Neural network-based three-phase state estimation for unobservable low voltage grids

Abstract

Building a smart grid at the low-voltage (LV) level by deploying numerous measuring devices is technically and economically challenging. With advanced metering infrastructures still in development, learning-based methods for state estimation present a cost-effective alternative for monitoring LV systems. This work introduces a three-phase, four-wire neural network-based estimator tailored for asymmetrical and unbalanced LV grids coupled with a three-phase probabilistic power flow method to generate synthetic training sets in the absence of historical measurement data. Validation against real pilot data from previously published work demonstrates promising results in detecting voltage fluctuations, ultimately enhancing observability in LV systems. Additionally, a sensitivity analysis conducted using a high-performance computing cluster determines the optimal placement of an additional measurement device to improve estimator performance. This ANN-based estimator, with augmented input measurements, can predict severe voltage drops in heavily loaded phases and voltage rises in neutral conductors due to imbalances, with a mean absolute error of less than 0.5V (0.0022p.u.) in the analyzed case study.

Keywords

• Grid monitoring
• Observability
• Distribution networks
• State estimation
• Neural networks