Convolutional neural network applied to detect electricity theft: A comparative study on unbalanced data handling techniques

Electricity theft is a problem that affects the efficiency and profitability of power companies. There are several studies and applications in order to detect electricity theft, including the use of artificial intelligence techniques and the most recent deep learning methods. For problems like it, the datasets utilized are completely unbalanced - consequently, the use of metrics as accuracy is not enough to properly evaluate the performance of the method for the application. In the present paper a Convolutional Neural Network (CNN) is applied to electricity theft detection problem using several techniques for balancing the classes of the dataset: Cost-Sensitive Learning, Random Oversampling, Random Undersampling, K-medoids based Undersampling, Synthetic Minority Oversampling Technique, and Cluster-based Oversampling. The objective is to compare and select the best unbalanced data-handling technique for CNN, utilizing a specific metric for problems with extremely unbalanced classes - the AUC (Area Under Receiver Operating Characteristic Curve). The results present that some techniques combined to CNN reach values of high quality, comparable to the obtained by other classifiers. Finally, the paper points studies related to electricity theft detection must deal with the unbalanced characteristic of the dataset in order to achieve better (or, in other words, correct) results.