M.S. AAI Capstone Chronicles 2024

ELECTRICITY DISTRIBUTION TOPOLOGY CLASSIFICATION

15

displayed training accuracy topping out at 80%. In contrast, the validation accuracy maxed out at 45%, which can be observed in Figure 8. Including an additional dropout layer helped alleviate the overfitting of training data. However, the model was still prone to overfitting. After approximately 200 epochs, the model begins to overfit more heavily, as observed in the Transformer model training process. The CNN model design would likely improve with increased training data instead of reducing the model’s complexity and requiring additional Based on the baseline model testing result, the traditional machine learning method under current data conditions achieves 0.65 accuracy compared to over 0.90 accuracy in the original research. This indicates that more meters of data representation will be needed for higher accuracy. The test showing the potential of the deep learning transformer model that is based on timestamp voltage data will only learn from the data; however, based on the nature of deep learning, it usually requires more data than the traditional machine learning data method, the potential of the deep learning model like transformer could benefit the electricity distribution topology classification task. The traditional machine learning method requires a long computation time on the similarity score; the deep learning method would be a better solution than the traditional machine learning method. One possible accuracy improvement on the deep learning model is mainly from more AMI meters deployed in the field, and more data needs to be collected from the meter. Another could be testing a pre-trained deep learning transformer model to perform transfer learning, where the base model will be trained and have an initial weight assigned to the pre-trained model. The additional data would enhance the model performance without overfitting through dropout layers and regularization. Conclusion and Recommendations

106