M.S. AAI Capstone Chronicles 2024

ELECTRICITY DISTRIBUTION TOPOLOGY CLASSIFICATION

5

distribution falls within expected norms. However, should transformers with higher representation skew the results, a resampling strategy might be considered to ensure accuracy and fairness in the models.

Background Information Understanding the problem and defining the success criteria will help shape the model selection and evaluation. The optimal solution from previous research appears to be using a deep learning approach to perform this classification task using a specific transformer ID from a series of data points consisting of various meter numbers over several months at approximately hourly intervals. The temporal aspect of the dataset and nonlinear correlation of the power usage make the traditional machine learning models relatively ineffective at predicting which meter number may be connected to the transformer. Several studies have approached the transformer classification use case with varied features and methods but are attempting to reach a similar goal. In a study by Cook et al. (2022), the primary focus of determining power grid distribution includes density-based topology clustering methods to incorporate both voltage domain data and geographic space. Data was obtained by combining the temporal power usage data from a partner electric utility company and spatial data with GPS coordinates of transformers, poles, and smart meters. The data clustering approaches included grouping properties, bounding clusters with a limit, and understanding the importance of cluster “density” (observed in Cook Figure 5.) The K-means clustering methodology considers the average distances for each group member. Still, the geographical variance of curved streets or irregular terrain may need to be clarified for the K-means algorithm in the true centroid of the houses, as the distance calculation

96