On-Line Fault Identification, Location, and Seamless Service Restoration using Transfer Learning-Based Convolution Neural Network for Low-Voltage DC Microgrid. Issue 8 (9th May 2023)

• Record Type: Journal Article
• Authors: Veerapandiyan, Shanmugapriya; Sugavanam, Vidyasagar
• Is Part Of: Electric power components and systems. Volume 51:Issue 8(2023)
• Page Start: 785
• Page End: 808
• Publication Date: 2023-05-09
• Subjects: DC microgrid -- fault diagnosis -- transfer learning -- convolution neural network -- deep learning; Electric machinery -- Periodicals; 621.3104205
• DOI: 10.1080/15325008.2023.2183997
• Languages: English
• ISSNs: 1532-5008

Abstract

Abstract: DC microgrid over the last decade has gained a global paradigm in the power system field. Through the effective integration of distributed energy resources, significant researchers have improved its advantages over conventional power systems. The new state-of-the-art infrastructure despite its numerous advantages possess challenges in implementing an appropriate protection system. Impact of selecting a definite threshold for voltage and current compromises with the accuracy and speed of detection in conventional fault detection methods. Although many machine learning methods are successful in fault detection and classification for DC microgrid they still suffer from overfitting problems and exhaustively time-consuming. This article intends to provide an Online fault protection method for a LVDC microgrid system based on a transfer learning-based convolution neural network (TCNN). With the help of transient voltages and currents at different buses, image data for faults at different buses serves as input to the convolution neural network layer. First, the pre-trained Alex-Net CNN initializes the weights and biases for the targeted offline CNN's. Secondly, the transferred layers from the offline CNN's, initializes the online convolution neural network for real-time fault detection and classification. This work aims to accurately identify and locate the faults without complex dataset and multiple thresholds while improving accuracy of fault detection and classification. To ensure reliability of the system the recognized faulty bus reconnects to the healthy bus via sectionalizing circuit breakers through the detected signals. The proposed TCNN framework has an accuracy of 99.78%. The proposed method results when compared with state-of-the-art machine learning techniques such as SVM, LSTM, RNN, multilayer perceptron, and wavelet-based ANN showed better results in terms of accuracy and has significantly reduced data abundance.