M.S. AAI Capstone Chronicles 2024

ELECTRICITY DISTRIBUTION TOPOLOGY CLASSIFICATION

8

Experimental Methods

Baseline Models To have experimental models evaluated, a baseline model must be used to perform the classification task on the dataset and gauge progress. After researching the current approach from other studies, this project will take the same approach conducted by a team from RMIT University, titled "Prediction of Customer Connectivity to Distribution Transformer via Application of Machine Learning Techniques to Smart-Meter Data," which employed the Euclidean distance measure to calculate similarity values between each smart meter's voltage profile and a reference meter from each transformer, which are then used to train a cubic SVM classifier (Tennakoon, R. N. et al., 2020). The limitation of taking this approach is that the dataset from this project does not match the research dataset. In Tennakoon, R. N. et al. 's research, they have 710 smart meters connected to five transformers, and there is an average of 142 meters of data per transformer. The dataset used by this project is based on 1258 meters connected to 262 transformers, 4-8 meters per transformer. Data quantity per transformer is limited. Therefore, this dataset may have a limitation on performance accuracy compared to the dataset used in the research of Tennakoon, R. N. et al. This approach was still selected as a good baseline for horizontal performance comparison with experimental deep learning models. The development of the baseline model involves comparing the voltage profiles of various meters with the average voltage profiles of transformers to assess the alignment of each meter's voltage patterns with typical transformer voltages. The process begins by merging meter data into a data frame that includes minimum, maximum, and average voltages for different meters and transformers, grouped by "RoundHour" to reflect average voltage profiles for each transformer during specific periods. A similarity scores dictionary is then initialized. The

99