IEEE Access (Jan 2024)

Electricity Theft Detection in Smart Grids Based on Omni-Scale CNN and AutoXGB

  • Sanyuan Zhu,
  • Ziwei Xue,
  • Youfeng Li

DOI
https://doi.org/10.1109/ACCESS.2024.3358683
Journal volume & issue
Vol. 12
pp. 15477 – 15492

Abstract

Read online

Electricity theft is a prevalent global issue that has detrimental effects on both utility providers and electricity consumers. This phenomenon undermines the economic stability of utility companies, worsens power hazards, and influences electricity costs for consumers. The advancements in Smart Grid technology play an essential role in Electricity Theft Detection (ETD), as they generate large amounts of data that can be effectively utilized for ETD through the application of Machine Learning (ML) and Deep Learning (DL) methodologies. The present study presents a novel approach for ETD by combining Omni-Scale CNN (OS-CNN) and AutoXGB. Firstly, the Piecewise Cubic Hermite Interpolating Polynomial (PCHIP) is employed as the data interpolation technique to address the limitations and missing data in the dataset. Additionally, a combination of the Synthetic Minority Over-Sampling Technique (SMOTE) and the Edited Nearest Neighbors (ENN), known as SMOTEENN, is utilized for data resampling to tackle the issue of class imbalance in the dataset. Secondly, the multi-layer Omni-Scale block stack is employed to effectively cover the receptive fields of diverse time series scales based on a straightforward rule. This enables the One-dimensional Convolutional Neural Network (1D-CNN) to acquire enhanced learning capabilities for both irregular electricity consumption data anomalies and periodic normal electricity consumption patterns in smart grid datasets, facilitating superior extraction of essential data features. The AutoXGB classifier is then utilized to classify the extracted features. AutoXGB possesses the capability of automatically optimizing the hyperparameters required by the model, ensuring that the classification model maintains optimal accuracy and settings. Finally, the method exhibits superior competitiveness compared to other methods on the same dataset. The experimental results demonstrate that the proposed model achieves an accuracy rate of 99.2%, a precision rate of 97.5%, and an area under the ROC curve of 98.4%. These results establish its significant superiority over alternative models.

Keywords