Small-Sample Event Identification Based on Adaptive Second-Order MDF and Triplet CNNs Using Distribution-Level Synchronized Measurements

Abstract

Distribution-level synchronized measurements serve as a valuable resource for event-type identification. The precision in identifying events offers crucial information for fault analysis, asset monitoring, state estimation, and ensuring the reliability of power delivery. Currently, a major issue in the event identification task arises from the scarcity and imbalance of event samples, attributed to the brief installation period of the distribution-level synchrophasor measurement system. To address this, a novel few-shot learning event identification method is proposed in this paper. The Adaptive Second-order Motif Difference Field (A2-MDF) is developed for feature transformation, which allows extracting higher-order patterns and structural information from time series data into images to extract fused features at different time scales. A dual strategy involving sample enhancement through Data Augmented Generative Adversarial Networks (DA-GAN) and sample balancing via Random Under Sampling (RUS) is also developed to mitigate the issues associated with sample scarcity and imbalance across different event types. Finally, a triplet network employing three Convolutional Neural Networks (CNNs) is proposed for classification in small-sample scenarios. Results using both simulation and field data demonstrate the advantages of the proposed method with limited labelled data.