中文摘要：

在信息物理系统（cyber physical systems,CPS）深度融合背景下,提出一种安全状态实时感知的相关向量机（relevance vector machine,RVM）数据驱动方法。RVM是贝叶斯概率框架下基于核函数的学习方法,通过多层先验的超参数设置获取模型参数的稀疏解,并采用伯努利分布获得分类后验概率。该方法首先根据日前市场的运行与调度规则,产生运行条件,构造安全评估特征集及事故安全分类;然后将基于距离的Relief算法用于特征排序,筛选出与分类紧密相关的特征子集;最后通过RVM分类学习对系统安全状态进行辨识。IEEE 30节点系统测试结果表明,RVM方法的极度稀疏性、高分类精度、概率输出在实时安全状态感知中具有显著优越性。

英文摘要：

Aim to the integration of cyber physical systems（CPS）, a relevance vector machine（RVM） based data-driven method was proposed for real-time static security situational awareness. RVM is a general Bayesian probabilistic framework to learn the kernel-based classification model, in which a set of hyperparameters are imposed to the hierarchical priors over model parameters for obtaining the sparse solutions, and the Bernoulli distribution is incorporated to output a consistent estimation of the posterior probability. The operation conditions were firstly generated according to the dispatches of the day-ahead markets and the pre-fault feature sets with contingency class memberships were obtained. Then a distance-based Relief algorithm was employed for feature rank and selection. Finally, RVM learning for classification was applied for security recognition. A case studied in the IEEE30-bus system shows the proposed method can provide exceedingly sparse solutions, high accuracy and probabilistic outputs, further clarifying its superiority in security awareness.