As advanced metering infrastructure (AMI) is deployed, the AMI data is used to detect an energy fraud. Along with context analysis of AMI data such as detection of an unreasonably low consumption AMI data mining is a primary solution for detecting abnormalities that cannot be detected using simple context analysis. Traditionally, abnormality detection based on AMI data mining compares a load profile with predefined normal prototypes. However, since a load profile can be normal in one condition and abnormal in another, the condition associated with the load profile should be considered as determining the normality. However, existing methods do not connect the normality of a prototype and a specific condition. In this study, the authors propose a mechanism that incorporates the conditional probability into determination of the normality of the prototype for comparison. The novelty of their study is its generating a two‐dimensional space using similarity and conditional probability, so that several multi‐dimensional classification methods can be applied. They compare the proposed mechanism with best‐fit and average prototype‐based abnormality detection methods. In conclusion, the proposed mechanism can distinguish fraud data with a higher precision than the traditional methods. They also explore the accuracy of the mechanism with various parameters. [ABSTRACT FROM AUTHOR]