Efficient Power Theft Detection using Smart Meter Data in Advanced Metering Infrastructure

Since its discovery, electricity is the basic necessity of human life. It has helped to boost the industrial revolution and any economy, especially the developing ones are highly dependent on a regular and cheap supply of power. But power theft is a major threat to electric companies, it is catastrophic to power distributors and is a reason for financial damage. Regular Manual Inspection of all customers is a solution but it is very time and cost consuming and also leaves scope for corruption. So, AMI (Advanced Metering Infrastructure) has been introduced, which acts as two-way communication between the supplying company and customer via a smart meter which automates the measurement of electric energy usage. But this new technique brings in the possibility of new kinds of thefts and the consumption data compiled from this AMI has the potential to help build a system for fraud detection. Here, a machine learning-based detection methodology is presented, which first applies GMM clustering to categorize data into consumption classes. Then we improve the working of an ANN-based Classification model using a softmax layer and ensembling. The robust dataset used here helps to train the model efficiently and it achieved an astonishing accuracy of 92%. Results indicate the surpassing performance of this method in comparison to other existing detectors.