Robust Methodology for Data-Driven Compact Learning Model-Free Voltage Calculation

Abstract

The ongoing energy transition demands that distribution system operators perform advanced calculations and analyses in near-real time. This is challenging in low-voltage distribution networks, where exact network topology is often unknown or insufficiently documented. Machine learning can bring significant advancement in addressing this problem, as it enables the use of smart meter measurement data to train various algorithms, such as deep neural networks, for effective model-free voltage calculations. Existing model-free approaches, that are reliant on machine learning algorithms training for input-output mapping, suffer from scalability issues when observed networks have high dimensionality and a lot of end-users. This paper assesses the robustness of various dimensionality reduction techniques, such as Principal Component Analysis, autoencoders and Isomap, in the compact learning model-free pipeline. The results demonstrate that all three of the observed dimensionality reduction methods produce fast, scalable and accurate voltage estimation in the compact learning model-free voltage calculation context, that is independent of network topology at any point. Final validation of the results produces an average estimation error below 1 V on a real Croatian distribution network, which can be considered negligible in real-world application.