Data-driven for neutral-to-ground voltage: an evaluation of dataset using deep learning.

Excessive neutral-to-ground voltage (NTGV) poses safety and operational risks in TT-grounded secondary distribution systems, particularly when caused by ground faults. This study investigates the effectiveness of data structure and density on a deep learning-based approach for classifying NTGV events as originating from either upstream or downstream fault locations. A deep learning model using LSTM is trained and tested using real-world datasets collected from multiple locations in a TT system. Results show that input data structure S3 which incorporating ground and phase current signals alongside NTGV measurements significantly improves model performance, achieving an accuracy of 98.15% and an F1-score of 97.85%. Notably, the proposed model maintains high classification accuracy even when trained using only 10% to 30% of the available dataset, demonstrating strong robustness under limited data conditions. These results indicate that feature-rich input structures enable reliable and data-efficient NTGV source classification with minimal additional computational cost, supporting practical deployment in real-world distribution networks.