The power data is usually stored in the form of structured tables in the data management platform, and it is difficult for people to intuitively perceive the numerical data, so the data is usually described in the form of short texts, and the data is classified and graded through the description to meet the needs of differentiated data management. In practical applications, classification and grading are usually judged by manual workers according to certain standards. In the face of a large amount of data, it not only consumes a lot of labor costs, but also causes classification errors due to subjective factors of different personnel, resulting in low classification efficiency. At present, the automatic classification methods for data usually take the data itself as the classification object and extract features from the perspective of the data for classification, lacking people’s subjective semantic information about the data. Therefore, this paper proposes a new automatic classification and grading method for power data, which is based on the short text describing the data information. The method of feature extraction and classification based on human perception of data is more in line with the actual classification and grading requirements.