This paper proposes a real time power quality classification based on improved incomplete S-transform and decision tree, which mainly focuses on classification accuracy and computing time. In order to reduce the restriction of Heisenberg’s uncertainty, different signal components are windowed by different Gauss windows according to the signal components frequency in the spectral, which reduces the response time and enhance the tolerance of noises. The feature extraction is implemented by using dynamics to the result of the improved incomplete S-transform. Finally, an optimal decision tree is constructed to classify the power quality disturbances through five distinctive features. A hardware based on DSP-FPGA is used to test the proposed method. Both simulations and experiments verify the practicability of the method.