Smart Grid Electricity Fraud Detection Using Machine Learning

With the digitization of every aspect of our lives, the smart grid is a new addition. It mainly enables two-way communication between consumers and electric companies. It ensures the network's ability to increase safety, security, resilience, and stability using sensors and smart meters. It can also monitor demand-supply changes. Every year, electricity-providing companies suffer huge losses all over the world. It can be technical loss that happens through power dissipation on the transmitters, substantial equipment, and overhead power lines. But the non-technical losses are caused mostly by the electricity thieves. If electricity consumption data are observed over a large time period, it is seen that honest regular customers have a certain pattern of electricity consumption. Whereas electricity thieves don't have any particular pattern, and their consumption is very irregular. By finding these anomalies in the consumption patterns, electricity fraud can be detected. This paper presents an improved model of state-of-the-art machine learning algorithms to detect electricity fraud with much better results. It proposes a comprehensive framework that leverages supervised learning techniques to analyze data collected from smart grids. This work uses the machine learning model 1D CNN to identify anomalous consumption patterns indicative of fraudulent activity. Additionally, it evaluates the performance of this algorithm against existing algorithms and literatures based on accuracy, RMSE, MAE, F1-Score and AUC using a dataset collected from the State Grid Corporation of China that includes both legitimate and fraudulent consumption records. This 1D CNN model outperforms 2D CNN, ANN, XGBoost, random forest, decision tree, and logistic regression with an accuracy of 95.47% and an F1 score of 97.42%. This work effectively shows how machine learning can be effectively incorporated with smart grids to prevent electricity fraud.