广东电力2025，Vol.38Issue(11)：34-45,12.DOI:10.3969/j.issn.1007-290X.2025.11.004
面向电厂设备异常检测的时空表征学习多元时序模型
Multivariate Time Series Model of Spatiotemporal Representation Learning Based on Anomaly Detection of Power Plant Equipment
摘要
Abstract
Anomaly detection on power plant equipment based on multivariate time series data from multiple sensors has become a research hotspot.To address the insufficient spatiotemporal feature mining and difficult root cause localization in the existing methods for anomaly detection on power plant equipment,this paper proposes a spatiotemporal uncertainty modeling oriented prediction framework integrating spatiotemporal representation learning to promote the prediction accuracy.The model captures spatial dependencies among sensors through adaptive graph learning and graph attention mechanism,while extracting multi-scale temporal dependencies using multi-scale temporal dilated attention and self-attention mechanism.It integrates spatiotemporal representations via cross-attention networks and introduces uncertainty constraints to reduce noise impact,focusing on learning the spatiotemporal features under normal operating patterns.Experiments on four public multivariate time series datasets show that the model outperforms current mainstream methods with an average 2.40%improvement in F1 score.The graph attention mechanism facilitates anomaly tracing,while multi-scale temporal feature extraction effectively captures unique patterns of different sensors.关键词
多元时序/异常检测/表征学习/注意力Key words
multivariate time series/anomaly detection/representation learning/attention分类
信息技术与安全科学引用本文复制引用
王刚,史恒惠,陈丹峰,肖楠,朱寒冰,凌贺飞..面向电厂设备异常检测的时空表征学习多元时序模型[J].广东电力,2025,38(11):34-45,12.基金项目
国家重点研发计划项目(2022YFB4100700) （2022YFB4100700）
