1D-CNN and A Weight-Based Balancing Technique for Improved Detection of Non-Technical Losses in Power Distribution Systems

Non-technical losses (NTLs) in power distribution systems, caused by theft, fraud, or faulty equipment, undermine energy management and energy security policies. Detecting these losses is challenging due to data imbalance, as abnormal cases are far fewer than normal ones. This paper presents a novel approach combining a weight-based balancing technique with a one-dimensional convolutional neural network (1D-CNN). Load profiles were collected from the smart meter database of the Provincial Electricity Authority (PEA) in Khon Kaen, Thailand, covering seasonal variations—winter, summer, and rainy seasons—to ensure comprehensive pattern representation. The weight-based balancing technique leverages real-world insights of abnormal patterns of load profiles to generate synthetic abnormal data with realistic characteristics and proportions. The 1D-CNN is applied to time-series data, utilizing its ability to extract temporal features and classify normal and abnormal behaviors. Experimental results demonstrate that the proposed method significantly outperforms conventional classifiers, such as Random Forest, with an F1-score improvement from 65.2% to 89.6%, representing a 37% increase. This approach not only addresses the issue of data imbalance and enhances the NTL detection rate but also supports energy security policies and improving energy management in power distribution networks.