An Integrated CNN-Bidirectional LSTM Model to Detect Electricity Theft Users in Smart Grids

Nowadays electricity theft is taking place in electricity grid throughout the world. It creates a significant threat to the economic operation of power systems and to genuine consumers. Simultaneously, it causes a non-technical loss to the power distribution system. Recently, Advanced Metering Infrastructures (AMI) is used to collect data from the smart meters. In this approach, a theft detection method is presented using a hybrid Recurrent Neural Network (RNN) and Bidirectional Long Short-Term Memory (Bi-LSTM). This method uses the features selected through minimum Redundancy Maximum Relevance (mRMR) scheme from the comprehensive time and frequency domain features extracted in the feature extraction stage. The challenges in the data set like missing data and limited availability of data are addressed through data re-sampling and data interpolation. We conducted experiments in the reduced feature set using RNN + Bi-LSTM classification model to achieve a better accuracy. The performance of this method is validated by comparing it with other existing results executed on the same dataset. On validation, this method achieved an accuracy of 93.6% and AUC of 93%.