Anomalous load profile detection in power systems using wavelet transform and robust regression. (August 2022)

• Record Type: Journal Article
• Authors: Karkhaneh, Mohammad; Ozgoli, Sadjaad
• Is Part Of: Advanced engineering informatics. Volume 53(2022)
• Publication Date: 2022-08
• Subjects: Anomaly detection -- Wavelet transform -- Robust regression -- Data cleansing; Computer-aided engineering -- Periodicals; Engineering -- Data processing -- Periodicals; 620.00285
• DOI: 10.1016/j.aei.2022.101639
• Languages: English
• ISSNs: 1474-0346

Abstract

Abstract: In modern power systems, a clean database of historical electric loads is crucial for system analysis, planning studies, load forecasting, and decision making. To this end, the clean database should be free of anomalies while preserving the power consumption behavior of special events like holidays. In recent years, model-based anomaly detection methods have been successfully used for anomaly detection and correction in electrical load. In these methods, the collected load data are first compared with the predicted load obtained from a forecasting model; if the difference is greater than a specific limit, the collected load data are replaced with the predicted value. One of the disadvantages of these methods is the consideration of load profiles of the days with higher levels of difficulty in load forecasting, such as holidays, as anomalous load profiles and replacing them with predicted values, which results in fading some important characteristics of the load series. Therefore in this paper, a two-stage anomaly detection method is proposed to alleviate this problem. In the first stage, the wavelet transform is used to identify suspected load profiles. In the second stage, a robust regression method is used to further inspect the suspected load profiles and decide whether they are anomalies or not. By performing experiments on three datasets (a utility of Iran, New York City, and a North American utility) and comparing the results with four other methods, it's shown that the proposed method decreases false detection of anomalies while accurately correcting anomalous load profiles. Results of load forecasting with the Iranian dataset corrected by the proposed method showed an average of 2.28% and 1.58% error on the holidays and total days, respectively, which is less than the other four methods.