M.S. AAI Capstone Chronicles 2024

ELECTRICITY DISTRIBUTION TOPOLOGY CLASSIFICATION

2

Introduction Utility companies increasingly adopt Advanced Metering Infrastructure (AMI) programs as a foundational component of their grid modernization efforts. AMI is also commonly referred to as a smart meter program. As a component of the smart grid, these AMI programs have recorded and transmitted the customer consumption and demand in the form of load profiles back to utilities. At the same time, more power quality data (voltage, current, and harmonic) was measured, timestamped, and transmitted to the utility by the AMI system. These AMI meter interval data can be utilized to create critical applications such as evaluation distribution line losses, load forecasting, and predictive equipment maintenance. However, utilities grapple with challenges related to the accuracy of Geographic Information Systems (GIS) due to poor data quality when deploying the critical applications mentioned above. For example, Pacific Gas and Electric Company (PG&E) has Electric Program Investment Charge (EPIC) project 3.20 research for predictive maintenance has pointed out that “GIS mapping of meters to transformers was found to be error prone” as one of the challenges that their data science team were facing (Pacific Gas and Electric Company, 2023). Distribution system topology must be solved before any critical application is created for utility using AMI system data. AMI voltage alone could be sufficient to solve topology with a suitable algorithm. Luan et al. have pointed out in their research paper that 1. similarity of meter voltage profiles and 2. relations between voltage decreases and downstream distance created based on Smart Meter voltage data could provide a viable solution for distribution topology verification (Luan et al., 2015, p1965). In further studies, another researcher used Smart Meter voltage data for distribution topology solutions with a different approach and great accuracy. For example, the research of Tennakoon et al. showed more than 90% accuracy using supervised

93