Comparative analysis of data transformation methods for detecting non-technical losses in electricity grids

• The energy consumption profile is transformed into a 2D representation (image). • The 2D representation uncovers features and patterns not captured in a 1D sequence. • Data transformation improves the detection of non-technical losses. • The classification methods’ performance varied based on the transformation used. • The quarterly consumption profile outperforms weekly and monthly profiles. Non-technical losses (NTL) pose a significant challenge for power companies, necessitating effective detection to minimize financial losses and improve energy system operations. Despite various proposed methods, effectively classifying normal and abnormal consumption patterns remains challenging. While feature-based detection methods perform well, they often require manual feature engineering and domain expertise. Convolutional neural networks (CNN) have emerged as effective tools for automatically extracting features from raw data, but raw data often lacks the structure needed for optimal feature extraction. This study explores various methods for transforming energy consumption patterns into two-dimensional (2D) representations to enhance feature extraction and NTL detection. Six encoding techniques for time series data were evaluated: Markov transition fields (MTF), Gramian angular summation field (GASF), Gramian angular difference field (GADF), Recurrence plots (RP), and time–frequency analysis methods, including short-time Fourier transform (STFT) and continuous wavelet transform (CWT). These transformations aim to improve the ability of CNN to identify meaningful patterns. The research findings demonstrate that data transformations significantly enhance CNN performance, reducing dependence on manual feature engineering. Performance improvements ranged from 0.3% to 6.4% when transitioning from raw energy consumption data to 2D transformed data. The high accuracy and low false positive rates enable utilities to recover unbilled revenue and reduce inspection costs, amplifying savings compared to raw data or traditional approaches. This study highlights the importance of data transformations in improving the performance of CNN for NTL and provides valuable insights for researchers and practitioners in the field.