Few-Shot Metering Anomaly Diagnosis with Variable Relation Mining

Abstract

Metering anomalies not only mean huge economic losses but also indicate the faults of equipment and power lines, especially within the substation. As a result, metering anomaly diagnosis is becoming one of the most important missions in smart grids. However, due to the insufficient and imbalanced anomaly cases, identifying the anomalies in smart meter data accurately and efficiently remains challenging. Existing methods usually employ few-shot learning models in computer vision directly, which requires the rich experience of human experts and sufficient abnormal cases for training. It blocks model generalizing to various application scenarios. To address these shortcomings, we propose a novel framework for metering anomaly diagnosis based on few-shot learning, named FSMAD. Firstly, we design a fault data injection model to emulate anomalies, so that no abnormal samples are required in the training phase. Secondly, we provide a learnable variable transformation to reveal inherent relationships among various smart meter data and help FSMAD extract more efficient features. Finally, the deeper metric network is equipped to support FSMAD in obtaining powerful comparison capability. Extensive experiments on a real-world dataset demonstrate the advantages of our FSMAD over state-of-the-art methods.