An Improved LSTM Method for Single-Phase Ground Fault Location in Distribution Networks

The faint characteristic signatures and strong susceptibilities to noise interference of a single-phase earth fault pose great challenges to accurate fault location. Existing traveling wave-based fault location technologies generally suffer from low wavefront recognition accuracy, weak anti-interference capability, and a reliance on manual experience. To address these issues, an improved Long Short-Term Memory (LSTM) network-based fault location method is proposed. In the proposed methodology, the traveling wave frequency feature maps and pseudo-labels are integrated. Firstly, a feature map construction method is developed using a double-layer sliding window slicing, a frequency domain transformation, and feature concatenation. Secondly, a pseudo-label is introduced to reweight the consecutive frames containing the wavefronts in order to reduce the impact of the edge frames. Finally, the exact fault distance is calculated based on the double-ended traveling wave location principle. Simulation results demonstrate that the proposed model achieves a location accuracy improvement of 89.19% and 87.93% as compared to the DNN and traditional LSTM models, respectively. Under extreme operating conditions, the location error of the proposed methodology is within ±8 m. The averaged absolute error for the measured data is 52.86 m, indicating good engineering practicality of the proposed methodology.