Fraud and Anomaly Detection in Electricity Usage Patterns Using BiLSTM

PLN, as the National Electricity Company in Indonesia, faces challenges from electricity theft, leading to financial losses and inefficiencies. Traditional fraud detection methods, with only 4.16% precision, have proven inadequate in addressing this issue. To enhance detection accuracy and streamline operations, this study proposes using a Bidirectional Long Short Term Memory (BiLSTM) model with L2 regularization by analyzing sequential electricity consumption patterns and addressing data imbalance through class weighting. The model achieved a fraud detection accuracy exceeding 6.42%, significantly outperforming traditional methods. The integration of L2 regularization further ensured robust generalization and reduced false positives. These results highlight the potential of AI driven approaches in combating electricity theft, offering a scalable solution to reduce non technical losses and protect revenues. Future research should explore developing hybrid models, incorporating smart meter data, and integrating realtime monitoring systems to improve hit rates and enhance detection accuracy while reducing operational costs.capabilities.