Noise pattern analysis based ground fault location approach for ungrounded electrical distribution systems.

摘要

Locating ground faults in ungrounded systems is inherently difficult because such faults do not provide considerable fault currents for tracing the fault location and the locating process has to be done during system operation. In this paper a novel fault location approach is presented which is based on the pattern recognition of high frequency noise introduced by the repetitive switching events of power electronic (PE) converters interacting with system parasitic elements (such as cable insulation capacitance and stray inductance). The proposed approach is applicable to ungrounded systems containing parasitic elements that form a ringing circuit and mechanisms to excite the ringing circuit. In addition, the faulty section and the measurement point should be high frequency coupled so that the fault location can change the ringing circuit.;The DC Zonal Electrical Distribution System (DCZEDS), which is suitable for applying the approach, is chosen as a representative example to investigate the effectiveness of the approach. The analysis is first conducted based on a notional, high-fidelity DCZEDS model in PSCAD/EMTDC. The results demonstrate the effectiveness of the approach even when severe input noise filtering is applied in the system. To investigate the feasibility of this approach for practical application, parameter sensitivity analysis is conducted based on the simulations. This technique examines the impact of varying loadings and fault impedances, varying switching frequencies, and varying operation modes, etc. To validate the analysis result based on simulations, a hardware model of a representative portion of the DCZEDS that includes shielded power cables and DC-DC converters with low switching frequencies is constructed and tested. Analysis of the experimental measurements matches that of the simulation results.;Problems with the application of this approach when using different types of converters are discussed. Recommendations for the feature extraction and feature selection algorithms are also provided.