Utilizing Generative Adversarial Networks (GAN) and Gated Recurrent Units (GRU) for Fault Area Identification in Shipboard Zonal Distribution Power Systems: A Biotechnological Approach

Abstract

To address the challenges of limited data availability and significant interference in fault area identification within shipboard zonal distribution power systems, this study introduces an innovative approach utilizing Generative Adversarial Networks (GANs) and Gated Recurrent Unit (GRU) networks. This method enhances traditional recurrent neural networks, which are often hindered by extensive parameter requirements and slow convergence rates. By applying variational modal decomposition, the intrinsic mode functions of zero-sequence currents across different frequency bands are extracted. The GAN framework, incorporating a Layer-Normalization layer within the discriminator, alternates training between the discriminator and generator to improve the quality and effectiveness of the data generated. Subsequently, both real and synthetically augmented data sets are utilized to train the GRU network, incorporating a Batch-Normalization layer prior to the fully connected layer to expedite network convergence. The architecture thus developed is tested under various operational conditions, demonstrating a significant enhancement in fault detection accuracy and robustness against noise interference, offering a promising solution for enhancing reliability in marine power systems.