Energy Theft Detection in Advanced Metering Infrastructure Using 1Directional-Convolutional Neural Network and Bidirectional Long-Short Term Memory

Abstract

Electricity theft poses a significant challenge to power utilities worldwide, resulting in substantial economic losses and operational inefficiencies. This study develops a hybrid deep learning system for detecting and confirming energy theft in Advanced Metering Infrastructure (AMI) using 1D Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) networks. The methodology processes three months of half-hourly consumption data for 25 customers, aggregating the data into daily sequences and simulating theft patterns for training purposes. The primary objective of the research is not to establish universal generalization across all electricity consumers, but rather to evaluate the feasibility and effectiveness of the proposed deep learning framework for theft detection under controlled conditions. The dataset contains high-resolution temporal consumption records, which significantly increases the number of individual training instances despite the limited number of customers.  The 1D CNN extracts local features from 14-day consumption sequences, while BiLSTM captures bidirectional temporal dependencies. An LSTM autoencoder detects anomalies through reconstruction errors, with ensemble prediction fusing outputs for final classification. The results show that the proposed ensemble system achieved an accuracy of 92.47%, a precision of 90.78%, a recall of 93.81%, and an F1-score of 92.27%. When compared with existing CNN-LSTM and rule-based detection methods, the hybrid system outperforms them in both accuracy and false-positive reduction. This approach directly addresses existing gaps, including insufficient confirmation frameworks, a heavy reliance on synthetic datasets, and the inability of traditional methods to detect sophisticated patterns of theft.