Learning Power Systems Waveform Incipient Patterns Through Few-Shot Meta-Learning

Abstract

Incipient faults (IFs) are abnormal states before the permanent failure of power equipment. IFs are typically transient and generally do not trigger the operation of relay protection devices. This leads the difficulty in capturing IF data from waveform monitoring or recording devices. However, traditional detection methods cannot achieve satisfactory performance when faced with limited data. Besides, some signal analysis methods based on waveform conversion to images cannot obtain understandable image data and cannot analyze both current and voltage signals simultaneously. To resolve these problems, a few-shot meta-learning framework for incipient fault detection (FSMLF-IFD) is proposed in this paper. For better data processing, a waveform image conversion strategy is proposed to convert waveforms into understandable images from the time domain perspective. Then, an adaptive image fusion strategy is developed to concurrently analyze voltage and current images. Next, at the meta-training stage, an adaptability-enhancing weighting initialization strategy is constructed to address the data differences between the meta-training stage and IF detection stage. Finally, an IF detection model based on convolutional neural networks (CNNs) is obtained through the fine-tuning process. In the numerical results, the IF detection and classification accuracy of FSMLF-IFD reached 0.9720 and 0.9840 based on simulation and field IF data, which validates the effectiveness of the proposed method.

Keywords

• Incipient fault detection
• power quality
• data scarcity
• waveform image
• few-shot learning
• meta-learning