MRNN-SA: A Multi-dimensional Time Series Fault Prediction Service for Power Equipment through Self-attention Network

Abstract

In recent years, as the business of the smart grid grows, the requirements for intelligent maintenance have become significant in the domain. One such typical application is fault prediction service for power equipment. However, traditional solutions to fault prediction have inherent limitations, because they cannot simultaneously employ patterns from global or partial segments and exclude irrelevant features from time series data. In this paper for power equipment, we propose a novel fault prediction service on multi-dimensional time series by a deep-learning model called MRNN-SA. Extensive experiments and a case study show our service can distinctly improve prediction performance on real-world sensory data from power transformers and database servers.