Anomaly Detection Algorithm For Smart Grid Time Series Data Based On Capsule Network

With the implementation of the State Grid Corporation’s smart grid digitization strategy,many innovative technologies such as virtualization,cloudification and artificial intelligence have been applied to the digitization of the power grid.Production and operation data,followed by real-time analysis of massive monitoring data generated by various equipment and systems,have become a new challenge for power IT operation and maintenance work.As the key technology of smart grid information operation and maintenance,anomaly detection technology can effectively detect operation and maintenance failures and give timely alarms to avoid damage to sensitive equipment.At present,some traditional anomaly detection methods detect few types of anomalies and low precision,resulting in untimely fault detection.To solve the above problems,this thesis proposes a multi-dimensional time series anomaly detection method based on capsule network.The main research contents of the paper are as follows:1.Aiming at the problem that there are few types of anomalies and the dataset lacks labeling,this paper proposes a method for labeling the grid operation and maintenance monitoring dataset.First,this paper defines abnormal types and injects abnormal data according to the actual scenarios and expert knowledge of power grid operation and maintenance;second,this paper uses four unsupervised algorithms(3-sigma,isolated forest,kmeans clustering and One-class SVM)to detect The abnormal data in the power grid operation and maintenance monitoring data set is used to mark the detection results using the idea of ensemble learning;finally,positive and negative samples are balanced using techniques of up-sampling and down-sampling.2.Aiming at the problem of low anomaly detection accuracy,this thesis proposes an anomaly detection method based on capsule network for classification and anomaly detection.Experimental results with five-fold cross-validation show that the proposed method achieves an average classification accuracy of 91.21%on a dataset containing 15 types of anomalies.At the same time,compared with the four benchmark models on the server monitoring data set containing 20,000 pieces,NNCapsNet has achieved better results on key evaluation indicators.The experimental results prove the effectiveness of the NNCapsNet algorithm.NNCapsNet can detect different types of abnormal patterns of typical business servers of power information systems,and the detection accuracy is high,avoiding physical hardware damage caused by untimely fault detection and operation caused by online business interruption.to maintain the stable and safe operation of power information equipment and systems.