Design and Evaluation of Low Voltage Neural Network-Based State Estimators in Scenarios With Minimal Measurement Infrastructure

Learning-based state estimators can represent a cost-effective opportunity for distribution system operators to perform grid monitoring and control in low-voltage grids where the measuring infrastructure is minimal, if not absent. This study lays the foundation for designing and evaluating neural network-based state estimators for low-voltage radial distribution grids. A simulation-based methodology is proposed for generating synthetic training data-sets relying only on minimal grid data. Additionally, a novel framework for performance analysis of low voltage learning-based state estimators is considered, which relies on a bi-dimensional evaluation of the absolute error and the parallel observation of relative metrics. The applicability and potential of these estimators have been tested and validated through various low-voltage radial case studies, showing promising results especially for large distribution grids. Finally, a propagation error study has been conducted to observe how these estimators handle errors in input measurements.