable comprises three categorical conditions: Theft detected, Normal, and Energy slightly higher. Four classifiers were implemented and compared: Decision Tree, Support Vector Machine (SVM) with Error-Correcting Output Codes (ECOC), Random Forest, and k-Nearest Neighbors (kNN). The models were trained and evaluated using MATLAB with an 80/20 hold-out validation approach. Performance was assessed using accuracy metrics and confusion matrices. Results demonstrated that SVM achieved the highest accuracy (86.67%), followed closely by Random Forest (83.33%) and kNN (82.33%), while Decision Tree yielded the lowest accuracy (73.33%). Confusion matrix analysis showed that all classifiers detected theft-based cases with high accuracy, whereas most classification errors arose from overlap and ambiguity between normal consumption and elevated energy usage conditions. The study adds to the expanding literature on data-driven energy management by providing practical evidence of how machine-learning techniques can strengthen grid security, minimize financial losses, and enhance overall operational efficiency.

Ask AI

Helpful

Bookmark

View Full Paper