Modern power systems with “double high” characteristics make power quality disturbance patterns more complex, and the accurate classification of multiple power quality disturbances becomes more difficult. In the feature extraction stage of traditional power quality disturbance classification methods, the extracted features are determined artificially. Thus, it is difficult to judge whether the extracted features are adequate for classification problems, and the coupling of multiple feature distribution will affect the separability of disturbance features. Therefore, this paper proposes a feature selection method based on granular computing to optimize the performance of the classification. Based on the original feature set, a multi-granularity space is constructed to reflect the difference in feature distribution. Then the optimal feature subsets under each granularity are mined to determine the effective and redundant classification features. The homogeneous base classifiers trained by optimal feature subsets corresponding to different granularity spaces are fused by the ensemble model. A new multi-granularity ensemble classification model for power quality disturbance is proposed. This method overcomes the problem of the existing techniques by searching for the optimal valuable information of a single granularity space in a multi-granularity classification, leading to other granularity spaces losing the useful information. The simulation results show that the multi-granularity feature selection algorithm can extract useful features for classification, and an integrated model can improve the classification performance of the model. This work is supported by the National Natural Science Foundation of China (No. 51807126).