Accurate Detection of Electricity Theft Using Classification Algorithms and Internet of Things in Smart Grid

Electricity theft is one of the most significant factors among non-technical losses. Because of electricity theft, genuine users have to pay more, supply quality decreases, and generation load increases. Development in the Internet of Things-based sensors has changed the way to monitor the electricity consumption pattern of the consumers. This electricity consumption data are processed by classification algorithms to identify electricity theft. The electricity consumption data are imbalanced in nature. The objective of this study is to design a machine learning model considering that data imbalance issue is resolved using six data balancing techniques, namely Synthetic Minority over Sampling (SMOTE), Adaptive Synthetic Sampling (ADASYN), Random over Sampler, Support Vector Machine-Synthetic Minority over Sampling, SMOTEENN (Edited Nearest Neighbor) and SMOTE Tomek Links. The designed model consists of two stages. In first stage, twelve classification algorithms (Decision Tree, Adaboost, Extra Tree, Logistic Regression, XGBoost, Light GBM, Multi-Layer Perceptron, Bagging, Random Forest, Support Vector Machine, and Naive Bayes, K-Nearest Neighbor) are applied on the data balanced by six techniques. In the next stage, two ensemble techniques, namely maximum voting and stacking, are applied to the best five performing algorithms using Python language. Dataset from the State Grid Corporation of China is considered, and algorithms are compared based on accuracy, MCC (Matthews Correlation Coefficient), f1-Score, and log-loss. We observed that SMOTEENN with stacking ensemble algorithm gives the highest performance: accuracy value-97.67%, MCC value-0.9434, log-loss value-1.01, and f1-score value-97.88% among all the experiments performed. The proposed model reported around 3% higher performance than the results presented in the literature. Moreover, the effectiveness of the model is validated by applying the ANOVA (Analysis of variance) one-way statistical test.