To solve the problem of redundancy and low recognition accuracy in traditional power quality disturbance recognition, this paper proposes a method based on an improved sparrow search algorithm (ISSA) to optimize feature selection and eXtreme gradient boosting (XGBoost). First, the power quality disturbance signal is transformed by S-transform, and 61 kinds of power quality characteristics are extracted. Then ISSA selects the optimal feature subset and the optimal parameters in XGBoost at the same time to eliminate redundant features and improve recognition accuracy. Finally, the power quality disturbance is identified according to the optimized feature subset and XGBoost. The simulation results show that the method proposed can effectively select the optimal feature subset, and efficiently identify 19 kinds of power quality disturbance signals in a noisy environment, and has high recognition accuracy.