M.S. AAI Capstone Chronicles 2024

ELECTRICITY DISTRIBUTION TOPOLOGY CLASSIFICATION

3

learning methods and just smart meter voltage data (Tennakoon et al., 2020). Moreover, in their study, Cook et al. point out that smart meter voltage data combined with meter GPS data will provide 97.87% accuracy on topology identification with unsupervised learning (Cook et al., 2022, p10). This project explored two deep-learning solutions for the topology classification system based on only AMI meter interval voltage data. This project used real-life AMI voltage data from the utility (SaskPower, 2024), which is formatted in hourly intervals of V max , V min , and V avg data from 1228 AMI smart meters for the time length of three months (November 2023, December 2023, and January 2024). Also, SaskPower provided the topology ground truth labels. The project's ultimate goal would be to create a deep-learning classification model that could identify the meter-to-transformer connection topology. Therefore, SaskPower could use the model to verify the GIS system and improve data quality.

Data Summary The dataset received from SaskPower contained 1619 files, including 1618 meter voltage

reading files consisting of time stamps, measurement types, and voltage readings between

November 1st, 2023, and January 31st, 2024. The dataset includes a ground truth table linking

meters to transformers and 1618 files of meter voltage readings. Each file contains data on the

type of measurement, the time of the reading, and the voltage measurement itself. This setup

allows for identifying the transformer connected to each meter. The dataset analysis revealed

several issues related to meter voltage measurements that need addressing before model training.

First, the timestamps recorded for voltage readings must be more consistent and often

missing, with gaps accounting for less than 5% of the data (Figure 1.). To manage this, an hourly

94