An approach to estimating distribution network losses: Clustering to overcome the challenge of load uncertainty and measurement error

Accurate estimation of energy losses in distribution networks remains challenging due to uncertain and time-varying consumer load profiles and measurement errors in smart-meter data. This paper proposes a robust and data-driven framework for distribution-network loss estimation based on K-means++ clustering. A nonlinear feeder-loss model is formulated, and the loss coefficient is estimated from representative clusters formed using smart-meter and master-meter measurements, thereby reducing the influence of individual measurement uncertainties. The proposed framework is evaluated on the IEEE 15-bus, 33-bus, and 141-bus distribution networks using quasi-realistic 30-day load profiles with 15-minute resolution and is benchmarked against Monte Carlo simulation. The proposed method achieves MAE/RMSE values of 1.0653/3.1898, 19.8914/32.4590, and 28.4989/43.1960 for the 15-, 33-, and 141-bus systems, respectively, with corresponding MAPE values of 0.89%, 8.96%, and 4.60%. Sensitivity analyses further demonstrate stable performance under different K-means++ initializations and cluster-weight perturbations. Moreover, the method maintains consistent loss-coefficient estimates under measurement-error levels ranging from 2% to 20%, symmetric and non-Gaussian error distributions, strongly skewed errors, and outlier-contaminated measurements. Comparative results with OLS, CPLS, and RLS demonstrate the superior robustness of the proposed framework under challenging measurement conditions. These results confirm that the proposed approach provides an accurate, robust, and scalable solution for distribution-network loss estimation under realistic load and measurement uncertainties. • A novel loss estimation approach for electrical distribution networks is proposed based on clustering method. • A nonlinear model for feeder loss is presented, with can accurately estimate the network losses. • The loss is estimated using the aggregated data in each cluster, as well as the agent of all clusters. • The performance is evaluated compared to the Monte Carlo method on the 15-, 33-, 141-bus IEEE systems. • The method is resilient to uncertainties stemming from consumer load profiles and measurement errors.