M.S. AAI Capstone Chronicles 2024

ELECTRICITY DISTRIBUTION TOPOLOGY CLASSIFICATION

9

Euclidean distance between each meter and transformer voltage profile pair is calculated for aligned profiles across various periods, using minimum, maximum, and average voltage values. These distances are averaged to derive an overall distance metric for each meter relative to each transformer. This overall distance is converted into a similarity score by taking its inverse. Finally, a Support Vector Classifier (SVC) model is trained and evaluated on this meter/transformer voltage similarity data, with an 80/20 split between training and validation data sets. In an attempt to provide additional insight for baseline testing, a simple Random Forest Tree Classifier was trained and evaluated. The data preparation followed a similar approach to the model listed previously and focused on the entirety of available features from the provided dataset. This model was designed to offer a simple classification strategy given the available features from the original dataset and several engineered features extracted from various fields within the dataset. This model design needs to account for the complexity of the temporal sequencing aspect of the entries, which is likely why the accuracy could have been higher. This allows for comparing gauging results when the timestamps are incorporated in a more

encapsulated method. Transformer Model

This project's experimental transformer model architecture was inspired by the transformer model approach designed by Wijkhuizen, M., in the Kaggle sign language competition (Wijkhuizen, M., 2023). However, due to the data type difference, Wijkhuizen, M. 's transformer architecture could not be directly followed. Using a similar idea (encoder transformer only), a new transformer model architecture design was created from scratch.

100