NeuralCompression: A machine learning approach to compress high frequency measurements in smart grid. (1st January 2020)

• Record Type: Journal Article
• Authors: Das, Laya; Garg, Dinesh; Srinivasan, Babji
• Is Part Of: Applied energy. Volume 257(2020)
• Publication Date: 2020-01-01
• Subjects: Compressive sensing -- Neural network -- Autoencoder -- Transfer learning; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2019.113966
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: An autoencoder-based approach is proposed for compression of smart grid data. Compression of public datasets demonstrate superiority for low compression ratios. Autoencoders exhibit generalizability with transfer learning. The proposed technique has low computational cost ad is easily deployable. Abstract: The smart grid features frequent communication of measurements collected at consuming and distributing nodes to other agents in the grid. While this increases grid visibility and improves situation awareness, the sheer volume of such data generated from the geographically vast grid will result in overloading of the communication infrastructure. In this regard, compressing data at the source and communicating compressed measurements has been explored in the literature. However, such techniques rely on the presence of a structure in data that is exploited for compression. In this article we propose the use of Autoencoder for data compression that extracts an appropriate structure from data that then allows for compression. The proposed approach also incorporates nonlinear transformations in the compression mechanism which is likely to improve the compression ratio for the same reconstruction accuracy. The proposed method is applied on four publicly available datasets and results show that the Autoencoder has merit over state-of-the-art Compressive Sensing for high compression ratios. Generalization of Autoencoder models to datasets from different geographical locations is also studied as a distinct feature of the proposed method. The generalizability of models is also improved with transfer learning by adapting pre-trained models to the idiosyncrasies of target dataset.