Unsupervised concrete feature selection based on mutual information for diagnosing faults and cyber-attacks in power systems

Removing the redundant features from massive data collected from power systems is of paramount importance in improving the efficiency of data-driven diagnostic systems. This work proposes a novel concrete feature selection based on mutual information, called CFMI, for selecting proper features to enhance diagnosing faults and cyber-attacks in power systems. The proposed technique is then compared with various state-of-the-art techniques and a comprehensive study has been performed on the selected features. All techniques are evaluated with respect to simulated scenarios on IEEE 39-bus system and a Three-Bus Two-Line experimental setup. The attained results, on one hand, verify the superiority of the proposed CFMI technique over other techniques. On the other hand, the selected features from both setups denote that current and voltage features are more informative than other features for diagnostic systems. Furthermore, the results of the comprehensive study shows that those features collected from generation buses are of higher priority for diagnostic systems.