Exploiting autoencoders for three-phase state estimation in unbalanced distributions grids

• Published in: Electric power systems research 2015-06, Vol.123, p.108-118
• Main Authors: Barbeiro, P.N. Pereira, Teixeira, H., Krstulovic, J., Pereira, J., Soares, F.J.
• Format: Article
• Language: English
• Subjects: Autoencoders; Distribution grids; Neural networks; Smart metering; Three-phase state estimation; Unbalanced loads
• ISSN: 0378-7796; 1873-2046
• DOI: 10.1016/j.epsr.2015.02.003

•Innovative method for complete three-phase state estimation in distribution grids.•The methodology is based on the use of a particular kind of ANN—the autoencoders.•Autoencoders use data from smart meters to learn the behavior of the grid.•Autoencoders avoid the need of characterizing the grid parameters and topology. The three-phase state estimation algorithms developed for distribution systems (DS) are based on traditional approaches, requiring components modeling and the complete knowledge of grid parameters. These algorithms are capable of dealing with the particular characteristics of DS but cannot be used in cases where grid topology and parameters are unknown, which is the most common situation in existing low voltage grids. This paper presents a novel three-phase state estimator for DS that enables the explicit estimation of voltage magnitudes and phase angles in all phases, neutral, and ground wires even when grid topology and parameters are unknown. The proposed approach is based on the use of auto-associative neural networks, the autoencoders (AE), which only require an historical database and few quasi-real-time measurements to perform an effective state estimation. Two test cases were used to evaluate the algorithm's performance: a low and a medium voltage grid. Results show that the algorithm provides accurate results even without information about grid topology and parameters. Several tests were performed to evaluate the best AE configuration. It was found that training an AE for each network feeder leads generally to better results than having a single AE for the entire system. The same happened when different AE were trained for each network phase in comparison with a single AE for the three phases.