Mat-Transformer-Based State Prediction Method for Information Equipment
Min Wang,chen zheng,Kun Wang,zhigang chen
DOI: https://doi.org/10.2139/ssrn.4055751
2022-01-01
SSRN Electronic Journal
Abstract:The smart grid is a major technological innovation that has provided a focused development trend for 21st century power systems.The adequate monitoring of smart grid information systems is crucial as personnel are required to control real-time operations, monitor status, and maintain all equipment so that effective and timely operational decisions can be made. Hence, accurate and actionable prediction results are vital to alarm responses. However, seasonal, climatological, operational, and maintenance behaviors often cause non-stationary and periodic interference with monitoring methods that rely on multidimensional time-series data. Therefore, accurate and efficient prediction of the operational state of information systems remains a vital challenge. Presently, short-term prediction methods based on small-scale datasets have become available. However, as power operations and maintenance data grow larger and more complex over time, the application of these methods to large-scale datasets will inevitably lead to low prediction accuracy while slowing computational speeds. To help counter this inevitability, this study proposes a mean anomaly translating (MAT) transformer-based machinelearning model that predicts the operational state of power information systems while improving generalizability when processing with massive amounts of data. The proposed MAT algorithm eliminates the influence of local outliers in the time series, rendering the training data distribution more conducive to mining by retaining its original characteristics while more efficiently accommodating the prediction model. This improves prediction accuracy when combined with the nonlinear correlation variable screening method. A real-world power industry dataset containing 170,000 information system status records was used to test the proposed method, and accurate predictions and generation of prediction intervals for CPU and memory metrics demonstrate the effectiveness of the method.