Electricity theft detection in IoT-based smart grids using a parameter-tuned bidirectional LSTM with pre-trained feature learning mechanism

Abstract

The most significant issue today is electricity theft (ET) which causes much loss to electricity boards. The development of smart grids (SGs) is crucial for ET detection (ETD) because these systems produce enormous amounts of data, including information on customer consumption, which can be used to identify ET using machine learning and deep learning (DL) techniques. However, the existing models majorly suffers with lower prediction accuracy because of over-fitting and dataset imbalancing issues. Therefore, to overcome these shortcomings, this paper proposes a novel DL approach for ETD in the Internet of Things-based SGs using parameter-tuned bidirectional long short-term memory (PTBiLSTM) with pre-trained feature learning model. The proposed system mainly comprises '4' phases: preprocessing, dataset balancing, feature selection, and ETD. Initially, the consumers’ electricity consumption data are collected from the theft detection dataset 2022 (TDD2022) dataset. Then, the data balancing is carried out by using Gaussian distribution, including fuzzy C-means approach to handle the imbalance data. Afterward, the meaningful features from the balanced dataset are extracted using the hard swish and dropout layer included residual neural network-50 (ResNet-50) model. Finally, the ETD is done, which utilizes a PTBiLSTM. The proposed models’ performance is evaluated using different performance metrics like accuracy, precision, recall, f-measure, the area under the curve, and kappa. The outcomes proved the efficiency of the proposed method over other related schemes in the ETD of SGs.