A data-driven algorithm to detect false data injections targeting both frequency regulation and market operation in power systems

• Published in: International journal of electrical power & energy systems Vol. 143; p. 108409
• Main Authors: Deb Roy, Siddhartha, Debbarma, Sanjoy, Guerrero, Josep M.
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.12.2022
• Subjects: AGC; Ensemble; FDI; Feature creation; Feature engineering; LFC; Machine learning; Market operation attacks; One class classifier; Power systems; SVDD; Feature creation; SVDD; Market operation attacks; Ensemble; Machine learning; LFC; FDI; Power systems; AGC; Feature engineering; One class classifier
• ISSN: 0142-0615
• DOI: 10.1016/j.ijepes.2022.108409

•A novel machine learning approach is proposed for detecting anomalies in AGC loop.•The proposed approach combines K-means, SMOTE, and multiple SVDD models.•For training, the proposed approach requires only one class data.•Feature Engineering concept is used to devise a new feature.•A financially oriented market operation attack is presented that earn illegal profits. This paper focuses on detecting cyber-attacks targeting the Automatic Generation Control (AGC) loop and market operation. To achieve this, a new data-driven learning algorithm is proposed that ensembles various learning tools such as K-Means clustering, Synthetic Minority Oversampling Technique oversampling, and Support Vector Data Description models as the base learners. Next, this paper devises a new feature variable, EF, which generates a relatively higher value for attack scenarios than normal grid states, thus aiding the classifier in predictions with low false alarms. The proposed approach can detect a wide range of cyber-attacks, including unseen attack cases. This paper further modelled and validated the detection of profit-oriented intelligent market operation attacks, absent in most of the previous works. Such attacks project themselves within the proximity of nominal grid states, making them difficult to predict. The algorithm's performance is finally compared with other learning models, and it is found that the proposed technique has superior prediction with True Positive Rate, True Negative Rate, and Geometric Accuracy as 98.13%, 99.85%, and 98.98%, respectively.