Robust resampling and stacked learning models for electricity theft detection in smart grid

Abstract

Electricity theft (ET) is a critical contributor to non-technical losses (NTLs) that significantly threaten the efficiency and reliability of power grids, leading to increased power wastage and financial losses. Despite the development of various artificial intelligence (AI)-based machine learning (ML) and deep learning (DL) approaches for electricity theft detection (ETD), existing methods often exhibit limitations in memorization and generalization, mainly when applied to large-scale electricity consumption datasets characterized by high variance, missing values, and complex nonlinear relationships. These challenges can result in models needing high variance and bias, reducing their effectiveness in accurately predicting electricity theft cases. To address these limitations, we propose a three-layer framework that employs a stacking ensemble model to combine the benefits of both ML and DL algorithms. During the first stage of data preprocessing, missing data is imputed through data interpolation, while the normalization is done through min–max scaling. To solve the high-class imbalance problem prevalent in most real-world datasets, we combine borderline synthetic minority oversampling techniques and near-miss undersampling strategies. In the final layer of our proposed ETD framework, we employ four ML base and five meta-classifiers. The outputs of base classifiers are aggregated and passed to a meta-classifier, where we evaluate recurrent neural networks (RNN) and convolutional neural network (CNN) as potential meta-classifiers. The RNN are long short-term memory (LSTM), gated recurrent unit (GRU), Bi-directional LSTM (Bi-LSTM) and Bi-directional GRU (Bi-GRU), respectively. Experimental outcomes show that the proposed Bi-GRU better achieves accuracy enhancement of detection in general than meta-classifiers and other state-of-the-art models used for ETD.