Power distribution networks in the United States are typically divided into three phases – labeled A, B, and C – with individual homes connected to a single phase. The distribution system operators aim to balance demand across these phases, which requires accurate information of which homes are connected to each phase.

With more distributed energy resources (DERs) and electric vehicle (EV) charging coming online, making sure the system is balanced becomes more challenging. These sources disrupt current models for phase detection and lead to more changes at the edge of the network. Yet, network models are often not updated to reflect updates made during routine maintenance, storm restoration work or other urgent changes.

Manual inspection of phase connections is costly and time-consuming, particularly for Eaton’s rural utility customers, who may have many thousands of miles of distribution lines. Furthermore, it can be difficult to justify conducting these efforts proactively as phase errors typically only affect a small percentage of meters, and it can be difficult to identify this small subset of meters.

Phase Detection Tool uses data to target out-of-phase meters

Eaton’s Center for Intelligent Power was engaged to find a new, data-based approach to this vexing issue They developed a novel machine learning approach to correct the errors in the phase labels the utilities currently have by starting with Advanced Metering Infrastructure (AMI) data already routinely collected by utilities using Eaton smart meters. Their approach relied on the similarities in voltage profiles across meters on a common phase.

Using synthetic voltage data with known phase labels, we tested different approaches to exaggerate the within-phase similarities and the across-phase differences. This enabled us to build a clearer picture of the characteristics of each phase. Then, relying on the fact that the inaccuracies are relatively rare, we can adjust the phase of each meter to the most common currently assigned phase among meters sharing similar characteristics.

Testing and refinement

In collaboration with Grand Valley Power (GVP), we conducted field validation to verify our results. From a total of over 19,000 meters, we identified just over 1,000 meters to be inspected. GVP were able to import our results into their own GIS system and quickly identify line sections which had phasing issues. This combination of data helped us improve the functionality of our algorithm which had not previously accounted for the presence of voltage regulators. Voltage regulators, as the name implies, modulate voltage fluctuations. Having accurately labeled, field-validated data was invaluable in allowing us to augment our approach to address this new variable.