Abstract:In recent years, cyber-attacks have become an imminent danger threatening the stable power supply. Data-driven approaches based on machine learning techniques are frequently used to detect cyber-attacks. To get satisfactory detection performances, these approaches have to establish an in-depth understanding of power system operating behaviors, not only under normal conditions but also under contingencies like short-circuit faults. Contingency data appears far more rarely than normal data in the historical measurement database. If normal and contingency data are uniformly sampled to generate training datasets, the included contingency examples may be too few for machine learning models to recognize the operating patterns thoroughly. In this paper, a data preparation method combining Random Matrix Theory (RMT) and Adaptive Synthetic Sampling (ADASYN) algorithm is proposed. RMT is applied to extract rare contingency data from the historical operating database and ADASYN is used to synthetically generate new contingency examples according to the existing ones. Case studies show that this method can facilitate the learning of the intrinsic statistical characteristics of power system operating data, and then enhance the detection of cyber-attacks.

A Data Preparation Method for Machine-Learning-Based Power System Cyber-Attack Detection