Abstract
Electricity theft remains a critical issue in many developing nations, contributing to non-technical losses and making essential financial strain for utility providers. To overcome this problem, this paper presents an integrated deep learning pipeline that combines complementary preprocessing, feature transformation, and feature fusion techniques to improve detection of electricity theft. The framework provides a structured pipeline in which missing values are addressed using the piecewise cubic Hermite interpolating polynomial method, and class imbalance is handled by the synthetic minority oversampling approach. By evaluating different approaches based on multiple criteria, principal component analysis is employed for dimensionality reduction. The one-dimensional electricity consumption data is transformed into two-dimensional: temporal domain features using the Gramian angular field and spatial domain features using the maximal overlap discrete wavelet transform. To extract significant structural information for anomaly detection, separate two-dimensional convolutional neural networks are applied to each domain. The determined features are fused and passed using a deep neural network classifier. Experimental outcomes show the strength of the proposed framework, achieving an accuracy of 98.96% and an area under the curve of 99.59%. This demonstrates the model’s capability to detect fraudulent consumption using advanced preprocessing, transformation, and deep learning frameworks, contributing to Sustainable Development Goal 7 by enhancing energy efficiency and reducing non-technical losses in smart grids.