Imaging Time-Series to Classify Energy Theft and Defective Meters in Automatic Meter Reading Using Convolutional Neural Networks

Abstract

Electricity distribution systems suffer significant non-technical losses (NTL) from theft and meter defects. While automatic meter reading (AMR) systems detect irregularities, they cannot distinguish specific anomalies. This study transforms AMR voltage and current time-series into 2D images using five methods; continuous wavelet transform (CWT), spectrogram plots (SGP), recurrence plots (RCP), Markov transition Field (MTF), and Gramian angular field (GAF), and classifies them into normal, defective, and theft categories via a lightweight convolutional neural network (CNN) with 5-fold nested cross-validation. On a real-world Provincial Electricity Authority (PEA) dataset, CWT achieves 95.3% accuracy and 0.95 AUC-ROC, outperforming other transformations, 1D NNs, and classical ML baselines. Grad-CAM confirms class-specific feature focus. Using only AMR data, augmentation, this practical, interpretable approach enables utilities to accurately differentiate theft from defects for targeted intervention.