Artificial intelligence model and balancing techniques for controlling non-technical losses in the electrical energy sector

Non-technical losses (NTLs) in electrical power distribution systems represent a critical problem due to their economic and operational impact, as they are primarily associated with fraud, tampering, and administrative errors. This article proposes the design and evaluation of a hybrid deep learning model for the automatic detection of NTLs, integrating convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and dense neural networks (DNNs). To mitigate the class imbalance characteristic of this type of problem, synthetic resampling techniques were applied, specifically SMOTE, BorderlineSMOTE, and ADASYN, analyzing their effect on model performance. Validation was performed using a real dataset from a power distribution company in the municipality of Aguachica, Colombia, consisting of 44,231 users and 365 variables per record. The results show that the BorderlineSMOTE + CNN–LSTM–DNN combination achieves the best overall performance, reaching an accuracy of 74.47%, along with improvements in key metrics such as recall, F1 score, and AUC. These findings demonstrate that integrating deep architectures with advanced balancing techniques is an effective strategy for PNT detection in real-world electrical environments.