The paper considers state-of-the-art efficiency of data oriented methods for non-technical electricity losses detection. It compares the main mathematical methods for detection of non-technical energy losses and provides the metrics for their assessment, the features of initial data processing, and models (patterns) of electric energy theft. The most commonly used types of artificial neural networks and their settings for detection of non-technical energy losses are examined.A computational case study is conducted and reproduced step by step proposing a new modified method for non-technical energy losses detection based on an convolutional autoencoder. The results of the paper are urgent for the development of a software non-technical electricity losses detection in distribution grids.