Detection of Faults and Attacks in Power Monitoring System via Semi-Supervised Learning with Pseudo Label Refinery

With the expansion of power grids and incorporation of new energy sources, fault and attack detection in power monitoring systems has become increasingly vital for maintaining grid stability and safety. A significant challenge for data-driven detection methods is the scarcity of labeled measurement data, which critically hampers their detection performance. This paper proposes a novel data-driven framework, SuPer, for detecting faults and attacks in power monitoring systems. SuPer employs a novel combination of self-adaptive thresholding-based pseudolabeling and prototype-based data purification strategies, which have superior detection performance under conditions of scarcity labeled data. SuPer generates reliable pseudo-labels for massive unlabeled measurements and selects the high-confidence ones to participate in the training process, which enhances the machine learning model's ability to identify anomalies accurately. Extensive experiments are conducted in a power system framework to demonstrate SuPer's effectiveness, particularly highlighting its improved performance in scenarios with limited labeled data.