Machine Learning-based Cybersecurity Defence of Wide-area Monitoring Systems

Due to the vulnerability of the wide-area monitoring systems (WAMS) communication, malicious data integrity attacks on WAMS records could be initiated by adversaries which may lead to disastrous events. In response to the cybersecurity challenges raised by WAMS, recently some machine learning-based methods have been developed to authenticate the source information of WAMS measurements. Most existing source authentication methods are designed for authenticating WAMS data from a small number of locations at a large geographical scale which may not reflect the complete operating condition of WAMS in practical networks. This paper aims to examine the feasibility of using machine learning-based methods to achieve reliable source authentication of WAMS measurements for practical power grids. Four "state-of-the-art" machine learning-based approaches (including both shallow learning and deep learning methods) are examined and their performance is compared using real-life data collected from a significantly large number of locations at a small geographical scale. The simulation results demonstrate that the continuous wavelet transforms - convolution neural network (CWT-CNN) based model outperforms other algorithms due to its high identification accuracy and low computational time which has the potential to be applicable for real-time data source authentication of smart grids.