Fault Detection and Location in DC Microgrid Using Autoencoder Neural Network

Abstract

The increasing number of DC microgrids added to contemporary power systems poses great challenges in fault protection and in the location of faults for DC microgrids, whose configurations are preferably ungrounded because the latter removes the need for additional bank configuration. In this thesis, a novel artificial neural network architecture is proposed to perform fault location in ungrounded DC microgrids using selective voltage measurements. The study covers the fundamental requirements for fault location in ungrounded systems and is based on how pole-to-ground voltages act in case of short circuits. The proposed methodology suggests a complex and sophisticated neural network architecture finely tuned for voltage-based fault detection and location. The architecture relies on measurements from four strategically positioned nodes in the DC microgrid infrastructure and provides a first-of-its-kind fault identification technique based on the principle that the measured voltage at the fault will be equal to zero. In the proposed neural network architecture, several dense layers are sequentially added, followed by applying batch normalization and ReLU activation functions to capture complex nonlinear mappings of fault characteristics. It particularizes the utilization of dedicated voltage normalization schemes and flexible thresholds to improve fault detection and discrimination from normal operating conditions. This work proposes a unified mathematical approach for the analysis and implementation of voltage-based fault location methods. The proposed system consists of multi-stage processing of data, from voltage normalization followed by feature extraction, fault detection, and fault localization. Robust performance across a wide range of operating conditions and fault scenarios is ensured through advanced optimization techniques and training methodologies. The experimental work was performed in an advanced test environment integrated with Hardware-in-the-Loop simulation and a laboratory-scale DC microgrid at 250V DC. It allowed examining the proposed technique under different faults, like fault resistance from 0.1Ω reaching 100Ω, single and double faults of transmission lines, and different operating conditions. Incorporating real-time digital simulation capabilities for the time step of 50 microseconds, the validation framework ensured accurate representation of fast transient phenomena during fault events. Experimental results show outstanding performance indexes that considerably push the limit of the modern technologies in DC microgrid protections. With a mean response time of 2.3 milliseconds, the system established 99.3% detection accuracy, as well as location accuracy above 97.7% for fault resistances up to 50Ω. An evaluation of the performance shows its strong performance within the proposed operating conditions with false positive and negative rates under 0.2%. It proves the practical viability of the proposed system in advancing industrial applications due to its high accuracy under normal, high impedance, and multi-fault scenario conditions. This study greatly advances theoretical knowledge and practical application for DC microgrid protection systems. The proposed mathematical framework opens new horizons of understanding fault location principles in a setting where voltages and associated strain are used for fault location, and a practical implementation demonstrates the applicability of neural network protection schemes.