Implementing AI Solutions for Advanced Cyber-Attack Detection in Smart Grid

Abstract

As the backbone of modern power systems, the smart grid is one of the most critical applications of the Internet of Things. The intelligent electricity system faces various types of unauthorized malicious access, that is, cyber-attacks. With the development and application of information and communication technologies in traditional power systems, the improvement of physical-cyber systems in the smart grid also increases. Nowadays, the deployment of defensive measures has surged in response to the growing number of attacks aimed at the smart grid. Therefore, in this paper, we investigate cyber-attack identification using artificial intelligence-based models and identify them by estimating the state vector of the electricity network. We accomplish simulations by extracting data from the five-bus IEEE network to test the effectiveness of the proposed algorithm. False data attack vectors are then injected into the healthy measurements. In this way, to check the detection power of the proposed algorithms, 2,000 measurement samples have been taken from the five-bus network, half of which are considered healthy data and the other half as manipulated data. After labeling healthy and false data, machine-learning algorithms such as decision trees and k-nearest neighbor (KNN) have been used to investigate and identify this type of attack. Comparative analysis of the two proposed algorithms against commonly used methods demonstrates significantly improved accuracy. Specifically, according to the best depth for the decision-tree algorithm and k for the KNN algorithm, it is drawn with k = 3 and the decision-tree algorithm with a depth of 9. According to the algorithms proposed in this article, the decision-tree algorithm with a depth of 9 in p-value of 0.45 and 0.64 and the nearest neighbor algorithm with k = 3 in p is equal to 0.72, 0.98, and 1 represent better accuracy. Also, the results indicate that the two proposed algorithms have performed much more favorably than other classification methods. Additionally, the detection accuracy increases with higher p-values for these two algorithms. This problem shows that the detectors can detect false data injection attacks that cause more severe disturbances in the system.