Detection of false data injection in smart grid using PCA based unsupervised learning

Advanced metering infrastructure (AMI) is one of the core aspects of the smart grid, and offers numerous possible benefits, such as load control and demand response. AMI enables two-way communication, but it is vulnerable to electricity theft. Due to the tempering of the smart meter, the abnormal pattern of fraud becomes difficult to detect, which introduces the increment of false data consumption. The majority of existing methods depend on factors such as predefined limits, extra information requirements, and the desire for labelled datasets. These factors are difficult to realize or have a poor degree of identification. This paper integrates two novel techniques to detect the False Data Injection attack. One is Principal Component Analysis, which is based on feature correlation, second is an unsupervised learning based technique Density-Based Spatial Clustering of Applications with Noise which helps to identify data patterns to detect outliers from a huge number of load profiles. The combination makes the proposed technique an appropriate tool for detecting an arbitrary pattern attack in high-dimensional data. Four different attack scenarios are analyzed on the Irish Science Data Archive smart meter data set. The efficacy of proposed theft detection method is evaluated by comparing the AUC, mAP, and time with those of other clustering approaches. The detection rate of the proposed scheme has been compared with the other approaches in the literature. The results demonstrate that the detection rate is substantially higher than that of other theft detection methods. Finally, the influence of varied abnormality ratios has been investigated.