A Novel Unsupervised Learning Approach for False Data Injection Attack Detection in Smart Grid

Smart grids, enhanced by integrating new technologies like home energy management systems (HEMS) and smart meters, face risks like false data injection attacks (FDIA). These attacks are particularly challenging in decentralized residential demand response (DR) structures, where security breaches are not immediately evident. Traditional security measures are ineffective against FDIAs in these systems due to the diverse data sources, unique household energy use patterns, and dynamic energy forecasts. Using unsupervised learning, we present a new way to find FDIAs in smart grids. It combines K-means clustering with the spectral residual method (KM-SR). This method accurately identifies attack timeframes, as evidenced by a 95% accuracy rate, 88.67% recall, an 88 % precision-recall curve, and a 97% ROC-AUC score. Tested with Austin, Texas, energy data, KM-SR outperforms existing detection methods, showcasing its effectiveness in safeguarding smart grids against FDIA threats.