We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
A new electricity theft detection method using hybrid adaptive sampling and pipeline machine learning.
- Authors
Tripathi, Ashish Kumar; Pandey, Avinash Chandra; Sharma, Nikhil
- Abstract
Electricity theft not only results in higher electricity costs for regular paying customers but is also a safety threat to the public due to illegal power connections made for cheating. Many governments and private companies have suffered substantial losses from electricity theft or fraud. Several approaches have been presented in the literature for electricity fraud detection. The existing methods for electricity fraud detection show poor performance as the electricity datasets are generally unbalanced or skewed. To handle skewness issues, SMOTE and ADASYN-based oversampling techniques are employed in some electricity fraud detection methods. However, existing oversampling methods typically increase the overlapping of classes and introduce additional noise. Therefore, in this paper, a new oversampling method named ADASYN-SGWO has been presented. The proposed ADASYN-SGWO employs ADASYN for oversampling and SSO-GWO for intelligent undersampling, which mitigates class imbalance issues in electricity datasets. After reducing the issue of class imbalance, random forest pipeline has been used to detect electricity fraud. To assess the efficiency of the proposed approach, precision, recall, F1 score, accuracy, Matthews correlation coefficient, and balance classification rate of the proposed and other state-of-the-art, including XG-boost classifier, Logistic regression, Gaussian naive Bayes, Random Forest classifier, Ada Boost classifier, LGBM classifier, and Quadrant discriminative analysis, have been computed and compared. Further, the proposed approach is also statistically validated using Friedman and post hoc analysis. The statistical analysis and experimental outcomes witnessed that the proposed ADASYN-SGWO-PRF method has outperformed all the other considered methods with 98.2% accuracy, which is significant for real-time applications of electricity theft detection problems.
- Subjects
ELECTRICITY; FRAUD investigation; THEFT; RANDOM forest algorithms; FRAUD
- Publication
Multimedia Tools & Applications, 2024, Vol 83, Issue 18, p54521
- ISSN
1380-7501
- Publication type
Article
- DOI
10.1007/s11042-023-17730-7