Univariate Time Series Anomaly Detection Based on Hierarchical Attention Network  

作　　者：Zexi Chen Dongqiang Jia Yushu Sun Lin Yang Wenjie Jin Ruoxi Liu 

机构地区：[1]State Grid Beijing Electric Power Company,Beijing 100031,China [2]Institute of Electrical Engineering,Chinese Academy of Sciences,Beijing 100190,China

出　　处：《Tsinghua Science and Technology》2024年第4期1181-1193,共13页清华大学学报自然科学版（英文版）

基　　金：supported by the Science and Technology Project named“Research on Risk Perception and Defense System for Medium and Low Voltage Distribution System Operation Based on Data Mining”of State Grid Beijing Electric Power Company(No.520202220002).

摘　　要：In order to support the perception and defense of the operation risk of the medium and low voltage distribution system, it is crucial to conduct data mining on the time series generated by the system to learn anomalous patterns, and carry out accurate and timely anomaly detection for timely discovery of anomalous conditions and early alerting. And edge computing has been widely used in the processing of Internet of Things (IoT) data. The key challenge of univariate time series anomaly detection is how to model complex nonlinear time dependence. However, most of the previous works only model the short-term time dependence, without considering the periodic long-term time dependence. Therefore, we propose a new Hierarchical Attention Network (HAN), which introduces seven day-level attention networks to capture fine-grained short-term time dependence, and uses a week-level attention network to model the periodic long-term time dependence. Then we combine the day-level feature learned by day-level attention network and week-level feature learned by week-level attention network to obtain the high-level time feature, according to which we can calculate the anomaly probability and further detect the anomaly. Extensive experiments on a public anomaly detection dataset, and deployment in a real-world medium and low voltage distribution system show the superiority of our proposed framework over state-of-the-arts.

关 键 词：edge computing anomaly detection univariate time series self-attention 

分 类 号：TP393.08[自动化与计算机技术—计算机应用技术]