A hybrid framework for the cyber resilience enhancement of frequency control in smart grids

Abstract

Modern power systems undergo a continuous digitalization for a more reliable, secure, and environmentally friendly operation. However, this advancement opens a door to a wide range of digital threats, making electrical grids vulnerable to cyberattacks. These malicious activities mainly affect the monitoring and control systems of smart power infrastructures. One of the most fundamental automation of power systems is the Load Frequency Control (LFC), which is responsible for maintaining the energy equilibrium in an electrical system by remotely adjusting the setpoints of the regulated generators. The criticality of LFC makes it a prime target for adversaries. Inspired by this threat, the present thesis introduces a novel set of active protection layers that detect, locate, estimate and mitigate the impact of cyberattacks against LFC. For each layer, both a model-based and a data-driven approach is designed, formulating a hybrid framework that increases the cyber resilience of LFC. The criteria for selecting the proper methodology at each layer are established according to the specifications of the system. The model-based layers of the proposed hybrid framework are based on a special type of mathematical systems known as state observers. The related detection and localization methodologies use novel pairs of sliding mode (SMOs) and Luenberger observers to identify cyberattacks against LFC. The main benefit of these methodologies is their ability to distinguish cyberattacks from other types of external disturbances. Regarding the attack detection thresholds, an adaptive design has been selected to minimize false positive alarms. After determining which LFC signals have been corrupted, the introduced attack estimation technique takes place. This method approximates the characteristics of the identified cyberattacks by utilizing an innovative combination of SMO and unknown input observers. The estimated attacks are then fed to the proposed attack-resilient control to neutralize the effects of malicious activities against the considered system. The developed observer-based estimation and mitigation approaches employ an H∞ method to minimize the effects of external disturbances on their performance. The data-driven techniques of the introduced hybrid framework apply advanced deep learning algorithms to strengthen the cyber resilience of LFC. For the corresponding detection and localization methodologies, an autoencoder is trained on time-series that represent various normal LFC states. After the training process, the model can replicate a given input with high accuracy under normal operation while it fails to achieve the same goal during a cyberattack. This feature makes the autoencoder a proper indicator for cyberattacks. Next, a deep neural network (DNN) is utilized for the proposed data-driven estimation and mitigation approaches. The DNN is trained on data that reflect the normal operation of LFC to estimate the healthy control signals through selected field measurements. The trained DNN is then deployed in the control center, along with backup communications channels that transfer the sensor readings and the approximated setpoints. When an attack is detected in the system, the original control loop is temporarily discarded and replaced by the proposed DNN, allowing the uninterrupted operation of LFC even under cyberattacks. For the performance assessment of the designed cyber defense layers, a series of detailed experiments is conducted. Firstly, the effectiveness and the scalability of the proposed methodologies are tested on several use cases of growing complexity. In the LFC modeling of these use cases, several practical features have been considered, such as nonlinearities, high-voltage direct current (HVDC), thyristor controlled phase shifter-equipped (TCPS) tie-lines, disturbances due to Renewable Energy Sources (RES), etc., to emulate the operation of real-world power systems. The performance of the introduced methodologies in realistic conditions is further investigated through software/hardware-in-the-loop techniques. Next, the robustness of the presented approaches against various system uncertainties, such as system parameter miscalculations, noisy settings, time delays, etc., is numerically evaluated. Finally, the introduced cyber defense layers are compared with other, state-of-the-art works of the research field to highlight the contribution and the innovations of the present thesis.