Applications for Autoencoders in Power Systems

This paper explores the use of autoencoders in enhancing the management and security of power systems applications. Research indicates that auto encoders are effective in dimensionality reduction capabilities, feature extraction, han-dling noisy data and non-linear feature sets, and bottleneck layer representations (i.e., input data in reduced form). This study examines six widely-used autoencoder models and their distinctions that show their adaptability to different types of data (i.e., time series, numerical, tabular, text, and image) and operational demands (i.e., cyber, physical) typical of power systems. The paper identifies autoencoder applications in: (1) attack/anomaly detection; (2) fault identification; (3) data im-putation; (4) dimensionality reduction; (5) state estimation; and (6) forecasting, but has detailed these applications specifically for the anomaly detection and fault identification types. Also, applicable selections of auto encoder types describing do's and don'ts is presented.