Electricity theft detection in unbalanced sample distribution: a novel approach including a mechanism of sample augmentation

• Published in: Applied intelligence (Dordrecht, Netherlands) Vol. 53; no. 9; pp. 11162 - 11181
• Main Authors: Yao, Ruizhe, Wang, Ning, Ke, Weipeng, Chen, Peng, Sheng, Xianjun
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.05.2023; Springer Nature B.V
• Subjects: Accuracy; Artificial Intelligence; Artificial neural networks; Behavior; Computer Science; Datasets; Design; Electric power distribution; Electricity; Electricity distribution; False alarms; Feature recognition; Generative adversarial networks; Machines; Manufacturing; Mechanical Engineering; Neural networks; Processes; Smart grid; Support vector machines; Theft; Unbalance; CT-WGAN; Bridged MCNN-BiGRU; Non-technical losses; Smart grids; Electricity theft detection
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-022-04069-z

Electricity theft is a major cause of non-technical loss (NTL) in smart grids. However, existing research on electricity theft detection (ETD) lacks a generalized analysis of the characteristics of theft behaviors and fails to effectively address the problem of unbalanced sample distribution; the electricity theft features they recognize are also singular. To address these problems, a novel approach consisting of a sample augmentation mechanism based on convolutional transformer-Wasserstein generative adversarial networks (CT-WGANs) and an electricity theft detection scheme using bridged multiscale convolutional neural network-bidirectional gate recurrent units (MCNN-BiGRUs) is proposed in this paper. First, the generalized characteristics of electricity theft behavior in multiple time dimensions are analyzed, and the data slices are constructed. Then, with the aim of reducing the influence of unbalanced sample distribution, CT-WGAN, which focuses on the generalized characteristics of electricity theft in various time dimensions, is designed and has better sample generation capability. Finally, a bridged MCNN-BiGRU is proposed to recognize the temporal, transition, and persistence characteristics of electricity theft to improve the efficiency of ETD. Experimental results on the State Grid Corporation of China (SGCC) and Irish Smart Energy Trial (ISET) datasets show that the proposed approach outperforms traditional schemes in terms of the area under curve (AUC), precision, recall, f1-score, and accuracy.