基于不平衡高维数据特征重要性的电力系统异常检测方法

An Anomaly Detection Approach Considering the Feature Imbalance and Importance in Power Systems

为提升电力系统对攻击和异常的检测能力,提出一种基于特征重要性的数据驱动异常检测方法。首先,通过合成少数类样本过采样技术对原始数据进行预处理,调整数据集的平衡性;其次,结合邻域成分分析特征提取和“统计特性融合”特征选择方法对数据集进行特征处理;然后,采用贝叶斯优化对机器学习分类器进行参数调整;最后,在Power System Attack Datasets多类数据集上进行实验验证。实验结果表明,所提方法在电力系统的三分类和多分类数据集上,分别实现了98.97%和95.59%的分类准确率,相较于其他异常检测方法,可以有效区分攻击和故障,且在准确率和误报率上均表现得更好。

To improve the ability of power systems to detect cyberattacks,a data-driven anomaly detection approach is proposed.Firstly,the raw data is balanced by using synthetic minority over-sampling technique.Secondly,neighborhood components analysis feature extraction technology with the “fusion of statistical importance” feature selection method is combined to process the features.Thirdly,the Bayesian optimization is used to fine-tune the parameters of the machine learning classifier.Finally,experiments on the Power System Attack Datasets is conducted.The results demonstrate that the designed system achieves a triple classification accuracy of 98.97% and a multi classification accuracy of 95.59%.Compared to other anomaly detection methods,the proposed approach not only distinguishes between natural faults and cyberattacks effectively but also outperforms in accuracy and false alarm rate.