Learning approach based DC arc fault location classification in DC microgrids

Abstract

A very challenging problem in direct current (DC) microgrid (μG) systems are the arc faults due to small air gaps caused by loose connections or degraded insulation. Generally, the voltage difference across these gaps initiates a plasma arcing channel with extremely high temperature. If not detected and extinguished in time, arc faults could endanger adjacent circuits and eventually cause fire hazards. Hence, to minimize the impact of DC arc faults through timely detection, this work proposes a novel detection scheme to improve the detection accuracy and to reduce unwanted false tripping. This is achieved by investigating the physical characteristics of the DC arc and by analyzing the arc current signals. The voltage-current characteristics of the DC arc as well as the high frequency arc noise are modeled and analyzed. This model enables an accurate simulation of the DC arc and its impact to a larger system is analyzed, i.e., a DC μG with multiple voltage sources and multiple resistive loads, and interactions between DC arc faults and two typical μG control strategies are carried out. A new detection scheme based on arc signatures from both time domain and time-frequency domain are proposed and verified through multi-layer perceptron neural networks (MLPNN) in numerical simulations. This strategy increases the detection accuracy and reduces the possibility of false trip. The training process of the proposed approach showed 97% accuracy and the testing process indicated 93.3% accuracy for a large sample of data.