For the big data on electric power, many specific applications, such as load forecasting and fault diagnosis, need to consider data changes during a period of time to determine their decision classes, as deriving a class label of only one data record is meaningless. Based on the above discussion, interval-valued rough set is introduced into big data classification. Employing algebra and information theory, this paper defines the related concepts and proves the properties for interval-valued reductions based on dependency and mutual information, and presents the corresponding heuristic reduction algorithms. The proposed methods can not only enrich and develop the interval-valued rough set theory, but also provide a new way for the analysis of big data. Pertaining to the distributed data storage architecture of big data, this paper further proposes the interval-valued global reduction in multi-decision tables with proofs of its properties. The corresponding algorithm is also given. In order for the algorithms to achieve better results in practical applications, approximate reduction is introduced. To evaluate three proposed algorithms, it uses six months’ operating data of one 600MW unit in some power plant. Experimental results show that the three algorithms proposed in this article can maintain high classification accuracy with the proper parameters, and the numbers of objects and attributes can both be greatly reduced.