Reconstructing Digital Twins of Power Distribution Networks using Machine Learning Techniques

Abstract

The reconstruction of digital twins for power distribution networks using machine learn- ing techniques presents a promising approach to enhancing the efficiency and reliability of modern energy systems. This thesis explores methodologies to enhance the creation of ac- curate digital representations of medium voltage/low voltage (MV/LV) power distribution networks by leveraging advanced machine learning models. The primary objective is to correct wrong or missing feeder connections, which is crucial for establishing the topology between a client and its corresponding transformation point. To achieve this, we first built a dataset from scratch, identifying relevant features that could be useful for a machine learning model to output probabilities for each potential feeder. Our problem is framed as a classification task, where the goal is to predict a likeli- hood probability for each candidate feeder and select the one with the highest probability. Once a feeder connection is proposed for a given client, the optimization algorithm de- scribed in Equation 16a computes the topological path. If a path satisfying all constraints is found, the client’s feeder information is corrected successfully and integrated into the power distribution network. We investigated the effectiveness of Logistic Regression, Random Forest, XGBoost, and Neural Network models as potential solutions. An in-depth analysis of these models was conducted to evaluate their respective performances and robustness, delving into the de- cision processes leading to their predictions. Our examination led to the selection of the Random Forest model due to its superior performance, achieving a macro average of 78% Precision, 80% Recall, 79% F1-score, and 94% accuracy. We also discussed challenging customer cases where no model could immediately pre- dict the correct feeder connection. Despite these challenges, we successfully corrected 94% of the clients with missing or faulty feeder information. Finally, we integrated this project into the production code, enhancing the digital twin of the power distribution network and improving its overall accuracy and reliability.