Fault Detection and Localization in an MT-VSC-HVDC Network via a Hybrid CNN-Transformer Model With a Multi-Head Attention Mechanism

In this paper, a non-unit protection approach is suggested for recognizing fault events (internal and external) and their location in a High Voltage Direct Current (HVDC) system employing a deep learning (DL) model. This work utilizes a small sample size of DC current signals from one end of an HVDC-fed system, achieving a balance between data manageability and preserving vital information related to faults. The simulation of fault samples is carried out with the maximum possible operating scenarios, such as different fault resistances, locations, types of faults, and noisy conditions. Afterward, combined data with varying noise values is trained by the proposed DL model. The testing is performed under individual noise levels to evaluate model robustness and accuracy in real-time scenarios. In the DL framework, the work focuses on two core architectures, Convolutional Neural Networks (CNN) and Transformer models, and assesses their performance individually. Furthermore, it explores the hybrid model CNN-Transformer with a multi-head attention mechanism to capitalize on their combined strengths. Afterward, a detailed comparative result evaluation with other DL models is carried out to gauge robustness and accuracy in fault detection and localization. Comparative outcomes reveal that the Transformer models performed more remarkably in varying noise conditions, while recurrent models excelled in sequential data processing. The hybrid DL methods provided a balanced trade-off, improving fault detection and localization. The study utilizes several performance metrics such as accuracy, precision, recall, F1 score, MSE, RMSE, MAPE, and R2 errors to provide a comprehensive assessment. The fault detection without noise conditions attained an accuracy of 100%; however, there was a slight decrease (1.5%) in 20db noisy conditions. The proposed approach archives excellent performance in fault location tasks, such as MAPE:0.0308, MSE:1.6415, RMSE:1.2814, and R2:0.9982. The performance of the proposed approach is also validated with varying system operating conditions and comparative performance analysis with other state-of-the-art techniques. The findings offer valuable insights into the optimal model configurations and their practical applicability for reliable HVDC system protection, contributing to more effective and real-time fault diagnosis solutions.