MAT-transformer-based state forecasting method for information devices

Abstract

Forecasting the state of information devices is critical to their smooth operation (e.g., smart grids). Transformer-based state forecasting methods have recently shown promising results in modeling sequential data, including time series data in the domain system of information devices. However, conventional Transformers are limited in their ability to handle very long sequences because of the quadratic memory complexity of the self-attention mechanism. This makes them less suitable for forecasting very long time series. In this study, we propose a machine-learning model based on the mean anomaly translating (MAT) Transformer that accurately forecasts the operational state of power information systems while improving the ability to forecast whether the server will fail when processing very long time series data. The proposed MAT algorithm eliminates the influence of local outliers in the time series and makes the distribution of the training data more conducive to mining by preserving its original characteristics and adjusting the forecasting model more efficiently. This improves forecasting accuracy when combined with the screening method for nonlinear correlation variables. A real-world power industry operation data set containing 170,000 information system status records was used to test the proposed method. Accurate forecasting and generation of forecasting intervals for CPU and memory metrics demonstrate the effectiveness of the method. Our proposed MAT-Transformer-based state forecasting method is a promising approach for accurately forecasting the state of information devices and has the potential to be applied in various practical scenarios.