Segmentation of Energy Consumption Using K-Means: Applications in Tariffing, Outlier Detection, and Demand Prediction in Non-Smart Metering Systems

The management of energy demand in systems lacking smart metering presents a significant challenge for electric distributors, primarily due to the absence of real-time data. This research assesses the efficacy of the K-Means algorithm when applied to the monthly billing records of 221,401 residential customers from Empresa Eléctrica Ambato Regional Centro Norte S.A. (EEASA) (Ecuador) over the period 2023–2024. The methodology encompassed data cleaning, Z-score normalization, and validation employing the Silhouette (0.55) and Davies–Bouldin (0.51) indices. Additionally, linear regression (LR) and Random Forest (RF) models were utilized to forecast demand, with the latter yielding an R2 of 0.67. The findings delineated eight distinct clusters, facilitating the formulation of more representative rates, the identification of outliers through the interquartile range (IQR) method, and the enhancement of consumption estimation. It is concluded that this unsupervised segmentation approach constitutes a robust and cost-effective tool for energy planning in network environments devoid of smart infrastructure.