Traveling wave-based fault location in power grids using neural networks

Abstract

downtime and improving public safety. One practical FL approach involves utilizing traveling waves (TWs) to locate the fault along a transmission line. TW-based FL methods are highly regarded for their speed and resilience to variations in fault parameters. However, many existing TW-based FL methods rely on constant propagation velocity assumptions, which can be challenging to apply in real-world applications. In response to these limitations, this thesis introduces two approaches integrating machine learning to locate faults in power grids and on hybrid lines accurately. First, a novel wide-area TW-based FL method is presented to handle complex grid networks, including multiple line type configurations and bus impedance variations. TW arrival times are collected via TravelingWave Recorders (TWRs) strategically placed across the network, capturing the FL information. A multi-task multi-layer perceptron (MLP) is then trained using TW arrival times throughout the network to simultaneously classify the faulted line and determine the exact FL within that line. Training the MLP avoids the need for explicit propagation velocity assumptions. This approach remains accurate with mixed transmission line types with varying line parameters and remains robust, even if the TW arrival times provided by the TWRs are slightly inaccurate. Furthermore, this method accounts for bus impedance effects, which affect the attenuation of the TWs, creating a more realistic scenario. Following the proposed wide-area method, the thesis addresses the unique challenges of hybrid lines (a series of non-homogeneous transmission lines). Conventional time-domain TW-based methods often face complications when the TW encounters varying propagation speeds, reflections at junctions, and distinct attenuation characteristics across nonhomogeneous line sections. To overcome these obstacles, the proposed FL technique uses the frequency-domain properties of TWs. Given that TWs are inherently tied to both distance and frequency, the proposed method captures the initial TW wavefronts at both line terminals, applies Clarke’s Transform and the Fast Fourier Transform (FFT) to extract critical frequency content (FC) from the recorded signals. The FC is then processed by neural networks (NNs), which identify the faulty section of the line and determine the precise FL. Utilizing the frequency domain removes time synchronization requirements and the estimation of propagation velocity. PSCAD/EMTDC simulations are conducted to validate both proposed methods. The IEEE 39-bus system is used to validate the robustness of the wide-area FL method by including different line configurations and added bus impedances to mimic measurement equipment. The method’s ability to handle inaccurate arrival time data and various propagation velocities will be examined. Simulations will also confirm the accuracy and reliability of the proposed frequency-domain approach using a hybrid line from Mainland British Columbia to Vancouver Island. Various fault conditions, such as fault resistances and inception angles, will be tested. The impact of noise and sampling frequency will also be examined.