In this paper, a new deep feedforward network (DFN)-based method for complex power quality disturbances recognition is proposed, aiming at solving the problem of low recognition accuracy and poor generalization performance. Firstly, original disturbance signals are processed by incomplete S-transform at several important frequency samples. Then,some distinctive features are extracted from the result of incomplete S-transform. Finally, a three-layer DFN classifier is constructed and trained, with Dropout regularization to improve the generalization and noise immunity. The simulation and experiment results show that the proposed method can effectively identify 17 types of disturbances, including 8 types of complex disturbances. The results in different noise levels indicate that the method also has commendable anti-noise and generalization performance. Compared with existing methods such as CART decision tree, extreme learning machine and random forest, the proposed method has higher recognition accuracy, better robustness and good application prospects.