Towards Accurate and Efficient Classification of Power System Contingencies and Cyber-Attacks Using Recurrent Neural Networks

Correct and timely responses to abnormal conditions in the power systems are crucial to their sound operation. In order for the operator or the automated response system to take prompt measures during system contingencies, it is critical to facilitate an accurate mechanism for the classification of the events and disturbances in the power grid. The massive amount of time-synchronized data recorded by the phasor measurement units can be combined with logs from other components in the power grid to create datasets for event and intrusion detection. This paper presents the results of applying deep learning techniques on open datasets recorded from a power system testbed to classify contingencies and cyber-attacks. Three different designs of recurrent neural networks (RNN) are investigated and tested for discriminating binary and multiclass events. Experiment results show 100% and 99.99% accuracy when applying the proposed classifiers on large scale binary and multiclass datasets respectively. It is also shown that one can improve the efficiency of the scheme by selectively eliminating 75% of the features in the dataset while maintaining as high as 99.96% accuracy in classifying multiclass events. Additionally, the feasibility of the design is validated by the low classification latency recorded on the low-end embedded system Jetson Nano. These promising results demonstrate the potential of employing RNN techniques in developing event and intrusion detection systems for power grids.