Employing Feature Extraction, Feature Selection, and Machine Learning to Classify Electricity Consumption as Normal or Electricity Theft

Abstract

One of the main causes of revenue loss in the energy sector across the globe has been non-technical losses. Electricity theft is a non-technical loss that affects the power supply’s quality. Detecting electricity theft is crucial for conserving energy and making use of it effectively. This research proposes a method through which higher accuracy in electricity theft detection can be obtained using fewer features. Different feature extraction and feature selection techniques are examined to find the best method for selecting the features that are more relevant in electricity theft detection. Various experiments are carried out using feature selection and feature extraction methods, such as mutual information, low variance filtering, and Principal Component Analysis. Various machine learning-based classifiers are used that include Random Forest, Support Vector Machine, K-Nearest Neighbours, Naive Bayes, and Decision Tree. Results of the experiments are tabulated based on standard performance measures, namely accuracy, precision, recall, and Area Under the Receiver Operating Characteristic Curve (AUC-ROC) score. According to experimental findings, the Random Forest classifier with 30 components for PCA outperformed other methods by producing the best accuracy of 95.82%, recall of 0.938, and AUC-ROC-score of 0.989. © 2023, The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd.