M.S. AAI Capstone Chronicles 2024

ELECTRICITY DISTRIBUTION TOPOLOGY CLASSIFICATION

14

During the transformer model experiments, different data formats were used to test the possibility of increasing the data quality; for example, one of the tests was a separate three month data set every week that retained the week of the month and month of the year in the feature beside voltage measurements. The final transformer model training result was training accuracy at 0.60, validation at 0.40, and test accuracy at 0.375. The training accuracy and loss curve are shown in Figure 6. Model training accuracy continues to improve with the loss getting lower, which shows the transformer model is learning. Before 200 epochs, the validation accuracy closely follows the training accuracy. However, after 200 epochs, the validation accuracy and loss have plateaued, indicating the model is starting to overfit after the 200 epochs of training. This is after the drop off layer and other regularization methods have been added to the model. This indicated that the dataset would require more data from different meters in the same transformer to improve the granularity and model accuracy. This project has also explored the synthetic sample data argumentation on the dataset. By applying the Synthetic Minority Oversampling Technique (SMOTE) on the training dataset, the training and validation accuracy has been boosted up to 0.58, and the risk of overfitting has been reduced, as shown in Figure 7. This could also mean that if more new meter data are collected, the model accuracy could increase significantly. CNN Model The approach for the CNN model design followed a stricter preparation than standard image classification tasks, as the sequence of the pixels in the image needed to remain in the generated location and not shifted, flipped or rotated to preserve the temporal characteristics. With the limited dataset, the model was prone to overfitting, and the initial training runs

105