Unsupervised Optimal Anomaly Detection Model Selection in Power Data

• Published in: 2022 China Automation Congress (CAC) pp. 5661 - 5666
• Main Authors: Yu, Guangrong, Yang, Qinsheng, Zhu, Yongjin, Zhang, Shiwei, Wu, Baotai, Liu, Shangdong, Ji, Yimu
• Format: Conference Proceeding
• Language: English
• Published: IEEE 25.11.2022
• Subjects: anomaly detection; Automation; Big Data; Computational modeling; Data models; Feature extraction; identification accuracy; power data; Training data
• ISSN: 2688-0938
• DOI: 10.1109/CAC57257.2022.10054730

Power data is complex and diverse. Different data types correspond to different power anomaly monitoring models. How to use a variety of feature combinations to automatically screen the optimal power anomaly detection model in the scenario of unsupervised power data anomaly detection is an urgent problem to be solved. First, extract the complex power data features into seven types of eigenvalues. Then, using the selection algorithm for unsupervised anomaly detection models based on the METAOD method, the optimal selection results of anomaly detection models under various power data sets are used to generate a selection database. Finally, divide the seven types of features into different combinations and use the reward principle and the corresponding abnormal detection results to combine and screen the optimal feature combination and the optimal power abnormality monitoring model for the existing data.