Non-technical loss identification by consumption profile analysis

Abstract

In this thesis the non-technical loss (NTL) identification problem is addressed both theoretically and in practice by the evaluation of a method. The theoretical part starts by characterizing the NTL problem, encompassing its definition, causes and impacts to society. In the sequence, the methods for its identification are reviewed, with special emphasis on methods based on consumption profile analysis. Then, some basic aspects of classification theory as a tool for probabilistic inference are presented, including performance metrics for problems with severe class imbalance. The theoretical part concludes with an overview of Neural Networks principles and their uses for binary classification problems. The method evaluation part starts with the preparation of the selected dataset, and includes a novel method for the treatment of missing values based on interpolation that is capable of better preserving the processes’ statistical moments. In the sequence, statistical information is extracted and diverse features are examined in respect to their capabilities of discrimination between the classes. Lastly, a published method is implemented and its performance is compared to a modified version which includes the features previously identified, and it is concluded that an improvement is achieved, especially in respect to the generalization capability.