Transfer Learning-Driven Electricity Theft Detection in Small-Sample Cases

Wenlong Liao,Ruijin Zhu,Abdulrahman Takiddin,Muhammad Tariq,Guangchun Ruan,Xueyuan Cui,Zhe Yang

DOI: https://doi.org/10.1109/tim.2024.3470061

IF: 5.6

2024-10-19

IEEE Transactions on Instrumentation and Measurement

Abstract:In recent years, there has been a growing trend in the utilization of machine learning techniques for electricity theft detection. However, most existing works share a strong hypothesis that the dataset is sufficiently large (i.e., tens of thousands or more samples) to train effective models, which is not applicable to small-sample cases commonly encountered in practical engineering. To address this issue, a transfer learning-driven approach is proposed to boost the detection accuracy of electricity theft in small-sample cases. Specifically, an overall framework and six strategies are proposed to transfer a model trained in one location called the source domain with many samples to another location called the target domain with few samples for electricity theft detection. These strategies are comprehensively investigated on widely used models and real datasets. Simulation results demonstrate that the proposed approach can improve the area under curve (AUC) and mean average precision (MAP) by about 3%–6% and 5%–24%, respectively, for small-sample cases. Moreover, the proposed approach is model-agnostic and, thus, applicable to various models (e.g., multilayer perceptron (MLP), gated recurrent unit (GRU), convolutional neural network (CNN), and their variants).

engineering, electrical & electronic,instruments & instrumentation