Phase identification in distribution grids with graph neural networks (GNN)

Many distribution system operators (DSOs) face the issue of unknown phase connections for the low voltage end-users. Knowledge of phases can enable phase balancing, reduce grid losses, detect unmetered demand and facilitate outage detection. In this study, a novel phase identification approach based on Graph Neural Networks (GNN) is presented. The method utilizes synthetic data generated through multiple Monte Carlo tests of power flow simulations. The accuracy of GNN is compared with the state-of-the-art techniques of k-Nearest Neighbors (k-NN) and least absolute shrinkage and selection operator (LASSO). From the assessment, it is concluded that the GNN outperforms both LASSO and k-NN under normal and noisy conditions, as in the case of PV penetration and unmetered consumption. Finally, the GNN presents robustness maintaining high accuracy of about 100%, when multi-phase loads are present.