Deep learning based detection of attacks against state estimation in smart grids

Electricity grids, one of the critical infrastructures, have become vulnerable to malicious attackers with the development of communication technologies. For this reason, power systems are among the issues that need to be examined in terms of cyber security. The safety of the state estimator using the measurements obtained from the electrical network is crucial. Due to the nature of the sensors in a power system, errors and communication problems may occur. The state estimator may fill in erroneous measurements due to sensor and communication errors by correcting minor errors in the measurements, but an attacker who has leaked the system's communication network could inject false data into the measurements and bypasses traditional detection algorithms and causes errors in the power system state variables. In this thesis, a deep learning-based detection algorithm is examined for the detection of the attacks that traditional residue-based approaches cannot detect. In addition, the False Data Injection (FDI) attack, which was previously proposed in the literature, is mathematically modeled in this study. FDI attack was simulated using open source MATGRID library in MATLAB by using IEEE 9 Bus test system. As a result of the study, the attack vector obtained mathematically and state variables were corrupted and the deep learning-based model performed well in detecting attacked measures.