A novel dc fault protection scheme based on intelligent network for meshed dc grids

• Published in: International journal of electrical power & energy systems 2023-12, Vol.154, p.109423, Article 109423
• Main Authors: Yousaf, Muhammad Zain, Khalid, Saqib, Tahir, Muhammad Faizan, Tzes, Anthony, Raza, Ali
• Format: Article
• Language: English
• Subjects: Artificial neural networks (ANNs); Bayesian optimization (BO); Dc line protection; Modular multilevel converter (MMC); MT-HVdc system
• ISSN: 0142-0615
• DOI: 10.1016/j.ijepes.2023.109423

•For dc fault diagnosis in MMC-MT-HVdc grids, improved features are collected using DWT-based MRA and Parseval's theorem.•An optimal ANN architecture is used to classify fault patterns based on transient features. The proposed protection scheme is robust against severe noise conditions and high impedance faults. Bayesian optimization is a well-suited technique to optimize ANN hyperparameters.•The Bayesian optimization method also significantly reduces the time required to evaluate and select ANN models.•It is further finding an efficient autonomous solution to tune excessive multiple hyperparameters with the help of Bayesian optimization and providing a basic understanding of it.•This algorithm is implemented in a single-end main and coordinated secondary unit containing control logic. By detecting internal dc faults as soon as possible, the proposed scheme covers the failure of the main unit with expedited backup. This paper proposes a fault detection and classification scheme for multi-terminal high voltage direct current (MT-HVdc) lines by integrating discrete wavelet transform (DWT) multi-resolution analysis with artificial neural networks (ANNs). Previously, such intelligent protection schemes used manual approaches or arbitrary rules of thumb to optimize a set of hyperparameters of the neural networks without applying any optimization algorithm. In order to improve accuracy, this work proposes an efficient Bayesian Optimization (BO) approach for evaluating and establishing the regulated hyperparameters for ANNs. The DWT multi-resolution analysis (MRA) and Parseval’s theorem are used to extract energy variation for various faults. The energy variation of fault signals at different scales is fed into a multi-stage model to optimize the hyperparameters of neural networks with minimal training setup time and compute effort. After training, the data-based algorithm is implemented in a single-end main and coordinated secondary unit with control logic. The proposed scheme intends to detect internal short-circuit dc faults as quickly as possible and cover the failure of the main unit with expedited backup action. The findings of the study reveal that the proposed scheme can accurately detect internal faults in a variety of testing conditions and remain stable against external faults or disturbances with an average recognition accuracy of 99.38%.