Application of integrated learning in power grid fake data intrusion detection

Abstract Cybersecurity is a critical component influencing the safe and stable operation of the smart grid, as the smart grid (SG) is an information-physical fusion system (also known as a Cyber-Physical System [CPS]) integrating sensing, communication, computation, decision-making, and control built on the foundation of the traditional grid. The most common cyberattack that compromises data integrity in the smart grid is the False Date Injection Attack (FDIA). If this type of attack is not discovered in time, it can take control of physical equipment or target a network transmission line to obstruct control decisions, which can result in power network failure or even a cascade failure in the grid. An integrated learning-based detection method is currently proposed to biclassify power grid data in order to address the issues of low accuracy, high false detection rate, and poor model differentiation ability when applying a single classifier in machine learning to detect false data. The integrated learning detection method is based on GBDT (Gradient Boosting Decision Tree), XGBoost, and Light GBM, with RF- Light GBM and Bagging classifier as the base classifiers, which are integrated by voting strategy after Bayesian tuning. Following simulation experiments, the algorithm is able to significantly outperform the checking rate and accuracy of traditional detection algorithms in detecting false data on the power grid by effectively addressing the issues of low checking rate and accuracy of single classifier detection as well as instability of single classifier detection.