A Semi-Supervised Clustering-Based Approach for Sensing and Classifying Unknown Attacks in Power Systems

To solve the problem of perceiving and classifying unknown attacks in power systems, the study proposes a method for perceiving and classifying unknown attacks in power systems based on semi-supervised clustering. The method first uses a hybrid algorithm of improved genetic algorithm (GA) and improved multi-grain cascade forest (GC-forest) power network system unknown attack for perception and detection, then increases the number of training data by marking extension algorithm, provides specific information of detected attack or event using the idea of semi-supervised clustering, and finally filters the clustered unknown attack or event using similarity metric method . The analysis of the performance results testing showed that the perceptual detection of unknown attacks by the improved GC-Forest and the improved Genetic Algorithm (GA) had a check-all rate, precision and accuracy of 97.1% and 96.9%, 98.9% and 98.6%, 99.0% and 98.7%, respectively, and the average computing time was 3.12s and 3.68, respectively. for the classification of unknown attack types, the semi-supervised clustering-based (SSL) classification method has 100%, 95% and 95% classification accuracy for normal, non-power network attacks and power network attacks, respectively, while the adaptive classification method has 100%, 92%-98% and 90%-95%, respectively. Taken together, these data illustrate that the research can efficiently and accurately sense and classify unknown attacks on power systems and provide some reference value for fault sensing and classification in other areas.