In the context of low-carbon development, the secure anonymization of big data in digital power grids faces technical challenges. Traditional methods mainly rely on simple data anonymization or generalization techniques, such as directly deleting or replacing sensitive fields. However, these methods do not fully consider the inherent correlation between data and the differences in attribute sensitivity, resulting in a significant decrease in the usability of processed data. In this regard, this paper proposes research on the secure anonymization processing technology of big data in digital power grids under the low-carbon background. The random forest algorithm is used to predict the sensitivity values of power grid data attributes, multiple decision trees are constructed and their prediction results are integrated to effectively capture nonlinear relationships and complex patterns in the data, thereby accurately identifying the sensitivity of each attribute. Subsequently, the K-means clustering algorithm is used to cluster the attributes of the power grid big data, grouping attributes with similar sensitive characteristics into one category. On this basis, the K-anonymity algorithm is used for generalization and suppression operations to conceal the attribute clusters of power grid big data to varying degrees, thereby balancing the relationship between data privacy protection and data availability. The test results show that after using the method proposed in this paper for anonymizing power data, the iterative value of data information loss is 0.25, which has a relatively ideal anonymization effect.