Home > Research > Publications & Outputs > Robust Data Driven Analysis for Electricity The...

Electronic data


    Accepted author manuscript, 5.25 MB, PDF document

    Available under license: CC BY

  • Acceptance email IEEE TPS

    Other version, 749 KB, PDF document

    Available under license: Unspecified


Text available via DOI:

View graph of relations

Robust Data Driven Analysis for Electricity Theft Attack-Resilient Power Grid

Research output: Contribution to Journal/MagazineJournal articlepeer-review

<mark>Journal publication date</mark>31/01/2023
<mark>Journal</mark>IEEE Transactions on Power Systems
Issue number1
Number of pages12
Pages (from-to)537-548
Publication StatusPublished
Early online date25/03/22
<mark>Original language</mark>English


The role of electricity theft detection (ETD) is critical to maintain cost-efficiency in smart grids. However, existing ETD methods cannot efficiently handle the sheer volume of data now available, being limited by issues such as missing values, high variance and non-linearity. An integrated infrastructure is also required for synchronizing diverse procedures in electricity theft classification. To help address such problems, a novel ETD framework is proposed that combines three distinct modules. The first module handles missing values, outliers, and unstandardized electricity consumption data. The second module employs a newly proposed hybrid class balancing approach to deal with highly imbalanced datasets. The third module utilises an improved artificial neural network (iANN) based classification engine, to predict electricity theft cases accurately and efficiently. We propose three distinctive mechanisms, including hyper-parameters tuning, regularization and skip connections, to improve the performance of standard ANN to handle more complex classification tasks using smart meter (SM) data. Furthermore, various structures of iANN are investigated to improve the generalization and function fitting capabilities of the final classification. Numerical results from real-world energy usage datasets confirm that the proposed ETD model has superior performance compared to existing machine learning and deep learning methods, and can effectively be applied to industrial applications.