The Effectiveness of Using AutoML in Electricity Theft Detection: The Impact of Data Preprocessing and Balancing Techniques

Electricity theft is a significant threat to the economy and society’s security. Machine learning models are applied to consumption data obtained from smart meters to detect and prevent it. The consumption data must undergo preprocessing and feature selection techniques to enhance detection accuracy. Moreover, the data is imbalanced, where it contains rare malicious samples. This study uses AutoML to detect electricity theft by utilizing its useful services, such as data preprocessing, feature engineering, algorithm selection, and model hyperparameter optimization. We will showcase and evaluate two prominent platforms, TPOT and H2O, and analyze the effectiveness and accuracy of each. The study will examine whether applying additional preprocessing for outlier detection and handling impacts TPOT’s and H2O’s performance. Additionally, we will examine the performance of each platform on the unbalanced datasets related to electricity theft and compare it with their performance after balancing the dataset using three different techniques - Random Oversampling, Synthetic Minority Oversampling Technique (SMOTE), and adaptive synthetic (ADASYN). After analyzing the results, the Random Oversampling Accuracy technique was the best among the three techniques for balancing the dataset. The TPOT achieved the highest accuracy of 99% on the balanced dataset with Random Oversampling, while H2O reached an accuracy of 96.8%. However, H2O outperforms TPOT when applied to the unbalanced dataset. The results demonstrate that inappropriate handling of outliers reduces accuracy, particularly with imbalanced datasets. However, accuracy is improved after balancing it.