Ensemble CorrDet with adaptive statistics for bad data detection

Abstract

Smart grid (SG) systems are designed to leverage digital automation technologies for monitoring, control and analysis. As SG technology is implemented in increasing number of power systems, SG data becomes increasingly vulnerable to cyber-attacks. Classic analytic physics-model based bad data detection methods may not detect these attacks. Recently, physics-model and data-driven methods have been proposed to use the temporal aspect of the data to learn multivariate statistics of the SG such as mean and covariance matrices of voltages, power flows etc., and then make decisions based on fixed values of these statistics. However, as loads and generation change within a system, these statistics may change rapidly. In this study, an adaptive data-driven anomaly detection framework, Ensemble CorrDet with Adaptive Statistics (ECD-AS), is proposed to detect false data injection cyber-attacks under a constantly changing system state. ECD-AS is tested on the IEEE 118-bus system for 15 different sets of training and test datasets for a variety of current state-of-the-art bad data detection strategies. Experimental results show that the proposed ECD-AS solution outperforms the related strategies due to its unique ability to capture and account for rapidly changing statistics in SG.

Keywords

• neural nets
• power system state estimation
• covariance matrices
• statistical analysis
• smart power grids
• learning (artificial intelligence)
• power engineering computing
• ensemble corrdet
• smart grid systems
• sg technology
• power systems
• sg data
• classic analytic physics-model based bad data detection methods
• data-driven methods
• multivariate statistics
• adaptive data-driven anomaly detection framework
• false data injection cyber-attacks
• ieee 118-bus system
• bad data detection strategies