Abstract
With the increasing use of microgrids, fault identification remains a significant challenge for microgrid protection. Overcurrent protection is the most widely used type of protection in the grids. Also, deep neural networks (DNNs) are used as suitable solutions to improve fault classification accuracy. However, some parameters can affect the performance of the deep learning-based fault detection and location (FDL) schemes. This paper presents a protective agent (PA)-based fault classification method using intelligent electronic devices (IEDs) and hybrid DNNs along with various parameters' impact investigation. First, an FDL scheme based on two types of DNNs (single and hybrid types of layers) and PAs will be presented. Then, we analyze various overcurrent-based protection scenarios and their different parameters' impact on the proposed scheme. In the next stage, the impact of different parameters including adding laterals to the microgrid, adding IEDs to the protection plan, improving the structure of the DNNs, protection scheme structure, and data transfer type on the proposed scheme are investigated in a data analysis and a comprehensive sensitivity analysis. The results demonstrate that the lowest and highest accuracy of the proposed algorithm in the studied scenarios in the fault detection section is 95.54% and 99.96%, in the fault type and phase detection section, is 95.56% and 99.86%, and for the fault location error is 6.51% and 1.27%, respectively. In all scenarios, due to the use of hybrid DNNs with two types of layers, the accuracy of DNN-2 was higher than DNN-1. Also, the results in the sensitivity analysis show that despite the high accuracy of the proposed scheme, some parameters significantly impact the accuracy of the DNNs outputs. We performed simulations using three software: DIgSILENT Power Factory, MATLAB, and Python (with Tensorflow and Keras libraries).