A Comparative Analysis of Supervised Classifiers Employing NCA for Feature Selection to Secure Generation Control

The advancement of modern grids and its excessive dependency on information and communication networks have paved the way for cyber-attacks in the power grid. Automatic Generation Control (AGC) system is one such field in the power grid which is extensively vulnerable to cyber-attacks. Attackers may disrupt the AGC functionality and degrade system stability by falsifying the tie-power and frequency deviation data sent by remotely installed sensors. There may exist several ways of data falsification, such as by injecting step, ramp, random, pulse, replay and stealthy signals to the original healthy data. To address the issue, this paper studies and compares the performance of nine different supervised classifiers in detecting such FDI attacks in the AGC loop. As per the experimental results, Gaussian SVM, KNN, AdaBoost and Random Forest showed similar and satisfactory performance, followed by Decision Tree. However, classifiers such as Linear SVM, Polynomial SVM, Logistic Regression and Naïve Bayes' failed to effectively classify normal and compromised instances. Further, the performance comparison is conducted by exposing the test data to various levels of channel noise. It is found that, when the SNR is below 30 dB, the overall performance of the classifiers are significantly affected; however, the Precision of Positive class is less impacted by the presence of noise.