Abstract
Fault distance estimation in multi-bus transmission lines is challenging but essential for reducing outage time and labor costs. Conventional fault distance estimation schemes face challenges, such as the impedance-based methods necessitating comprehensive knowledge about the line parameters, and the traveling wave-based method requires a high sampling frequency for fault distance estimation. While using artificial intelligence, neither the fault resistance influences the fault location, nor is any information about the source impedance needed. Machine learning requires manual feature extraction and data preprocessing, which adds extra work. This paper employs a hybrid deep learning model based on Transformer and extreme gradient boosting (XGBoost), where the Transformer extracts the prominent feature from raw fault data and XGBoost predicts the fault distance in power transmission systems. The proposed model achieved a mean absolute error (MAE) ranging from 0.0008 to 0.6406 across different lines and an average MAE of 0.27025 km for all lines using an 80:20 train/test split, demonstrating strong performance. Ablation studies further validated the model’s robustness against noise, bad data, sensitivity to line parameter variation, load variation and different train/test splits. For validation, the data is generated from a real-time simulator, and unseen fault data is provided to a Raspberry Pi microcontroller deployed with the proposed model for fault distance estimation.

This publication has 52 references indexed in Scilit: