Non-technical losses (NTL) constitute a major issue in many countries, both developing and developed. NLT can be considered as a bad data detection problem. Thus, classical approaches like the weighted least square method and statistical tests can be used for the detection and identification of bad data. They are suitable tools when the topology of the network and its parameters are known. While this assumption is widely accepted in transmission grids, it is not clear to hold in distribution grids, where grid reconfiguration is common and parameter values have an important dependence on the ambient conditions.

In this paper, we leverage the latest advances in mathematical and computational tools to detect NTL in distribution grids. Thus, NTL detection can be implemented in an automated system that does not require human interaction. For dealing with it, we use off-the-shelf machine learning algorithms. In particular, we introduce a new architecture that combines a deep learning network and convolutional and recurrent neural networks.

A thoughtful set of simulations over a realistic dataset is performed and compared among other model-free machine-learning approaches, namely, support vector machine, random forest, and gradient boosted trees.