Anomaly Detection of Power Time-Series Data Based on MultiDimensional Transformer Network

Abstract

Power time-series anomaly detection has always been one of the important means to mine security threats in the power grid. The traditional detection method based on machine learning has certain limitations. It is difficult to capture the dependence between data and is prone to missed detection and false detection. In view of the above problems, this paper proposes a method for detecting abnormal data of power time-series based on Swin transformer. Firstly, the transformer model can predict the development trend of power data more effectively by using its ability to extract global information. Secondly, considering the deployment of the model in the actual scene, in order to improve the running speed and efficiency of the model, Swin transformer uses its unique patch to cut the data into small windows and reduce the sequence length. In addition, the design of the mobile window enhances the information exchange between the data and achieves the ability of global modeling. Finally, the input lightweight gradient boosting tree of the model is further improved in accuracy. The experimental results show that the proposed method can greatly reduce the calculation time under the premise of effectively identifying abnormal data.