Security performance, attack detection and attack identification in smart grid

Abstract

Smart grid is a new form of electricity network which integrates the traditional electrical power grid with information and communication technologies to form a bi- directional power and information flow infrastructure. It is a kind of cyber physical system (CPS), and the measurement meters are installed on the buses and the power lines connecting these buses. In this thesis, we focus on the security issues in smart grid. First, we focus on the communication network of smart grid. Encryption algorithms are used to ensure that the transmitted information in the network is secure. To increase the security level, more complex algorithms may be used, which takes longer time to process the information. To tradeoff between security level and network performance (delay), unequal security encryption (USE) is proposed, which employs different encryption schemes according to the importance level of the information. Using analytical model and simulation results, we analyze the trade-off, and find that USE achieves a good balance between the network security and performance. It just sacrifices a little on security level and has much better performance. Secondly, we focus on the integrity attack in smart grid. Data Framing Attacks (DFA) have been proposed in recent years, and have attracted much interest in the research community on smart grid security. DFA compromises Bad Data Identification and Removal (BDIR), leading BDIR to remove secure data, which will result in incorrect state estimation. Therefore, successful detection of DFA is significantly important for the control and operation of the power grid. A study on the utilization of machine learning to detect DFA is presented in the thesis. Since the detection problem can be formulated as a classification problem between secure measurements and attacked measurements, and mature machine learning technology is chosen in this work. The proposed methods are examined on the 118-bus IEEE test system. The experimental analyses indicate that machine learning can detect DFA with good performance. Finally, based on the second work, we focus on attack identification. The traditional method to do the identification is maximum residual, which treats the maximum measurement residual as the source of bad data. Since DFA misleads the BDIR, the largest residual test does not work anymore. We propose machine learning-based approaches to identify the errors. The measurements are separated into small groups. Machine learning method is applied in each group, and we can find where the error is. This method is examined on the 118-bus IEEE test system and 14-bus IEEE test system. The experimental analyses show that machine learning-based approach can perform the identification well.