Machine learning-based forecast of secondary distribution network losses calculated from the smart meters data

Abstract

With greater integration of smart meters, an opportunity to increase the observability of the traditionally unobservable low-voltage distribution networks is created. The purpose of such meters is mainly to collect and store data on end-user consumption for billing purposes, and it is these large data flows that open up a wide range of analyses to Distribution System Operators. Leading in this is the prediction of end-user consumption, which finds its application especially in determining network losses for more efficient planning and operation of distribution networks. Due to the complicated features of the collected load series data, the application of synthetic curves for the consumption forecasting problem is abandoned and energy utilities are turning to more complex solutions, most often based on machine learning algorithms. Therefore, this paper presents a machine learning-based model for forecasting losses in a low-voltage distribution network. Power flow simulation tools are frequently used to estimate and predict active power losses but are applied only in the case of available network topology and elements data. Hence, in this paper, special emphasis is placed on a model that does not rely on network data, but only on historical measurements collected from smart meters. The model is tested on a real-world distribution network with more than 150 end-users. The results show the effectiveness of the model in forecasting active power losses of the observed network, but also highlight the sensitivity of the model to errors, which is a good basis for the implementation of additional algorithms and variables as a means to enabling near real-time operation planning of distribution networks.