In this paper two different methods for non-technical losses (NTL) detection are analyzed and new approach is proposed, based on the noticed drawbacks. It is shown that NTL can be successfully detected by a neural network trained by “artificial”, i.e., generated samples. This approach eliminates the need for many hard-to-obtain real life samples and the network can easily be trained to detect some new, non-typical occurrences in the system. This makes the proposed solution suitable for large companies that supply many different consumers who possibly change their consumption habits.