This chapter presents an improved hybrid architecture for the intelligent diagnosis of non-technical losses in electrical distribution networks, based on the profiling of energy service users. The research builds upon the previous CNN–LSTM–DNN–XGBoost model, which achieved an accuracy of 85.62%, and proposes an optimized version called CNN–CBAM–BiLSTM–DNN–XGBoost, designed to reduce false positives and improve the system’s generalizability. The new approach integrates spatial and channel attention mechanisms using the CBAM (Convolutional Block Attention Module), captures bidirectional temporal dependencies with BiLSTM, and incorporates a calibrated, cost-sensitive XGBoost classifier that can dynamically adjust decision thresholds according to ROC metrics and asymmetric loss functions. This strategy weights classification errors based on class frequency and operational impact, prioritizing the correct identification of fraud and anomalies in contexts with high data imbalance. The experimental results demonstrate a weighted accuracy of 96.3%, with AUC values exceeding 0.90 in all classes, showing a substantial improvement over the baseline model. Overall, the proposal outlines a robust methodological framework for automating the detection of non-technical losses, thereby enhancing operational efficiency and supporting informed decision-making within energy sector companies.