AI-Based Subscriber Classification in Power Distribution Networks Using Load Profiles

Abstract

Accurate classification of electricity consumers is vital for demand forecasting, anomaly detection, and tariff optimization in modern power systems. This study investigates the effectiveness of various machine learning and deep learning models in classifying electricity subscribers based on their active power consumption profiles. Using a real-world dataset containing over 21 million hourly consumption records from 25,476 subscribers in Southeastern Turkey, four consumer types were defined: General Subscriber, Lighting, Transformer, and Industrial Facility. Two modeling approaches were explored: (i) traditional machine learning models (e.g., LightGBM, Random Forest) using statistical and temporal feature extraction, and (ii) deep learning models (1D-CNN, CNN-LSTM, LSTM, ROCKET) trained directly on raw time series data. The 1D-CNN and ROCKET models outperformed others with a classification accuracy of 98%, while LightGBM achieved 96% accuracy among classical models. Evaluation metrics, including confusion matrices and F1 scores, indicated that 1D-CNN was particularly effective in learning temporal patterns from imbalanced data, especially in minority classes. Unlike most prior studies focusing on load forecasting, this work demonstrates how time series can be utilized for direct consumer classification. The results provide a practical and scalable framework for electricity distribution companies to enhance operational efficiency, consumer profiling, and energy management strategies. © 2025 IEEE.