The security of smart grid is a major challenge in grid modernization. While many existing solutions rely on human-defined features to develop machine learning (ML) based attack detectors against prominent exploits, such features are becoming more expensive and less effective in the smart grid. To supplement more high-quality features for ML-based threat monitoring, this paper proposed a stacked autoencoder (SAE) based deep learning framework to develop machine-learned features against transmission SCADA attacks. Compared with the state-of-the-art ML detectors, the proposed framework leverages the automaticity of unsupervised feature learning to reduce the reliance on system models and human expertise in complex security scenarios. Simulations with data collected from a high-fidelity smart grid testbed demonstrated that the machine-learned features effectively enabled more accurate discrimination against SCADA exploits in power transmission systems.
David Wilson, Yufei Tang, Jun Yan, Zhuo Lu, In Proc. IEEE PES General Meeting 2018, Portland, Oregon, August 2018
KEYWORDS: feature extraction, smart grids, relays, monitoring, training, phasor measurement units, security, SCADA systems, power system security, grid modernization