一种基于特征映射与深度学习的虚假数据注入检测方法

智能电网逐步发展为大型电力信息物理系统，信息与物理系统的交互降低了其抵御虚假数据攻击（false data injection attacks, FDIA）的能力。针对这一问题，研究并提出了一种基于多层递阶融合模糊特征映射方法（multi-layer hierarchical fusion fuzzy feature mapping, MLHFFFM）与条件深度信念网络（deep belief network, DBN）相结合的智能电网虚假数据注入检测方法。首先，对FDIA原理进行分析，基于MLHFFFM结合主成分分析法对智能电网负荷数据进行聚类，选取日负荷与预测日类似的近似日；然后，提出利用条件深度信念神经网络对近似日智能电网负荷进行分析，通过选取不同参数对日负荷特征进行动态捕捉从而检测FDIA；最后，结合某省实际负荷以IEEE33节点系统为例进行分析。案例分析结果表明，所提模型相比于其他模型，在不同攻击强度下准确率均保持在95%以上，错报率在5%以下，能够有效检测出虚假数据的注入。

A Method to Detect False Data Injection Based on Feature Mapping and Deep Learning

The smart grid is gradually being developed into large power cyber physical system, however the interaction between cyber and physical system reduces its ability to withstand the false data injection attack (abbr. FDIA). In allusion to this issue, a method to detect the fake data injection for smart grid based on multi-layer hierarchical fusion fuzzy feature mapping (abbr. MLHFFFM) combined with deep belief network (abbr. DBN) was researched and proposed. Firstly, the principle of FDIA was analyzed and based on MLHFFFM and combining with principal component analysis (abbr. PCA) the smart grid load data was clustered to select the approximate day with daily load similar to that of the forecasted day. Secondly, the conditional DBN was used to analyze the daily grid load of approximate day, and in the meantime, by means of selecting different parameters the daily grid load characteristics was dynamically captured thereby the FDIA was detected. Finally, combining with actual load data of a certain province in China, IEEE 33-bus system was taken for computing example to verify the proposed method. Experimental results show that compared with other models the accuracy rate of the proposed model under different attack intensities keeps above 95%, and the error rate is lower than 5%, thus, the injection of fake data can be effectively detected.