Compensation strategies for improved real-world deployment of deep learning-based partial discharge localization from digital twin data

Abstract

Measurement errors caused by real-time electromagnetic (EM) noise or the sensitivity of measuring equipment significantly affect Partial Discharge (PD) localization accuracy in open-space substations. This study proposes a Deep Neural Network (DNN) approach combined with an augmented Time Difference of Arrival (TDOA) dataset to improve PD coordinate estimation. A digital twin of a 10 m × 10 m × 2 m space was used to generate over one million synthetic data points, substantially reducing data collection time. The trained DNN demonstrated excellent localization performance in the digital environment, with more than 90% of errors below 5%, as validated by 3D scatter plots. Despite relying on augmented TDOA data from the digital twin, the DNN model exhibited high confidence when applied to real-world measurements, achieving an average localization error of 0.3610 m across 21 test points, outperforming Random Forest Regression (RFR), Gaussian Process Regression (GPR) and 1-dimensional Convolutional Neural Networks (1DCNN). Additionally, an in-depth analysis of the augmented TDOA dataset synthesis was conducted to optimize the DNN model. Key factors investigated included the impact of dataset size on localization accuracy, training time and performance across different training durations. Finally, a benchmarking comparison with existing methods was summarized in tabular form, highlighting the advantages of the proposed work over conventional iterative algorithms and other machine learning (ML) models.