In order to solve the problem of feature selection and classification accuracy in complex power quality disturbances recognition, a method of feature selection and classifier construction based on incomplete S-transform and Gradient Boosting Decision Tree (GBDT) is proposed. Firstly, the time-frequency matrixes of the disturbances are obtained by the incomplete S-transform at the specific frequencies. Then the original feature set of 53 features is extracted from the time-frequency matrix, and the importance level of each feature is measured by GBDT to select optimal features. Finally, the disturbance classifier is trained and constructed by GBDT based on the selected feature set. The simulation results show that the classification accuracy of the proposed method is higher than the CART, the Random Forest (RF), the Multilayer Perceptron (MLP), and so on, for a total of 17 disturbance classes including 8 classes of combined disturbances. Moreover, the classification results in different noise conditions indicate that the method has good anti-noise performance and algorithm robustness, which shows good prospects for application. This work is supported by National Natural Science Foundation of China (No. 51277080).