盛振明, 郭耀松, 刘超,等.基于深度学习的电网指标数据异常检测方法研究[J].河南理工大学学报（自然科学版）,2025,44(4):59-65.

SHENG Z M, GUO Y S, LIU C, et al. Anomaly detection method for power grid indicator data based on deep learning [J]. Journal of Henan Polytechnic University (Natural Science) , 2025, 44(4): 59-65.

基于深度学习的电网指标数据异常检测方法研究

盛振明1, 郭耀松1, 刘超1, 成阳1, 韩肖1, 方巍2

1.国电南瑞南京控制系统有限公司，江苏 南京  211106；2.南京信息工程大学 计算机学院，江苏 南京  210044

摘要: 目的 基于深度学习的时间序列异常检测在电力系统运维、电力设备故障检测、电网运行故障监测等智能运维场景中起关键作用。尽管现有方法取得了显著效果，但它们主要关注于特定窗口的时间序列，忽视了不同特征维度时间序列之间存在的相关性，在某些情况下会导致误检，异常检测精度下降。 方法 本文提出一种基于时间-特征维度融合的异常变压器（temporal-feature  fusion  anomaly  transformer，  TFFAT）模型，用于无监督多变量电网指标数据异常检测。TFFAT利用图注意力机制并行地从时间维度和特征维度学习多变量时间序列之间的复杂依赖关系，采用异常Transformer模型处理融合的隐藏特征并生成异常得分。 结果 结果表明，在3个公开的时间序列异常检测数据集上，TFFAT检测精度分别达到89.73%，92.12%，97.14%，显著优于现有基准方法。 结论 TFFAT能够充分捕获时间和特征维度的时间序列之间的相关性，从而精确捕获时序数据的异常关系，在电网运维中具有重要的应用价值，能够显著提高电网故障检测的准确性，减少误检，增强电网的稳定性和可靠性。 

关键词:深度学习;电网数据;异常检测;电网运维;电网监控;时间序列

doi: 10.16186/j.cnki.1673-9787.2024070038

基金项目:国家自然科学基金资助项目（42475149）；2023年国电南瑞南京控制系统有限公司科技信息项目（524609230052）

收稿日期:2024/07/09

修回日期:2024/10/08

出版日期:2025/06/19

Anomaly detection method for power grid indicator data based on deep learning

Sheng Zhenming1, Guo Yaosong1, Liu Chao1, Cheng Yang1, Han Xiao1, Fang Wei2

1.NARI-TECH Nanjing Control Systems Co.， Ltd.， Nanjing  211106， Jiangsu， China；2.School of Computer Science， Nanjing University of  Information Science & Technology， Nanjing  210044， Jiangsu， China

Abstract: Objectives Time series anomaly detection based on deep learning has plays a key role in intelligent operation and maintenance scenarios such as power system operations， equipment fault detection， and power grid fault monitoring. While existing methods have achieved notable success， they often focus on fixed time windows and overlook correlations between different feature dimensions of the time series， which can lead to false positives and reduced detection accuracy. Methods This paper proposes a Temporal-Feature Fusion Anomaly Transformer （TFFAT） model for unsupervised anomaly detection in multivariate power grid indicator data. TFFAT leverages a graph attention mechanism to capture complex dependencies from both the temporal and feature dimensions in parallel. It employs an anomaly transformer to process the fused hidden features and compute anomaly scores. Results Experimental results on three publicly available time series anomaly detection datasets show that TFFAT achieves detection accuracies of 89.73%， 92.12%，  and 97.14%， respectively， significantly outperforming existing benchmark methods. Conclusions TFFAT effectively captures interdependencies across temporal and feature dimensions， enabling more accurate detection of anomalies in time series data. It demonstrates strong potential for application in power grid operation and maintenance， significantly improving fault detection accuracy， reducing false positives， and enhancing the stability and reliability of the power grid. 

Key words: deep learning; power grid data; anomaly detection; power grid operation and maintenance; power grid monitoring; time series